USER INTERFACE FOR PROVIDING SOURCE TRACEABILITY WITHIN AN INFORMATION EXTRACTION SYSTEM

[0001] This application is a Continuation-in-Part of U.S. patent application Ser. No. 19/327,863 filed on Sep. 12, 2025, which is a continuation of U.S. patent application Ser. No. 18/831,434 filed on Jan. 24, 2025 (Now U.S. Pat. No. 12,437,155), both of which are herein incorporated by reference in their entirety.

[0002] This disclosure generally relates to more efficiently providing data to a large language model, and more particularly, to sending the table portion of a PDF to a large language model.

[0003] A PDF is a portable document format. A PDF is often used as a format for saving documents that should not be modified, but the document still needs to be shared and printed. PDF documents may be categorized in three different types. The types may include true PDF pages, image-only PDF pages (or scanned PDF pages) and searchable PDF pages. The PDF category may depend on the way the file was originally created. The way the document was originally created also defines whether the content of the PDF (e.g., text, images, tables) can be accessed or whether the content may be inaccessible (or “locked”) in an image of the page. Optical character recognition (OCR) can be applied to PDF files to generate machine-encoded text.

[0004] OCR is an electronic conversion of images of text into machine-encoded text. Thus, the use of OCR is necessarily rooted in computer technology. In its most common application, OCR is performed on a scanned or photographed document to detect the text of the document. After the text is detected using OCR, the text may be selected, searched, or edited by software executed by a computer.

[0005] An embodiment of the present disclosure relates to a system for providing source content for information extracted by language models. The system includes one or more processing circuits configured to generate a plurality of chunks from one or more documents, each respective chunk associated with a corresponding reference to a portion of the one or more documents that corresponds to the respective chunk. The one or more processing circuits are also configured to identify one or more relevant chunks from the plurality of chunks based on a search criterion. The one or more processing circuits are also configured to transmit, to a language model, one or more prompts including (i) a first request to extract a value for a data field from the one or more relevant chunks and (ii) a second request for the language model to identify one or more used chunks of the one or more relevant chunks used to extract the value. The one or more processing circuits are also configured to generate instructions for a user interface including the value extracted and the corresponding reference for the one or more used chunks.

[0022] Different types of businesses often carefully curate and extract a large volume of documents. For example, a large set of insurance documents or accounting documents (in the form of images and/or PDFs) are typically sent to an insurance broker or a tax preparer, who then has the task of identifying and extracting relevant information from the accounting documents. To provide more efficiency, businesses have tried to automate this workflow by incorporating template-based optical character recognition (OCR). Businesses have also used rigid, specific rule-based methods. For example, businesses would often perform optical character recognition to use the expected positioning of text on a document to both identify the document type and to further extract and annotate data from that document.

[0023] Template-based OCR often includes trained humans to create each template. A human with detailed knowledge of the OCR system and document variability must review every document to specifically create sets of rules detailing exactly how to extract data from each of the documents. Template-based OCR also usually requires trained humans to maintain each template. However, templates often degrade in performance as documents change. While some variability can be explicitly declared in the template, any unaccounted-for changes usually require humans to modify a template to account for the differences or to create a new template.

[0024] Moreover, template-based OCR approaches typically cannot adapt to dynamic documents. In particular, some documents may be structured, but the documents have variability in table length and data positioning. While a human is often able to detect these similar regions of interest from document to document, it is difficult to create generic rules/templates that are able to capture these regions and specify how to process the regions.

[0025] These concerns about PDFs and OCRs are even more pronounced when systems feed documents to a large language model (LLM) using a retrieval-augmented generation (RAG) architecture. Previous methodologies index (e.g., embed, vectorize, etc.) portions of the documents, using chunks that include tabular information and document text so that the portions of the document can be provided with the prompt (e.g., the document augments the prompt) to the LLM. OCR output of a PDF with text and tables may include the tabular information intertwined with the document text such that the order of the text is no longer appropriate, and the semantic meaning of the document is lost if processed in the natural order of the language (e.g., left to right and top to bottom in English). For example, documents may have tables inline with the text and/or documents may have tables at the top or bottom of a page for which the text description is on a different page.

[0026] RAG-based systems often identify relevant documents or portions thereof based on a distance metric for a vectorized embedding of the text of the document. The embedding models are designed such that text with similar semantic meaning will have vector embeddings that are near each other (e.g., as measured by the distance metric). Including the table text and body text together can damage the semantic meaning of the document and the RAG-based system may not find appropriate documents or portions thereof to provide to the LLM leading to poor response accuracy.

[0027] Previous methodologies of RAG-based data annotation and/or extraction also provide portions of the document (e.g., chunks) that include the tabular information and the text information to the LLM. Even if the retrieval augmentation identifies an appropriate document, the LLM may not be able to sufficiently distinguish the tables, data from the tables, text and other content. If the text, as provided to the LLM, does not have appropriate meaning when processed in sequence by the LLM, a suitable response may not be generated. In such scenarios, the LLM may note the error and not return the requested data; or worse, the LLM may return a nonsensical response as if it is correct (e.g., the LLM may hallucinate the output).

[0028] Moreover, the number of computations performed by the LLM (and thus the fees associated with using an LLM) depend on the number of letters or words (e.g., tokens) submitted to the large language model. Providing the LLM with tabular data and document text can increase computations and therefore is not energy or cost efficient. In addition, incorrect responses for the reasons described above may result in the prompt and/or the retrieved portions of the documents being adjusted and resent to the LLM leading to further computational inefficiencies.

[0029] The present disclosure improves the technological field of RAG-based generative artificial intelligence (AI) systems by separating the table data and the table text from the body text using markdown language tags provided by the OCR process. Separating the tabular information from the document text improves the quality of the index created allowing for improved document retrieval. In addition, by separating the tabular information from the document text prior to providing it to the LLM allows the LLM to consider both tabular data and text data independently. The LLM can understand the row and column relationship used by the table and generating responses including data obtained from the table. Further, with the tabular information removed from the text, the LLM can process the text semantically and generate appropriate responses using information from the text.

[0030] As a result of the improvements to the RAG-based generative AI systems and methods described herein, a larger portion of the data to be populated can be accurately determined and extracted from the documents leading to reduction in labor associated with data correction. The improved index leads to improved identification of the information retrieved using semantic search and provided to the LLM. Improved retrieval has a direct effect on the accuracy of data population. Separated tabular information and document text prevents the two coupled data sources from confounding the LLM, again leading to improved functioning of the computer hardware executing the LLM in the form of enhanced accuracy that reduces the need for reprocessing of prompts and/or retrieval of additional documents. Computational effort by the LLM is thus reduced. Moreover, the text data and tabular information are not required to both be provided to the LLM. The LLM can focus computational effort on only the portions of the document that are necessary to extract the information for the given prompt again reducing computational effort, energy usage, and operational costs of the system.

[0031] The system may also provide auditability, traceability and/or compliance when submitting data to a large language model (LLM). In particular, the system may use chunk identifiers and indexing to label and trace the text used by the LLM, so that the system may determine where the text originated. In that regard, the system may determine if an LLM is hallucinating because the system may determine that the LLM generated texted that was not provided by the source document.

[0032] FIG. 1 shows a data extraction and population system 100 configured to leverage a large language model (LLM) to extract data from documents and populate data elements (e.g., of a data model, ontological data store, etc.), according to some embodiments. The data extraction and population system 100 is shown to include one or more UI clients 102, one or more data sources 104, an OCR system 106, an LLM 108, a text embedder 110, and a data extraction manager system 200 communicably connected via a network 112. FIG. 1 shows a non-limiting example of a possible configuration of the data extraction and population system 100. It is contemplated that the various components of the data extraction and population system 100 may be distributed across discrete systems and/or hardware in different ways. For example, the large language model 108 and the text embedder 110 may be configured within the same hardware or same node in a computer cluster or the data extraction manager system 200 may be distributed across multiple elements of computer hardware.

[0033] In some embodiments, the general operation of the data extraction and population system 100 is to extract data from documents and populate various data elements, according to some embodiments. The data extraction manager system 200 may gather documents from the one or more data sources 104 and generate a searchable index of documents or portions thereof from the one or more data sources 104 using the text embedder 110. The index generation may be based on the semantic meaning of the documents from the one or more data sources 104, allowing comparison between the entries of the index and a prompt for data (e.g., the prompt also embedded by the text embedder 110). To populate the data elements, the data extraction manager system 200 may generate prompts for the data, identify relevant portions of the documents by searching the index, and provide both the prompt and the relevant portions of the documents to the LLM 108. The LLM 108 may then process the prompt with the provided portions of the document to extract (e.g., identify, parse, summarize, combine, generate, etc.) the data requested by the prompt so that the data extraction manager system 200 can store the data (e.g., in an object, a data model, ontological model, an ontological data store, etc.).

[0034] In some embodiments, the data extraction and population system 100 gathers large amounts of data from the one or more data sources 104. The one or more data sources 104 may be internal (e.g., on the company intranet) or external (e.g., stored on another company's web server). The one or more data sources 104 may include dedicated databases for particular types of data or webpages from which documents may be compiled, scraped, etc. The one or more data sources 104 may include documents (e.g., files, records, reports, articles, forms, data, etc.). The documents in the database may contain text, tables, columns, rows, charts, graphics, images, and/or other content. The documents may include PDF files or other image-based files for which the text of the document is not readily available for searching, copying, etc. Such image-based files, may be processed by the OCR system 106 prior to processing by other components of the data extraction and population system 100. While this disclosure may discuss the use of tables in a PDF, one skilled in the art will appreciate that similar systems and methods may be applied to any arrangement of data on a PDF document such as, for example, tables, charts, graphs, columns, rows, pie charts, bar charts, trend lines, and other graphical items. The system may also be used for tables within tables. The documents may include a variety of content such as, for example, in the insurance industry, applications, broker correspondence, financials, summary of claims, historical claims filed under business insurance policies (“Loss Run”) and historical claim losses.

[0035] The OCR system 106 may be configured to convert the contents of the document to plain text. The OCR system 106 may include, for example, any commercially available OCR system. Additionally or alternatively, the OCR system 106 may be a component of the data extraction manager system 200 (e.g., using available OCR software). The system may use this type of private OCR system 106 for increased security. The text extraction tool may convert an image-based document (e.g., PDF file, PostScript, tagged image file format (TIFF), etc.) plain text that can be processed by a computer (e.g., the American Standard code for Information Interchange (ASCII)). In some embodiments, the plain text is stored in a plain text file format for later processing. For example, the plain text may be stored in plain text file formats such as TXT or markup languages such as hypertext markup language (HTML), JavaScript Object Notation (JSON), extensible markup language (XML), tau epsilon chi (TeX), etc. (e.g., into a text format (e.g., JSON). JSON is a text format that is completely language independent, but uses conventions that are familiar to programmers. JSON may also be better than OCR because JSON retains positional relationships in the text (positional encoding).

[0036] The documents processed by the OCR system 106 may include non-text-based information (e.g., charts, graphs, trend lines, flow charts, or other graphical elements) and/or special text structures (e.g., tables, rows, columns, etc.). This information may be recognized by the OCR system 106 as different from the text of the body of the document and may indicate the presence of special structures (e.g., non-text-based information and/or special text structures) in the output.

[0037] The OCR system 106 may return output in the JSON text format. The output may include an object for any special structures in the document with a key-value pair for the location of the special structure within the original document. The key-value pair for the location may include, for example, the X-Y position of each of the four corners for each of the tables in the document or the X-Y position of each cell in the tables, or the key-value pair for the location may include the two X limits of the table and the two Y limits of the table. Each PDF analyzed by a text extraction tool may have the same orientation and coordinates. The X-Y positions may describe a table, row structure, column structure, and/or cell structure.

[0038] In some embodiments, the OCR system 106 returns an output with tables inline with the text using a markdown language. The system may use the same markdown symbols to indicate different locations or different markdown symbols to indicate different locations. For example, the first appearance of the markdown symbol indicates the start (or top) of a table and a second appearance of the same markdown symbol indicates the end (or bottom) of the table. The markdown symbols may also indicate a first (e.g., left) side of the table and a second (e.g., right) side of the table. Markdown symbols (e.g., within text) may provide characteristics of the table. The markdown system may provide information to the system, so the system may render the table. For example, the vertical bar or pipe character, ‘|’, may be used to mark the start of a new column within a row of the table, and the vertical bar followed by a newline character (e.g., ‘|/n’) may be used to represent a new row. The markdown language may also use hyphen characters, ‘-’, to separate a header row from a content row within a table. When analyzing the position of each cell, the system may consider each cell as having a single row of text, regardless of the number of lines of text in each cell. For more information about markdown symbols, see www.markdownguide.org/extended-syntax/.

[0039] In some embodiments, the OCR system 106 returns an output in a first format, and the data extraction manager system 200 may convert the text into a second format (e.g., a common format) prior to processing by other components of the data extraction and population system 100. For example, the data extraction manager system 200 may convert the JSON output (e.g., with location data) to markdown language that includes markdown symbols. The JSON web language may be translated to markdown text indicating one or more boundaries of the table. Modularity is provided by converting to a common text format (e.g., the markdown language) allowing the data extraction and population system 100 to substitute other various OCR systems 106 if there is a cost advantage, computational advantage, or an improvement by one provider of OCR technology.

[0040] The data extraction manager system 200 may be configured to coordinate the operations of the data extraction and population system 100. For example, the data extraction manager system 200 may initiate (e.g., at the request of a user of the one or more UI clients 102) document gathering from the one or more data sources 104. The data extraction manager system 200 may communicate (e.g., send, deliver, transmit, etc.) the PDFs or other image-based documents to the OCR system 106 for conversion to plain text. The data extraction manager system 200 may separate the document text from the tabular information before chunking (e.g., splitting text into word lengths that are suitable for retrieval augmentation of, for example, 500 words, 1000 words, 1000 characters, etc.). The data extraction manager system 200 may communicate the chunks (both tabular chunks and text chunks) to the text embedder 110 to build an index for semantic search.

[0041] Upon receiving a request from a user of the one or more UI clients 102, the data extraction manager system 200 may generate several prompts for data extraction (e.g., identification, summarization, generation, etc.) for processing by the LLM 108. In some embodiments, the data extraction manager system 200 is configured to embed each prompt (e.g., using the text embedder 110 or similar embedding model) and compare the prompt vector embedding to that of the index to identify and retrieve potentially related or relevant chunks (e.g., portions of the documents). The prompts, along with the identified relevant chunks, may be communicated to the LLM 108 by the data extraction manager system 200. In some embodiments, the data extraction manager system 200 is also configured to store the results of a prompt from the LLM 108. Thereby, the data extraction manager system 200 manages the population of the particular data elements by retrieving both structured and unstructured data, text, tables, etc. from various sources across the local intranet or the internet.

[0042] The data extraction manager system 200 may also generate user interfaces for the data extraction and population system 100. For example, the data extraction manager system 200 may communicate instructions (e.g., JavaScript, Cascading Style Sheets, etc.) to generate a user interface to the one or more UI clients 102. The user interface may provide interactive capability with the systems of the data extraction and population system 100. For example, the user interface may provide the ability to initiate data population, configure the data to populate or extract, view results, trace errors, view source material, and/or other interactions that may be appropriate for a particular use case.

[0043] The text embedder 110 may be configured to generate a vector embedding for a chunk of text. The vector embedding may refer to a vector representation of the semantic content of the chunk of text. Vectorization gives text numerical values that can be searched, with computational efficiency, for similarity (e.g., using a distance metric); thereby, text with similar semantic content can be identified for retrieval. Similar words would have similar numerical values. For example, hot and cold may have vectors pointing in different directions. The system may not find the word “cat”, but with vectors, the system will determine that lion is similar to cat or big+cat. The text embedder 110 may be trained to understand the meaning of the words (female+king=queen).

[0044] After the vectors are created, the text embedder 110 may communicate the vector embeddings of the text chunks to the data extraction manager system 200 for storage in an object (e.g., a vector store). In some embodiments, the text embedder 110 may be included as a component of the data extraction manager system 200.

[0045] The LLM 108 may be any type of artificial intelligence (AI) configuration. For example, the LLM 108 may include generative pre-trained transformers (GPT), bidirectional encoder representations from transformers (BERT), text-to-text transfer transformers (T5), recurrent neural networks (RNN), or any other AI architecture suitable for a large language model. The LLM 108 may be configured to output a text response from a textual prompt. For example, the LLM 108 may convert text of a prompt into tokens representing a unit of information (e.g., a character, word, prefix, punctuation, etc.) and use the input sequence tokens to predict each output word (or token) consecutively. The prompt communicated to the LLM 108 may include chunks from the documents gathered from the one or more data sources 104 so that the LLM 108 is able to use that information to generate its response. For example, the LLM 108 may be provided a prompt including a request to determine the range of the market capitalization of a company over the last 6 months and one or more table chunks or text chunks that include information that may be relevant for such a question.

[0046] The LLM 108 may be a publicly available LLM such as Claude. The LLM 108 may be pre-trained on massive corpora of text data, allowing it to learn the statistical properties of language and predict output text based on the prompt. In some embodiments, the LLM 108 may be fine-tuned, for example, to extract specific data from tabular and/or textual input. Fine-tuning a LLM may refer to the process of taking a pre-trained model and further training it on a specific dataset to adapt it to a particular task or domain. Fine-tuning may allow the LLM 108 to leverage its existing knowledge while improving its performance on the new, specialized data. For example, by focusing on the correlations found in the particular task or domain.

[0047] The one or more UI clients 102 may provide users, administrators, and/or developers of the data extraction and population system 100 access to its features. In some embodiments, the one or more UI clients 102 are used to generate a user interface that allows for interaction with the components of the data extraction and population system 100. For example, the one or more UI clients 102 may be used to initiate data population, configure the data to populate or extract, view results, trace errors, view source material, and/or other interactions that may be appropriate for a particular use case. The one or more UI clients 102 provide various inputs (e.g., selecting user interface objects, entering text into fields, etc.) and various outputs (e.g., display, print, email, or transmission to another system) to/from the data extraction and population system 100.

[0048] The network 112 can include routers, switches, antennas, computers, and any other hardware required to communicate information between the components of the data extraction and population system 100 (e.g., from the data extraction manager system 200 to the LLM 108). A portion of the network 112 can be wireless and/or a portion of the network 112 can be wired. The network 112 can include one or more networks with routers to facilitate data transfer between the different networks.

[0049] In one use case where the data extraction and population system 100 is particularly useful is to extract data for the underwriting process of insurance policies. For example, directors and officers liability insurance and/or environmental insurance require extracting large amounts of information for which there is no central repository. The information may be collected about the company, the directors and officers, and/or any business locations. Manually searching for this information is error prone and requires a large time investment for the underwriters. Moreover, much of the data that is to be extracted for insurance underwriting may be found in financial tables of image-based documents (e.g., PDFs) making the systems and methods of separating tabular information and text information described herein particularly useful in such scenarios.

[0050] Continuing with the example of insurance underwriting, the user of the data extraction and population system 100 may be an insurance underwriter. They may have a specially curated set of data elements that they require to perform the underwriting process of different types of insurance policies. A type of insurance policy may be considered a task for which the data extraction and population system 100 is configured to populate the data elements of an ontological data store related to that type of insurance policy. The insurance policy may be associated with one subject (e.g., companies, people, buildings, etc.) for which the insurance policy is to be underwritten. After data is populated, the underwriter may review the information and or generate a report. For regulatory purposes, the data used to generate the report may require citation to the source of the information. Systems and methods described herein may allow for such traceability.

[0051] FIG. 2 shows a block diagram of the data extraction manager system 200, according to some embodiments. In some embodiments, the data extraction manager system 200 is configured to coordinate the processes performed by the data extraction and population system 100 during the data extraction and population. The data extraction manager system 200 of FIG. 2 is shown as a single entity (e.g., hardware). However, it is contemplated that the components and/or instruction sets included in the data extraction manager system 200 could be distributed over any number of computer hardware devices and in any manner of architecture (e.g., local network, cloud-based, etc.).

[0052] The data extraction manager system 200 is shown to include a communications interface 202, and one or more processing circuits 204 having one or more processors 206 and memory 208.

[0053] The communications interface 202 may be configured to facilitate communication between the data extraction manager system 200 and other components of the data extraction and population system 100. For example, the communications interface 202 may transmit information onto the network 112 and/or receive information from the network 112.

[0054] The one or more processors 206 may be general purpose or specific purpose processors, an application specific integrated circuit (ASIC), one or more field programmable gate arrays (FPGAs), a group of processing components, or other suitable processing components. The one or more processors 206 may be configured to execute computer code and/or instructions stored in the memory 208 or received from other computer readable media (e.g., CDROM, network storage, a remote server, etc.). The one or more processors 206 may be configured in various computer architectures, such as graphics processing units (GPUs), distributed computing architectures, cloud server architectures, client-server architectures, or various combinations thereof. A first set of the one or more processors 206 can be implemented by a first device, such as an edge device, and a second set of one or more processors 206 can be implemented by a second device, such as a server or other device that is communicatively coupled with the first device and may have greater processor and/or memory resources.

[0055] The memory 208 may include one or more devices (e.g., memory units, memory devices, storage devices, etc.) for storing data and/or computer code for completing and/or facilitating the various processes described in the present disclosure. The memory 208 may include random access memory (RAM), read-only memory (ROM), hard drive storage, temporary storage, non-volatile memory, flash memory, optical memory, or any other suitable memory for storing software objects and/or computer instructions. The memory 208 may include database components, object code components, script components, or any other type of information structure for supporting the various activities and information structures described in the present disclosure. The memory 208 may be communicably connected to the processors and can include computer code for executing (e.g., by the processors) one or more processes described herein. For example, many of the components of the data extraction manager system 200 illustrated in FIG. 2 may be implemented as instruction sets stored by the memory 208 and executed by the one or more processors 206.

[0056] In FIG. 2, the data extraction manager system 200 is shown to include a coordinator 212, a data manager 220, an ingestion manager 240, a generative AI manager 260, an interface manager 280, and enabling services 290, according to some embodiments. The coordinator 212 may be configured to control the timing and flow of data through the other circuitry of the data extraction manager system 200. For example, the coordinator 212 may cause the modules or circuits to execute in a specific order to perform the function of the data extraction manager system 200. In some embodiments, the coordinator 212 may route the information and/or outputs of other modules that are dependent on the information or use the information as an input.

[0057] The data manager 220 may be configured to manage the data gathering process of the data extraction and population system 100, including gathering documents from the one or more data sources 104. The ingestion manager 240 may be configured to identify image-based documents (e.g., PDFs) and coordinate the processing of the image-documents with the OCR system 106. The ingestion manager 240 may also be configured to separate text from other information that may be in documents (e.g., tables, graphs, etc.) and manage the creation of a semantic search index using the text embedder 110. The generative AI manager 260 may be configured to generate prompts (e.g., from templates) to cause the LLM 108 to extract data from retrieved documents. The generative AI manager 260 may coordinate the retrieval of relevant portions of documents (e.g., table chunks and/or text chunks) to supply as part of the prompt to the LLM 108. The interface manager 280 may provide for interaction with a user of the data extraction and population system 100 and/or an administrator of the data extraction manager system 200. In some embodiments, the enabling services 290 provide deployment support, security, and monitoring for the data extraction and population system 100.

[0058] In some embodiments, the data manager 220 includes a request manager 222, a data scraper 224, internal data storage 226, and an ingestion initializer 228.

[0059] In some embodiments, the request manager 222 coordinates the document gathering for a particular task. The request manager 222 may be configured to receive a request to begin data gathering for a particular task. The request manager 222 may, at the request of a user (e.g., through the user interface on the one or more UI clients 102), cause the data scraper 224 to begin searching the one or more data sources 104 for documents that may contain information to be used to populate the data elements or data model. In some embodiments, the request manager 222 may communicate information related to the particular sources of the one or more data sources 104 that should be searched for information. For example, the request manager 222 may receive a set of particular sources of the one or more data sources 104 that should be searched. Additionally or alternatively, the request manager 222 may receive a type of request for which the data scraper 224 has a predetermined list of potential sources. In some embodiments, the request manager 222 may report status back to the user (e.g., to the one or more UI clients 102) in the form of a percent complete. The request manager 222 may also accept individual sources from the user (e.g., from the one or more UI clients 102). For example, the user may provide a data source that the data scraper 224 is not preprogrammed to search. Additionally or alternatively, the user may upload documents to the data extraction manager system 200 that can be stored by the request manager 222 using the internal data storage 226.

[0060] The data scraper 224 may be configured to gather information from various sources, including the one or more data sources 104, additional data sources linked by a user (e.g., from the one or more UI clients 102), and/or documents uploaded by the user (e.g., after scanning a hard copy, receiving an email, etc.). The data scraper 224 may search databases, webpages, emails, and other internal and/or external sources of documents (e.g., text, data, image-based documents, etc.). The data scraper 224 may include a list of particular sources of the one or more data sources 104 that are to be searched for a particular task. For example, if the task includes gathering financial information, the data scraper 224 may gather data from Dun and Bradstreet using a POST request or by navigating to a particular web page. Additionally or alternatively, the data scraper 224 may use a web-based search engine (e.g., Google, Bing, etc.) and gather documents (e.g., text, PDFs, etc.) from a number of the top search results (e.g., top 10, top 50, etc.). In some embodiments, the data scraper 224 searches pre-approved websites that are returned from the search engine (e.g., websites that have been vetted to maintain currency and accuracy).

[0061] To gather documents, the data scraper 224 may visit a webpage and perform a keyword search or a semantic search to find information that may be used for a particular task. For example, the data scraper 224 may perform a keyword search or a semantic search against the file names of any documents stored in the one or more data sources 104. For plain text documents and/or webpages, the data scraper 224 may identify a keyword or a section that is semantically related to the task and gather the text for a number of words, characters, or sentences before and after the identified area of the text. The data scraper 224 may combine such text from multiple identified areas if the resulting text is overlapping. By gathering data both before and after the identified area the data scraper 224 may gather any information that may be useful for populating the data model both in its current form and potentially gathering information for future versions of the data model.

[0062] The data scraper 224 may be configured to store the gathered text and/or documents in the internal data storage 226 for processing by the OCR system 106 and/or chunking. In some embodiments, the data scraper 224 searches through all the one or more data sources 104 prior to an index for retrieval augmentation being built. Alternatively, each document may be added to the index as it is gathered, for example, to speed up operations by processing in parallel (e.g., gathering data while building the index) and/or to use internal data storage 226 more efficiently by discarding information that is deemed not useful. In some embodiments, the data scraper 224 may search the one or more data sources 104 until it finds an amount of documentation, or a number of documents related to each search or data that is to be populated. As such the data scraper 224 may be configured to ensure that the data extraction and population system 100 has a level of information available that is expected to successfully populate all or a threshold percentage of the data.

[0063] In some embodiments, the data manager 220 may be configured to periodically (e.g., based on a schedule) search for updates of the documents from the one or more data sources 104. The schedule may be entered by a user (e.g., via a user interface on the one or more UI clients 102). As updates to the documents are found and/or new documents are found, the ingestion manager 240 may add the new information to the retrieval index.

[0064] In some embodiments, the internal data storage 226 includes storage for both processed and unprocessed documents. The internal data storage 226 may include a data model and/or an ontology that includes structured storage for documents with properties for the document name, type (e.g., imaged-based, plain text, etc.), source (e.g., from which of the one or more data sources 104), if the document has been chunked, etc. The internal data storage 226 may include storage for each chunk of the documents, with properties that link the source document to enable traceability, the page of the source document from which the chunk is from, a chunk ID (e.g., sequential number, globally unique identifier, GUID, hash code, etc.), if the chunk is a table chunk or a text chunk, etc. The internal data storage 226 may include a vector store to store the vector embeddings of the chunks for the index. The vector store may be maintained separately from the other objects of the data model so as to allow efficient semantic search during retrieval augmentation. The internal data storage 226 may include prompt templates for a particular task or data elements to be populated. For example, the given data population task may include several data elements that are to be populated by the data extraction and population system 100 and the internal data storage 226 may include prompt templates that are used to cause the LLM 108 to extract the data from the documents for the particular data element (and thus allowing the data extraction and population system 100 to extract the data elements from the one or more data sources 104). The internal data storage 226 may include the data elements that are to be populated by the data extraction and population system 100. For example, at the initiation of a request (e.g., by the request manager 222) the data elements to be populated may be provided to the internal data storage 226 and populated during the data extraction and population process.

[0065] The internal data storage 226 may include storage for all the requests of the data extraction and population system 100 in a single data lake. Additionally or alternatively, a data lake may be generated for each request, providing data isolation and the ability to move the data between systems on a per request basis. The internal data storage 226 may be organized based on request, user id, or any other key to provide efficient operation.

[0066] The internal data storage 226 may be any type of non-transitory, computer readable storage medium. For example, internal data storage 226 may store data in magnetic hard disk, solid state drives, optical drives, RAM, and/or any other suitable storage medium. The internal data storage 226 may be distributed across one or more computer system, for example, communicably connected over the network 112.

[0067] The system may include remote access to data, standardizing data and allowing remote users to share information in real time. The system may allow users to access data (e.g., data from the database, text from the documents, table data, etc.), and receive updated data in real time from other users. The system may store the data (e.g., in a non-standardized format) in a plurality of storage devices, provide remote access over a network so that users may update the data that was in a non-standardized format (e.g., dependent on the hardware and software platform used by the user) in real time through a GUI, convert the updated data that was input (e.g., by a user) in a non-standardized form to the standardized format, automatically generate a message (e.g., containing the updated data) whenever the updated data is stored and transmit the message to the users over a computer network in real time, so that the user has immediate access to the up-to-date data. The system may allow remote users to share data in real time in a standardized format, regardless of the format (e.g. non-standardized) that the information was input by the user. This standardization of data improves communication between devices, improves the functioning of the system and improves the sharing of the data. In particular, the communications are streamlined without having to conduct data conversions because the users and systems may share data (e.g., in real time) in a standardized format.

[0068] In some embodiments, the ingestion manager 240 may include an OCR manager 242, a markup decoder 244, a table chunker 246, a text chunker 248, a chunk tracer 250, and an indexer 252. These components may provide functionality allowing the data extraction manager system 200 to identify image-based documents (e.g., PDFs) and coordinate the processing of the image-based documents with the OCR system 106 and prepare the text for retrieval within the RAG architecture of the data extraction and population system 100.

[0069] The OCR manager 242 may coordinate the interaction with the OCR system 106. The OCR manager 242 may be configured to receive image-based documents and output plain text files for those image-based documents. For example, the OCR manager 242 may request all unprocessed imaged-based documents from the internal data storage 226 and generate requests for processing by the OCR system 106. The OCR manager 242 may include instructions for communicating the documents to the OCR system 106, tracking their progress, and returning results back into the internal data storage 226. In some embodiments, the OCR manager 242 may have error handling code if the OCR system 106 is not able to appropriately process the documents. For example, the OCR manager 242 may flag the document as unusable, generate a request for the data scraper 224 to obtain additional documents from the one or more data sources 104 that include similar information, and/or use a secondary or back-up OCR system to perform the conversion to plain text.

[0070] In some embodiments, the OCR manager 242 may convert the output of the OCR system 106 into a standardized format. The OCR manager 242 may convert the output of the OCR system 106 into plain text using a markdown language to indicate various text structures and/or tables. For example, the OCR system 106 may return plain text in JSON format, and the OCR manager 242 may convert the JSON format into markdown. In some embodiments, more than one OCR system 106 is used, for example, as an alternative if an error occurs or the system is down. The OCR manager 242 may convert all outputs from an OCR system 106 into the format of the primary OCR system 106 or into a common format. In some embodiments, the text information from the one or more data sources 104 contains tables that are not image-based (e.g., Word documents or spreadsheets). Such documents may be provided to the OCR manager 242 for processing into the common markdown even if the document does not require OCR. For example, the OCR manager 242 may be able to read data directly from the Office Open XML (OOXML) structure of the documents. Additionally or alternatively, the OCR manager 242 may be configured to use inter-process communication, object linking and embedding, and/or component object model automation to extract plain text and tables from non-image-based, rich text formats.

[0071] The markup decoder 244 may be configured to separate tabular information from text. In some embodiments, the markup decoder 244 uses markdown language to determine information that is tabular and separate from text information. For example, the OCR manager 242 may communicate plain text returned from the OCR system 106 to the markup decoder 244. The plain text may use certain markdown symbols to indicate data as part of a table. In some embodiments, the plain text output of the OCR system 106 includes the vertical bar or pipe character, ‘|’, to mark the start of a new column within a row of the table, and the vertical bar followed by a newline character (e.g., ‘|/n’) may be used to represent a new row. The markdown language may also use hyphen characters, ‘-’, to separate a header row from a content row within a table.

[0072] The markup decoder 244 may be configured to find certain patterns in the plain text (e.g., with markdown symbols) to determine where a table begins. Regular expressions can be used with wildcards in order to identify a table in plain text (e.g., via a text-based search). For example, the regular expression ‘\|.*?\|\n\n’ may be used to find text (e.g., data, etc.) that is in a row of a table. After finding a row from a table, the markup decoder 244 may generate a new entry in the internal data storage 226 (e.g., a table entry) to store the rows of the table. For example, the rows of the table may be cut from the plain text and moved to the table entry until the next text that does not satisfy the regular expression. After this process, the plain text may have the tabular information removed (e.g., and is ready to be broken into text chunks) and the table entry may have the tabular information.

[0073] The table chunker 246 may be configured to generate table chunks from the table entry generated by the markup decoder 244. For example, a table chunk may include the entirety of the table entry. Alternatively, the table chunker 246 may be configured to generate a table chunk including a number of rows of the table entry. For example, the table chunker 246 may break the tables into 50 row chunks or 100 row chunks. The number of rows may be tailored (e.g., through configuration of the data extraction manager system 200) based on a trade-off between the ability for the retrieval process to identify the correct information to send to the LLM 108 and the amount of data that is provided to the LLM 108 and therefore the computational cost, monetary cost, and energy cost of using the LLM 108.

[0074] In some embodiments, the table chunker 246 is configured to generate a separate table chunk for the table header. It is contemplated that the table header typically has the most text in a table. In addition, the table header may have text that can be vectorized into an embedding to allow for semantic search of the tables. For example, semantic search may be performed on the headers of each table, and if a header satisfies a similarity criterion during the search, the table or a portion thereof associated with the header may be provided to the LLM 108 during processing of the prompt. The LLM 108 may be configured to understand tabular information in a certain format (e.g., the markdown provided by the OCR system 106, a JSON format wherein each cell is an object with text content or data in ASCII format, a row index, and a column index, or another suitable tabular representation). The table chunker 246 may convert (e.g., transform) the tabular representation of the OCR system 106 or the markup decoder 244 to the tabular representation used by the LLM 108.

[0075] The text chunker 248 may be configured to generate chunks of text from the plain text remaining after tables have been removed from the document. A number of text chunks may be generated from a single document. The text chunks may be of a fixed length (e.g., 500 words, 500 characters, 1000 tokens, etc.). The text chunks may be overlapping. For example, the contribution of a set of words to the semantic meaning of a chunk may be higher if the words are in the center of the chunk (e.g., because they are able to use the context of more nearby words) than at the end and therefore chunks may overlap by 50% of the length of the chunks. In some embodiments, the amount of overlap of text chunks is optimized (e.g., offline) and used to configure the data extraction manager system 200. Accuracy of the semantic search retrieval may be calculated for a set of training data (e.g., multiple documents) and used to determine a best amount of overlap or a best fixed length.

[0076] The length of the text chunk may be optimized based on an objective that includes a trade-off of the semantic search accuracy the accuracy of the data population LLM 108, and the processing time, computation cost, energy cost, or real cost used to execute the LLM 108. For example, longer text chunks may allow the LLM 108 additional background information during processing, but increase computational expense. Additionally, the accuracy of the semantic search may be poor for both chunks that are short (e.g., too little information) and chunks that are too long (e.g., so much information that the semantic meaning cannot be summarized in the vector embedding). In some embodiments, the length of the text chunk is adaptive, for example, based on the type of request, the data to be populated, the type of document, etc.

[0077] In some embodiments, the chunk tracer 250 is configured to add metadata to the text chunks and/or the table chunks. The metadata may be added to improve the document retrieval and/or provide traceability of the data that the LLM 108 extracts. For example, the chunk tracer 250 may associate a flag (or tag) with a chunk indicating the chunk is a table chunk. The flag may be a separate property in the data store (e.g., data model, ontology, etc.) used to store the chunks or the flag may be embedded in the chunk itself. The flag may be a binary flag that includes a True (1) or False (2) value next to a chunk, wherein a value of (1) indicates that the chunk is a table chunk. The flag may found using a regular expression (regexp), for example, “TABLES” may be added to table chunks. The flag may identify which chunk is a table chunk or a text chunk, based on the chunks having a similar table pattern. Adding metadata that indicates whether the table chunk allows the retrieval process to search only tabular information for certain data (e.g., that is known to be stored in tables for the particular field of use, task, etc.).

[0078] The chunk tracer 250 may be configured to store a chunk identifier, a document identifier, and/or a page identifier so that if data extraction fails or is questionable, the user is able to trace the source documentation that was used to populate a specific data element. The metadata used for tracing a chunk may be stored as part of the data store and/or the metadata may be stored in the vector store of the index (e.g., keyed based on the location within the vector). Upon failure or request by the user or the one or more UI clients 102, the chunk tracer 250 may return the document chunk identifier, a document identifier, and/or a page identifier. Additionally or alternatively, the chunk tracer 250 may be configured to retrieve the entirety of the chunk text or the table using the identifiers for viewing, verification, or reporting purposes. In some regulated industries, it may be necessary to include the reference material (e.g., as a footnote or citation) to show that the system is accurately populating the data elements and/or is unbiased.

[0079] Source documents (e.g., from the one or more data sources 104) may update or change over time. Therefore, it may be advantageous to periodically obtain documents for a specific task (e.g., data population job, etc.). However, if the documents change after some data has been extracted, traceability may be lost. To prevent loss of traceability, the chunk tracer 250 may include with the chunks a creation timestamp and an access timestamp. In some embodiments, the chunk tracer 250 may link chunks from different versions of the same document. The user may be provided with all chunks (e.g., original and updated) related to extracted information, the times the chunks were created, and the times the chunks were accessed, allowing the user to view historical information related to the information extracted and decide if the information should be updated or data extraction should be repeated.

[0080] The indexer 252 is configured to create a searchable index of the chunks generated by the table chunker 246 and/or the text chunker 248. In some embodiments, the indexer 252 generates vector embeddings of the text of the chunks. The indexer 252 may coordinate with the text embedder 110 to generate a vector embedding for a text chunk. The vector embedding may refer to a vector representation of the semantic content of the text chunk. Vectorization gives the text chunk numerical values that can be searched, with computational efficiency, for similarity (e.g., using a distance metric); thereby, text chunks with similar semantic content to a prompt can be identified for retrieval. Similar words would have similar numerical values. For example, hot and cold may have vectors pointing in different directions. The system may not find the word “cat”, but with vectors, the system will determine that lion is similar to cat or big+cat. The text embedder 110 may be trained to understand the meaning of the words (female+king=queen).

[0081] In some embodiments, the table chunks are also indexed by the indexer 252 based on semantic meaning, for example, of their header row. Additionally or alternatively, the indexer 252 may generate an index including full text for the table headers. Full text of table headers allows for more specificity in a search of tabular data. For example, specific headers may always be available in certain types of tables and can be found by keyword search and or regular expressions.

[0082] The indexer 252 may return an index including a vector data store for the vector embeddings and/or a separate index for table chunks including the full text of the table headers. The index may be stored in the internal data storage 226 until used by the retrieval augmentation process.

[0083] In some embodiments, the generative AI manager 260 includes a prompt manager 262, a semantic searcher 264, a keyword searcher 266, an LLM manager 268, a response validator 270, and response storage 272. These components may provide functionality allowing the data extraction manager system 200 to use the LLM 108 to extract specific data from the documents found by the data manager 220 and processed by the ingestion manager 240 and store that data in the data store.

[0084] The prompt manager 262 may populate prompt templates that are stored within the internal data storage 226. For example, the prompt manager 262 may be configured to insert retrieved documents (e.g., by the semantic searcher 264 and/or the keyword searcher 266) into the prompt before the prompts are sent to the LLM 108 (e.g., via the LLM manager 268). The prompt manager 262 may sequentially process prompts stored in the internal data storage 226 or the prompts may be processed in parallel, e.g., by multiple of the one or more processors 206 on the same or different computer hardware. The internal data storage 226 may store a number of prompt templates, (e.g., to extract data from the documents for each of the data elements to be populated). The prompt manager 262 may select the appropriate prompt templates for the current data population task (e.g., as provided by the user via the one or more UI clients 102).

[0085] The prompt manager 262 may use the semantic searcher 264 and the keyword searcher 266 to retrieve chunks (e.g., both table chunks and text chunks) to augment the prompt sent to the LLM 108. The semantic searcher 264 may search based on a similarity criterion or ranking using a distance metric (e.g., Euclidean distance, cosine distance) within the index of vector embeddings produced by the indexer 252. The keyword searcher 266 may search based on one or more other criteria or scores. For example, the keyword searcher 266 may search based on the number of keyword matches or the number of regular expression matches and choose the documents that have the largest number of matches. In some embodiments, the keyword searcher 266 is used for searching the table chunks, whereas the semantic searcher 264 is used to search the vector embedding index. Alternatively, both the keyword searcher 266 and the semantic searcher 264 may be used to search both table chunks and text chunks. For example, a weighted function that combines the similarity scores of the semantic searcher 264 and the matching score of the keyword searcher 266 may be used to score both table chunks and text chunks.

[0086] In some embodiments, the search criteria, score, and/or distance metric is modified based on the prompt (e.g., the particular data the prompt is requesting the LLM 108 to extract). For example, the prompt template may include search (e.g., query, retrieval) parameters such as a type of search and/or parameters for the search that are to be used while performing retrieval augmentation (e.g., while querying for relevant chunks) for a particular prompt. Advantageously, by storing the parameters for the semantic searcher 264 and/or the keyword searcher 266 with the prompt template, the retrieval augmentation can be tailored for each data element that is to be populated by the data extraction and population system 100. For example, a prompt template may indicate that only table chunks should be searched.

[0087] In some embodiments, the search performed by the semantic searcher 264 and the keyword searcher 266 is hierarchical. Multiple sets of search parameters may be associated with the prompt or the particular data to extract. The semantic searcher 264 and the keyword searcher 266 may first use a primary (e.g., first, most narrow, etc.) set of search parameters to identify relevant chunks for retrieval augmentation. If the generative AI manager 260 determines that the relevant chunks do not satisfy a retrieval criterion, the semantic searcher 264 and the keyword searcher 266 may use a secondary (e.g., second, broadening, etc.) set of search parameters. For example, the retrieval criterion may include a threshold number of chunks that must be exceeded, a threshold number of words that must be included in the chunks, chunks from at least a number of different document types, or any other desired criterion that may ensure accuracy of the LLM 108 response. In some embodiments, the semantic searcher 264 and the keyword searcher 266 continue to use increasingly broad search/retrieval parameters from the multiple sets until the retrieval criterion is achieved.

[0088] After identifying one or more relevant chunks using the semantic searcher 264 and/or the keyword searcher 266, the generative AI manager 260 may provide the one or more relevant chunks to the LLM 108 with the prompt. In some embodiments, a search reach criterion may also be used by the generative AI manager 260. The search reach parameter defines a number of chunks related (e.g., adjacent, nearby) to the one or more relevant chunks. For example, for each identified relevant chunk, the generative AI manager 260 may include all the chunks that are from the same page as the identified relevant chunk or all the chunks that satisfy the search reach criterion with the identified relevant chunk. Advantageously, in such a system the chunks generated and stored in the index can be smaller, for example, to have a concise semantic meaning for improved retrieval, and the LLM 108 is provided with contextual information adjacent to the relevant chunk to help with information extraction.

[0089] The LLM manager 268 may coordinate the interaction between the data extraction manager system 200 and the LLM 108. The LLM manager 268 may be configured to receive populated prompts to communicate to the LLM 108. The LLM manager 268 may include instructions for communicating the prompts to the LLM 108, tracking the progress in processing the prompts, causing the results to be validated by the response validator 270, and storing the response (e.g., in the internal data storage 226 and/or the response storage 272). The LLM manager 268 may post jobs (e.g., tasks, prompts, etc.) to the LLM 108 using an API provided by the LLM 108. Additionally, the LLM manager 268 may use the API to request the response to a particular prompt.

[0090] In some embodiments, the LLM manager 268 provides the prompt for information extraction, the one or more relevant chunks (e.g., found by the semantic searcher 264 and/or the keyword searcher 266), and a request for the LLM 108 to identify the used chunks that were used by the LLM 108 to extract the information. To provide traceability each chunk may be given a unique identifier (e.g., a chunk identifier, a document and page identifier, etc.) and the LLM 108 can include in its response the identifier of the chunks used during processing. The identifiers provided to the LLM 108 may be globally unique or may be unique only to the current prompt (e.g., if 23 chunks are provided to the LLM 108, the integers 1-23 may be used as unique identifiers related to the scope of that prompt). The used chunks may be stored with the response of the LLM 108 to be displayed, reported, cited, etc. for traceability and/or regulatory reasons. Additionally or alternatively, the used chunks may be stored and/or displayed responsive to an error or other undesired condition identified with the LLM 108 or the response to the current prompt.

[0091] The response validator 270 is configured to check the accuracy of the responses obtained from the LLM 108. The response validator 270 may include various guardrails to ensure that the response is appropriate. Each prompt template may store information about the expected response (e.g., type, length, acceptable range if numeric, etc.) and the response validator 270 may execute checks stored in the prompt template and/or a set of common checks that are executed against all responses. For example, the prompt template may indicate that the response should be numeric, and if the LLM 108 returns a response that is not numeric, the response validator 270 can flag the response before storing it in the response storage 272.

[0092] In response to detecting a potential error, the response validator 270 may store additional tracing information with the response from the LLM 108. Tracing information may include the chunk identifier, the page identifier, and/or the document identifier (e.g., as stored by the chunk tracer 250) from any of the chunks that were provided to the LLM 108 as part of the retrieval augmentation process. In some embodiments, the response validator 270 may store the tracing information with all responses even if no error occurs, for example, for display or regulatory purposes.

[0093] Responses may be stored in response storage 272 and/or internal data storage 226. In some embodiments, the data extraction manager system 200 stores all data in the internal data storage 226 and there is no independent data store for the data that is being populated by the data extraction and population system 100. The response storage 272 may be of the same type or a different type from the internal data storage 226. The response storage 272 may store data in magnetic hard disk, solid state drives, optical drives, RAM, and/or any other suitable storage medium. The response storage 272 may be distributed across one or more computer system, for example, communicably connected over the network 112.

[0094] The interface manager 280 may be configured to allow interaction with the data extraction manager system 200. The interface manager 280 is shown to include a client interface generator 282, an admin interface generator 284, and APIs 286. The client interface generator 282 and/or the admin interface generator 284 may provide instructions to the one or more UI clients 102 (e.g. JavaScript, Cascading Style Sheets) that instruct the one or more UI clients 102 how to generate the user interface within a client application (e.g., an internet browser, a proprietary application, etc.). In some embodiments, the interface manager 280 can provide APIs 286 that cause various functionality of the data extraction manager system 200 to be triggered. For example, the client interface generator 282 may cause the one or more UI clients 102 to generate a user interface that includes checkboxes (e.g., to select the task or the data elements to be populated) and a button to send the request to begin processing. Upon interaction with the button (e.g., a click, etc.) the user interface may use the APIs 286 to post a request to begin processing of the selected task or data elements to be populated.

[0095] The client interface generator 282 may include instructions to generate a user interface for user centric operations. The user of the data extraction and population system 100 may also be responsible for validating the data, making decisions based on the populated data, generating reports using the data, etc. and the client interface generator 282 may focus on such operations. The client interface generator 282 may provide instructions for a user interface from which particular data that is to be populated can be selected. In some embodiments, certain task includes groups of data that is to be populated. For example, a task could be “analysis number 1,” which includes a particular set of data elements that is to be populated. The client interface generator 282 may provide instructions to allow the user (e.g., via the one or more UI clients 102) to add additional data elements to the list of data that is to be populated.

[0096] The client interface generator 282 may also include instructions to allow the user to select an appropriate subject of the analysis. Example subjects include, companies, people, places, or any other subject for which it would be useful to gather large amounts of data from disparate sources. For example, a task may be to extract data to underwrite an insurance contract with a company or to collect financial information related to a publicly traded company. The client interface generator 282 may be configured to allow the user (via the generated user interface) to run a task against several subjects (e.g., for comparison). In some embodiments, the client interface generator 282 provides instructions to generate a user interface that allows the user to schedule requests for extracting the data. For example, the data extraction may be done periodically to account for changes in the data that may have occurred and/or to allow time varying data to be displayed on trendlines, bar charts, radar plots, etc. Additionally or alternatively, the client interface generator 282 may allow the user to schedule multiple subjects to be processed at different times (e.g., to avoid initializing additional cloud computing resources and being charged peak rates).

[0097] In some embodiments, instructions communicated to the one or more UI clients 102 from the client interface generator 282 include the ability to view errors that have occurred during the processing of a task. For example, errors detected by the response validator 270 may be displayed on the UI along with any tracing information that may be stored by the chunk tracer 250 with the retrieved chunks used by the LLM 108.

[0098] The admin interface generator 284 may have much of the same functionality as the client interface generator 282, for example, with additional configuration ability. For example, the instructions provided by the admin interface generator 284 may allow for the chunk size to be configured during processing. Additionally or alternatively, the admin interface generator 284 may change the parameters (e.g., weighting of a distance metric or a match metric) of the semantic searcher 264 and/or the keyword searcher 266 to adjust how the chunks are retrieved.

[0099] The enabling services 290 provide various enabling services, according to some embodiments. The enabling services 290 are shown to include a deployment manager 292, a system monitor 294, and a security manager 296. The components of the enabling services 290 together ensure smooth operation of the data extraction manager system 200 and the data extraction and population system 100.

[0100] The deployment manager 292 may be configured to allow developers to deploy new versions of the data extraction manager system 200 while maintaining the data extraction manager system 200 operational. Deployments of the data extraction manager system 200 may be container based, allowing the data extraction and population system 100 to scale the number of servers implementing the data extraction manager system 200 to scale as user demand changes. Requests for processing may be communicated to a first version of the data extraction manager system 200 while an updated second version of the data extraction manager system 200 is generated (e.g., initiated). Once the second version of the data extraction manager system 200 is fully operational, the first version may be decommissioned.

[0101] The system monitor 294 may be configured to monitor the operations of the data extraction and population system 100. For example, the system monitor 294 may monitor the request queue and/or memory usage and decide if additional computing environments should be provisioned. For example, the system monitor 294 may determine to add computing resources to the data extraction manager system 200, purchase additional processing or prioritized processing of the OCR system 106, the LLM 108, or the text embedder 110. In some embodiments, the system monitor 294 is configured to automatically provision the additional computational power. Additionally or alternatively, the system monitor 294 may generate alerts indicating that the queue is large or processing could otherwise be improved with additional resources. Such alerts may be displayed on the admin interface generator 284.

[0102] In some embodiments, the security manager 296 is configured to secure data stored within the data extraction manager system 200. The security manager 296 may maintain login information with the request identifiers that are associated with a particular user. In addition, the security manager 296 may associate various roles (e.g., user, admin, developer) with a login.

[0103] The security manager 296 may include a filtering tool that is remote from the end user and provides customizable filtering features to each end user. The filtering tool may provide customizable filtering by filtering access to the data. The filtering tool may identify data or accounts that communicate with the server and may associate a request for content with the individual account. The system may include a filter on a local computer and a filter on a server. The filtering tool may identify information or accounts that communicate with the server and associate a request for content with the individual account. The system may include a filter on a local computer and a filter on a server.

[0104] FIG. 3 shows a swimlane diagram 400 illustrating certain operations within a method for data extraction and population and indicating the components or systems that perform the steps, according to some embodiments. The first swimlane is labeled “client device” and may refer steps that are performed by a user of the data extraction and population system 100, for example, using the one or more UI clients 102. The second swimlane is labeled “data extraction manager” and may refer to steps that are performed by the data extraction manager system 200. The third swimlane is labeled ‘external systems” and may represent steps that are performed by the OCR system 106, the LLM 108, or the text embedder 110. In general, the flow of the swimlane diagram 400 is from top to bottom. However, some steps can be performed in different orders and/or in parallel.

[0105] The client device may initiate request to begin data ingestion for data sources related to a subject (e.g., topic, company, person, place, etc.) in step 402. A user may, from the one or more UI clients 102, select a task, one or more data elements to be populated, and/or a subject about which to populate the data. The user interface may activate one of the APIs 286 of the interface manager 280, causing the data extraction manager system 200 to begin processing the request. The data extraction manager system 200 may gather data from internal and external systems (e.g., the one or more data sources 104) in a step 404. For example, data may be gathered using the data scraper 224 as described herein. The external systems (e.g., in this case the OCR system 106) may perform OCR on image-based documents to return a response payload with tables indicated by markdown language in operation 406. For example, some of the gathered documents may be image-based (e.g., a PDF) that require conversion to plain text, while other documents may be already text based (e.g., from a website, etc.). The OCR system 106 ensures that text and tables are in a machine-readable format prior to further processing.

[0106] The data extraction manager system 200 may separate the response payload into a first portion having the one or more tables and a second portion having the document text in the step 408. In some embodiments, the step 408 is performed by the markup decoder 244. The markdown provided by the OCR system 106 may use symbols to represent a tabular structure (e.g., the vertical bar or pipe character, ‘|’ may indicate the start of a table row and a new column within that row). The markup decoder 244 may search for certain patterns in the plain text (e.g., with markdown symbols) to determine where a table begins. In some embodiments, a text-based search or regular expressions can be used with wildcards in order to identify a table in plain text. For example, regular expression ‘\|.*?\|\n\n’ may be used to find text (e.g., data, etc.) that is part of a table. After finding a row from a table, the portion of the table may be moved into another entry of the data store (e.g., the internal data storage 226). After this process, the plain text (e.g., the first portion of the response payload) may have the tabular information removed, and the second portion of the response payload may have only the tabular information.

[0107] One or more table chunks from the first portion of the response payload and one or more text chunks from the second portion of the response payload are formed in step 410. For example, the table chunker 246 and the text chunker 248 may be used to generate table chunks and/or text chunks as described herein. Step 410 may include generating the table chunks that include the whole table, or a number of rows or columns of the table. Text chunks may include a number of characters, words, or tokens (e.g., 2000 characters, 500 words, 1000 tokens, etc.). In some embodiments, the token length is optimized based on a trade-off between the amount of information that is communicated to the LLM 108 (e.g., related to the cost, number of computations, or energy usage) and the accuracy of the result.

[0108] In some embodiments, the table chunks and text chunks are converted into a vector embedding in step 412. For example, the data extraction manager system 200 may use the text embedder 110 to generate a vector embedding of the table chunks and/or text chunks. Embedding the chunks may convert the text into a vector or array of numbers that represent the semantic meaning of the text. The table chunks and the text chunks may be converted into vector embeddings and stored in the index for semantic search during retrieval augmentation. Alternatively, only the text chunks are converted into vector embeddings, and the table chunks may be searched by text-based keyword search of the header column and/or the first row. After step 412 is performed, the ingestion process (e.g., the gathering and preparation of documents for the RAG system of the data extraction and population system 100) may be complete and the data extraction and population system 100 ready to respond to requests for data population.

[0109] In step 414 of the swimlane diagram 400, the user, by way of the one or more UI clients 102, may initiate request to perform data population. For example, the user may choose one or more data elements to populate, develop an ontology or data model, or otherwise indicate what data is to be extracted from the documents prepared in the ingestion process before initiating the request. In some embodiments, the request to begin data ingestion of step 402 and the request to perform data population of step 414 are included together, and the other components of the data extraction and population system 100 perform all steps to extract the data without user interaction.

[0110] The steps 416-426 of the swimlane diagram 400 describe how one or more data elements are extracted using a single prompt. In some embodiments, the steps 416-426 are repeated for a number of prompts to extract a number of data elements requested by the user. The steps 416-426 may be performed sequentially, in parallel, or in a combination of both sequential processing and parallel processing.

[0111] In step 416 a prompt associated with a data element to be populated may be generated. Prompt generation may be performed by the prompt manager 262 and may include selecting an appropriate template prompt for the data element from the internal data storage 226. The swimlane diagram 400 may continue with identifying relevant chunks for the prompt based on a search criterion in step 418. For example, the semantic searcher 264 and the keyword searcher 266 may generate scores indicative of the relevance for the various chunks indexed in step 412. Separating the tabular information from the text information, among other advantages, allows the table chunks and text chunks to be searched differently. For example, certain prompts may only search for table chunks by keyword, while other prompts may search based on a weighted score of both a semantic search process and a keyword search process. Step 418 may include identifying all chunks for which the generated score is exceeds a threshold (e.g., less than a threshold for a distance metric or greater than a threshold for a similarity score) or choosing a number of the highest scoring chunks. The identified chunks may be augmented with the prompt in step 420 and sent (e.g., communicated), to the LLM 108.

[0112] In some embodiments, step 422 includes processing the prompt and communicating a response including data for the data element to be populated. For example, the LLM 108 may send the response to the data extraction manager system 200. The response may be validated in step 424. Accuracy of the responses obtained from the LLM 108 may be checked by the response validator 270. Each prompt template may store information about the expected response (e.g., type, length, acceptable range if numeric, etc.) which may used to determine if the response is appropriate for the type of data requested by the response. For example, in step 424, if a result is expected to be numeric, it is possible to check the semantic meaning of the response and determine if it is a number. Errors, for example, no response and/or data flagged in step 426 may be subjected to additional processing. For example, the identifier of the chunks identified in step 418 or the document and page of the source information for the chunk may be stored with the prompt so the user can trace the reason for the response and validate the data or note the reason for the error and populate the data manually.

[0113] After validation in step 424, the data of the response may be stored in an data store associated with the data element to be populated. For example, the data may be stored as a key value pair where the data element is the key, and the value is the response from the LLM 108 generated in step 422. Stored data may be delivered to a user interface and may be viewed by the user in step 428. In the event of an error, the user may adjust prompt format, and/or fill in missing data using chunk traceability in step 428.

[0114] FIGS. 4-6 show various flows of operations representing various aspects of the present disclosure. Each of the flows of operation may illustrate all or a portion of the process of extracting data using a large language model with retrieval augmentation, according to some embodiments. FIGS. 4-6 may emphasize various aspects of some embodiments and therefore some steps (e.g., operations) may be omitted from the flow of operations, the flow of operations may start after some steps have been completed, may end assuming some operations are performed after completing the flow of operations. In particular, FIG. 4 is related to improvements to both data extraction using a large language model and document retrieval by appropriate processing of both tabular and textual data within a RAG framework; FIGS. 5A-C are related to improvements to accuracy by allowing parameters of the retrieval process to be associated with a particular prompt (e.g., query parameters are associated with a prompt or request to extract particular information, a data element, etc.); and FIG. 6 is related to providing traceability to source documentation within the RAG framework, allowing a user to see exactly where information is sourced.

[0115] FIG. 4 shows a flow of operations 500 for coordinating data extraction and population, according to some embodiments. The flow of operations, for example, may be performed by the data extraction manager system 200 of the data extraction and population system 100. The flow of operations 500 may include receiving a response payload that includes document text of the document and one or more tables of the document represented using markdown language in operation 502. The response payload may be generated from an optical character recognition tool (e.g., the OCR system 106). The data extraction manager system 200 may receive from the OCR system 106 a response payload with tables inline with the text using a markdown language. For example, the first appearance of the markdown symbol indicates the start (or top) of a table and a second appearance of the same markdown symbol indicates the end (or bottom) of the table. The markdown symbols may also indicate a first (e.g., left) side of the table and a second (e.g., right) side of the table. Markdown symbols (e.g., within text) may provide characteristics of the table. The markdown system may provide information to the system, so the system may render the table. For example, the vertical bar or pipe character, ‘|’, may be used to mark the start of a new column within a row of the table, and the vertical bar followed by a newline character (e.g., ‘|/n’) may be used to represent a new row. The markdown language may also use hyphen characters, ‘-’, to separate a header row from a content row within a table. When analyzing the position of each cell, the system may consider each cell as having a single row of text, regardless of the number of lines of text in each cell. Additionally or alternatively, the response payload from the OCR system 106 may use JSON to indicate the location of the tabular data. A component of the data extraction manager system 200, for example, the OCR manager 242, may convert JSON into a format in which the tables are represented by markdown symbols, which can be received by the processors for further processing during later operations of the flow of operations 500.

[0116] The flow of operations 500 may include separating, using the markdown language, the response payload into a first portion having the one or more tables and a second portion having the document text in operation 504. The operation 504 may be performed by the markup decoder 244. During operation 504 certain patterns in the plain text may be found (e.g., with markdown symbols) to determine where a table begins. For example, the regular expression ‘\|.*?\|\n\n’ may be used to find text (e.g., data, etc.) that is in a row of a table. After finding a row from a table, the markup decoder 244 may generate a new entry (e.g., a location to store the first portion of the response payload having the tabular data) in the internal data storage 226 (e.g., a table entry) to store the rows of the table. The rows of the table may be cut from the plain text and moved to the table entry until the next text that does not satisfy the regular expression. After this process, the plain text (e.g., the second portion) may have the tabular information removed (e.g., and be ready to be broken into text chunks) and the table entry or first portion may have the tabular information.

[0117] The flow of operations 500 may include forming, by the one or more processors using a chunking methodology, one or more table chunks from the first portion of the response payload and one or more text chunks from the second portion of the response payload in operation 506. The operation 506 may be performed by the table chunker 246 and text chunker 248 as described with reference to those components of the data extraction manager system 200. For example, the table chunks may include a fixed or adaptive number of rows, the entire table, etc. and the text chunks may include a fixed or adaptive number of characters, words, etc.

[0118] The flow of operations 500 may include generating, by the one or more processors, an index for the one or more table chunks and the one or more text chunks in operation 508. Generating the index may include converting the one or more table chunks and the one or more text chunks into vector text embeddings using a text embedding model. For example, the operation 508 may be performed by the indexer 252. The indexer 252 may coordinate with the text embedder 110 to generate a vector embedding for a text chunk. Vectorization gives the text chunk numerical values that can be searched, with computational efficiency, for similarity (e.g., using a distance metric); thereby, text chunks with similar semantic content to a prompt can be identified for retrieval. By generating vector embeddings of the text chunks and/or the table chunks, an index may be created for which chunks can be searched (e.g., queried for retrieval) based on their similarity to a prompt for data extraction.

[0119] In some embodiments, the flow of operations 500 includes associating a document identifier and a page identifier associated with table chunks and text chunks. The chunk identifier, document identifier, and/or page identifier may be stored with the chunk. Advantageously, the retrieved chunks (e.g., the sources used by the LLM 108 during prompt processing) may be cited for regulatory reasons, in the scenario of an erroneous response, or a response that the user of the system finds questionable.

[0120] The flow of operations 500 may include identifying a relevant table chunk of the one or more table chunks or a relevant text chunk of the one or more text chunks based on a search criterion related to a prompt for a large language model in operation 512. Identifying a relevant table chunk or a relevant text chunk may include performing a semantic search (e.g., using a distance metric to compare an embedding of the prompt to an embedding of the chunk in the index), a keyword search (e.g., by counting a number of keyword or phrase matches), or a combination of both a semantic search and a keyword search. For example, the operation 512 may be performed by the generative AI manager 260 using the semantic searcher 264 and/or the keyword searcher 266 as described herein.

[0121] The flow of operations 500 may include sending (e.g., communicating, transmitting, etc.) the prompt and the relevant table chunk or the relevant text chunk to a large language model in operation 514. For example, the operation 514 may be performed by the LLM manager 268. The prompt may include a request for extracting a data element from the documents (e.g., that have been converted to text chunks and table chunks). The LLM 108 may generate a response to the prompt that includes the data element. The flow of operations 500 may include storing a response from the large language model to the prompt and the relevant table chunk or the relevant text chunk in the data store in operation 516. For example, the data element may be populated in the data store with the information from the response. In some embodiments, a request for the LLM 108 to identify the chunks used during data extraction is also provided with (e.g., as part of) the prompt. The LLM 108 may return the identifiers of the used chunks. The used chunks and/or the text or tables thereof may be displayed or reported with the extracted information. Providing the user access to the information used by the LLM 108 may allow inaccuracies and/or hallucinations by the LLM to be detected, traced, and analyzed for root cause.

[0122] FIGS. 5A-C are related to improvements to accuracy by allowing parameters of the retrieval process to be associated with a particular prompt (e.g., query parameters associated with a prompt or request to extract particular information, a data element, etc.). FIG. 5A shows a flow of operations 520 for retrieval augmentation according to retrieval parameters associated with a prompt (e.g., a request to extract particular information from one or more source documents or a request to populate particular information within a data store). The flow of operations 520 may include acquiring, by the one or more processors, an extraction prompt configured to cause a large language model to extract requested data from retrieved chunks of the one or more chunks in operation 522. The prompts and/or prompt templates may include various additional data associated with the prompt. For example, an expected data type for the extracted information may be associated with the prompt. Additionally or alternatively, one or more retrieval parameters may be associated with the prompt. In some embodiments, the retrieval parameters are used to specify specific filters, techniques, etc. for searching a RAG index. Each prompt (e.g., request to extract different information) may retrieve relevant chunks in a specific (e.g., unique, tailored, custom) manner by way of different retrieval or search parameters. For example, the operation 522 may be performed by obtaining the current prompt from the internal data storage 226 by the prompt manager 262.

[0123] The flow of operations 520 may include identifying, by the one or more processors, one or more relevant chunks according to retrieval parameters associated with the extraction prompt, the one or more relevant chunks identified from an index of one or more chunks from one or more documents, the index including vector text embeddings of the one or more chunks in operation 524. The operation 524 may be performed by the generative AI manager 260 using the semantic searcher 264 and or the keyword searcher 266. Different retrieval parameters may be used to tailor the identification of chunks for extraction of particular information. For example, a chunk type designation, a document type designation, a search type designation, regular expressions, a weighted hybrid search, and/or a search reach criterion may be used independently or in combination to customize a search. In some embodiments, more than one set of retrieval parameters is provided in a hierarchy. Subsequent sets of retrieval parameters may broaden the search criteria and be used if the relevant chunks found using the first set of retrieval parameters does not satisfy a retrieval criterion (e.g., number of chunks identified, etc.).

[0124] A chunk type designation may be used to specify if the relevant chunks (e.g., retrieved chunks or chunks provided to the LLM) are to be retrieved from table chunks, text chunks, or any other type of chunk that is referenced in the index, or a combination thereof. A document type designation may be used to specify the type of document from which the relevant chunks should originate. For example, each chunk may have an associated source document type property stored with the index. During the search (e.g., as part of the query), chunks may be filtered based on the document type. A search type designation may be used to specify if the search is to be performed using a semantic search (e.g., comparing the vector embeddings of the chunks), a keyword search, or a combination of the two search types. In some embodiments, if both semantic search and keyword search are to be used together the retrieval parameters may include weighting parameters describing how to combine the results of the keyword search and the semantic search so that an overall relevance score can be used to rank the chunks and/or compare to a threshold to determine the relevant chunks.

[0125] After one or more relevant chunks are identified, those relevant chunks may be provided to the LLM 108 with the prompt. In some embodiments, a search reach criterion is also be used to provide additional chunks related to the one or more relevant chunks. The search reach parameter defines a number of chunks related (e.g., adjacent, nearby) to the one or more relevant chunks. For example, for each identified relevant chunk, the generative AI manager 260 may include all the chunks that are from the same page as the identified relevant chunk or all the chunks that satisfy the search reach criterion with the identified relevant chunk. Advantageously, in such a system the chunks generated and stored in the index can be smaller, for example, to have a concise semantic meaning for improved retrieval, and the LLM 108 is provided with contextual information adjacent to the relevant chunk to help with information extraction.

[0126] The flow of operations 520 may include sending, by the one or more processors, the prompt and the one or more relevant chunks to a large language model in operation 526. For example, the operation 526 may be performed by the LLM manager 268. Advantageously, the high degree of specificity provided by the retrieval parameters (e.g., while executing a query) will reduce the number of computations necessary to complete the search and retrieve the relevant documents for the LLM 108, provide information to the LLM 108 with increased relevance, and may reduce the amount of data that is sent over the network to the LLM 108. The flow of operations 520 may include storing a response from the large language model to the extraction prompt and the one or more relevant chunks in operation 528. For example, data may be stored in internal data storage 226 allowing a user of the data extraction manager system 200 access to the extracted information (e.g., data elements, properties of an ontology, etc.) for viewing, report generation, etc.

[0127] FIG. 5B shows detailed operations included in some embodiments of the operation 524. For example, more than one set of retrieval parameters may be associated with a prompt or data to extract. A hierarchical list of retrieval parameters may be used to iteratively broaden the search until the relevant chunks satisfy a retrieval criterion (e.g., identified more than a threshold number of chunks, etc.). The operation 524 may include identifying, by the one or more processors, the one or more relevant chunks according to a first set of retrieval parameters in operation 530.

[0128] In some embodiments, the operation 524 includes determining, by the one or more processors, whether the one or more relevant chunks satisfy a retrieval criterion in operation 532. If the data scraper 224 was not able to find many documents from the one or more data sources 104 a small number of chunks or no chunks may be identified in operation 530. If no chunks are provided to the LLM 108 the LLM 108 may be unable to extract the requested information. The retrieval criterion may be based on a number of chunks determined to provide consistently accurate responses from the LLM 108. If the retrieval criterion is satisfied at block 534, the flow may continue to sending the one or more relevant chunks to the large language model (e.g., in operation 526 of the flow of operations 520). If the retrieval criterion is not satisfied at block 534, a second set of retrieval parameters may be used, potentially to identify more relevant chunks and satisfy the retrieval criterion in operation 536. The operations 532-536 may continue with broadening retrieval parameters until the retrieval criterion is satisfied. During the second and subsequent identification steps, it is contemplated that the search may be performed relative to the previous search for computational efficiency. For example, if the second search adds table chunks to a search that previously included only text chunks, it is not necessary to search the text chunks again with the same retrieval parameters.

[0129] FIG. 5C shows a flow diagram for the operation 524 in more detail, according to some embodiments. In some embodiments, the retrieval parameters may include a chunk type designation. The chunk type designation may be used to cause filtering, by the one or more processors, of one or more table chunks having tabular data from one or more text chunks having text data according to a chunk type designation in an operation 538. The chunk type designation may indicate that one or more relevant chunks are to be retrieved from the one or more table chunks, the one or more text chunks, or both the one or more table chunks and the one or more text chunks. Operation 538 may reduce the number of candidate chunks that are provided to the LLM 108 (e.g., if the chunk type designation specifies only table chunks or only text chunks). In some embodiments, the retrieval parameters may include a document type designation. The document type designation may be used to cause filtering of one or more chunks according to a document type designation in an operation 540. The document type designation may indicate one or more document types from which the chunks are to originate, thereby reducing the number of candidate chunks that may be provided to the LLM 108 (e.g., if the document type designation does not indicate all document types).

[0130] In operation 542, the remaining candidate chunks may be searched according to a search type designation indicating the one or more relevant chunks are to be searched using a semantic search, a keyword search, or both the semantic search and the keyword search. Performing a semantic search may include generating, by the one or more processors, distance metrics between the vector text embeddings and a vector text embedding of the extraction prompt in operation 544, and performing a keyword search may include generating, by the one or more processors, keyword scores between the one or more chunks and a keyword associated with the extraction prompt in operation 546. For example, a keyword score may be equal to a number of keyword matches or a function thereof. Additionally or alternatively, regular expressions can be used during a keyword search. In some embodiments, weighting parameters are provided as part of the retrieval parameters. The weighting parameters may be used to define a weighted function of the keyword scores and the distance metrics of the candidate chunks by which to rank or select the relevant chunks. For example, the operation 524 may include comparing, by the one or more processors, a weighted function of the keyword scores and the distance metrics of the one or more chunks according to weighting parameters in an operation 548.

[0131] In some embodiments, a search reach criterion is also be used to provide additional chunks related to the one or more relevant chunks as shown in operation 550. The operation 524 may include identifying, by the one or more processors, one or more reached chunks that satisfy a search reach criterion with a relevant chunk. The search reach criterion may define a number of chunks related (e.g., adjacent, nearby) to the one or more relevant chunks that are to be provided to the LLM 108. For example, for each identified relevant chunk, the generative AI manager 260 may include all the chunks that are from the same page as the identified relevant chunk or all the chunks that satisfy the search reach criterion with the identified relevant chunk. Advantageously, in such a system the chunks generated and stored in the index can be smaller, for example, to have a concise semantic meaning for improved retrieval, and the LLM 108 is provided with contextual information adjacent to the relevant chunk to help with information extraction.

[0132] FIG. 6 shows a flow of operations 560 related to providing traceability to source documentation within the RAG framework, according to some embodiments. The flow of operations 560 may include generating, by one or more processors, a plurality of chunks from a document in operation 562. The plurality of chunks may include table chunks, text chunks, or any other type of chunk suitable for a data extraction process. For example, the plurality of chunks may be generated by the table chunker 246 and the text chunker 248.

[0133] In some embodiments, the flow of operations 560 includes associating, by the one or more processors, (i) a document identifier for the document and a page identifier or (ii) a chunk identifier for each chunk of the plurality of chunks in operation 564. The document identifier and page identifier or the chunk identifier allow source content of the chunk to be retrieved under certain scenarios (e.g., responsive to an error, during report generation, etc.). The document identifier and page identifier or the chunk identifier may be associated with a chunk by storing the information in a database with the chunk. For example, the data model for a chunk may include properties for storing the document identifier, page identifier, and/or chunk identifier. The operation 564 may be performed by the chunk tracer 250 during the data ingestion process.

[0134] Source documents (e.g., from the one or more data sources 104) may update or change over time. Therefore, it may be advantageous to periodically obtain documents for a specific task (e.g., data population job, etc.). However, if the documents change after some data has been extracted, traceability may be lost. In some embodiments, the flow of operations 560 includes maintaining, by the one or more processors, a usage history and/or a version history for each chunk of the plurality of chunks in operation 566. For example, each time a document changes, new chunks may be created, and the new chunks may store each revision of their respective information or new chunks may be created. Using the revision history and usage history, it may be possible to provide the date and the content of a document that was used to extract the information, or if new chunks are generated when a document changes, the old chunks may be stored (e.g., for traceability), but decommissioned (e.g., no longer searched for retrieval purposes). In addition, the usage history of chunks or the number of times a chunk has been used (e.g., usage counts) may be displayed on a UI to determine which of the one or more data sources 104 are often used for information extraction. For example, the usage history may allow one to optimize the one or more data sources 104, potentially eliminating subscriptions to less useful of the one or more data sources 104. The chunk tracer 250 may perform the operation 566.

[0135] In some embodiments, the flow of operations 560 includes identifying, by the one or more processors, one or more relevant chunks from the plurality of chunks based on a search criterion related to a prompt for a large language model, wherein the prompt includes a request to extract particular information using the one or more relevant chunks in operation 568 (e.g., as performed by the semantic searcher 264 and/or the keyword searcher 266). The one or more relevant chunks may be combined with a prompt for the LLM 108 to extract particular information from the chunks (and therefore from the source documents). The flow of operations 560 may include recording, by the one or more processors, a timestamp for each chunk used by the large language model in operation 570. As the one or more chunks are identified for retrieval a timestamp may be associated with the chunk (e.g., stored with the chunk) indicating when the relevant chunk was chosen for retrieval. In some embodiments, the timestamps allow traceability by comparing the timestamp a chunk was used to the version history of the chunk.

[0136] The flow of operations 560 may include transmitting a prompt to the large language model in operation 572. The prompt may include a request to extract particular information using the one or more relevant chunks and the prompt may also include the one or more relevant chunks. In some embodiments, the prompt may also include request for the large language model to identify used chunks of the one or more relevant chunks used to extract the particular information. To provide traceability each chunk may be given a unique identifier (e.g., a chunk identifier, a document and page identifier, etc.) and the LLM 108 can include its response the identifier of the chunks used during processing. The identifiers provided to the LLM 108 may be globally unique or may be unique only to the current prompt (e.g., if 23 chunks are provided to the LLM 108, the integers 1-23 may be used as unique identifiers related to the scope of that prompt). In some embodiments, the prompt may also include a request for the LLM 108 to report any errors encountered by the LLM 108.

[0137] The flow of operations 560 may include storing the particular information from a response to the prompt from a large language model with (i) the document identifiers associated with the one or more used chunks and the page identifiers associated with the one or more used chunks or (ii) the chunk identifiers for the one or more used chunks in operation 574. The document identifiers, page identifiers, and/or chunk identifiers may be used to provide an association between the extracted, particular information and the source documentation. The association may be used to provide traceability between the data elements populated with the particular information and the source documentation, allowing for error correction and citation generation in user interfaces and or generated reports.

[0138] The flow of operations 560 may include generating, by the one or more processors, a user interface including the particular information and/or a citation to the document generated from (i) the document identifiers associated with the one or more used chunks and the page identifiers associated with the one or more used chunks or (ii) the chunk identifiers for the one or more used chunks in operation 576. For example, the interface manager 280 may create such a display to allow a user to view the extracted information with the source information (e.g., to allow for human-in-the-loop validation). The flow of operations 560 may also include generating, by the one or more processors, a citation list based on the document identifiers and page identifiers associated with the one or more used chunks in operation 578. A citation list may be used at the end of a report, presentation, or other such document that may require information sources to be cited. The citation list may also include each extracted, particular information with the citation to the source information (e.g., for regulatory purposes). Providing the user access to the information used by the LLM 108 may allow inaccuracies and/or hallucinations by the LLM to be detected, traced, and analyzed for root cause.

[0139] The detailed description of various embodiments herein makes reference to the accompanying drawings and pictures, which show various embodiments by way of illustration. While these various embodiments are described in sufficient detail to enable those skilled in the art to practice the disclosure, it should be understood that other embodiments may be realized and that logical and mechanical changes may be made without departing from the spirit and scope of the disclosure. Thus, the detailed description herein is presented for purposes of illustration only and not for purposes of limitation.

[0140] In some embodiments, the one or more LLMs 108 of data extraction and population system 100 include one or more multi-modal language models (MMLMs). The one or more MMLMs may be designed to process and/or integrate information from various modalities of input (e.g., text, images, audio, video, etc.). In some embodiments, the input layer of an MMLM includes a channel for each available modality. For example, there may be an audio channel and an image channel. The image channel may also support text represented visually in the document (e.g., on a page, etc.). The MMLMs may encode the different modalities into a common format that can be processed by one or more hidden layers within the one or more MMLMs. For example, an MMLM can include convolutional layers for imaged-based data and/or transformer layers or other attention mechanisms to process textual data. The one or more MMLMs can also include layers that combine (e.g., fuse, integrate, etc.) information across different input modes to generate an output. The output may include similar modalities as the input data. For example, the output may include text, images, audio, video, and/or other relevant formats based on the task and/or the prompt to the one or more MMLMs.

[0141] The one or more MMLMs may be configured to use the image-based input modality to better understand context of any text on the page. For example, image-based input to the one or more MMLMs may allow the one or more MMLMs to understand the flow (e.g., reading order) of the text within a document. The image-based input may also allow the one or more MMLMs to recognize relationships between figures and/or tables and text within a document. The image based one or more MMLMs may be configured to segment various areas of the document or a page within the document based on relationships between the text, figures, and/or other visual cues. For example, the one or more MMLMs may distinguish handwritten characters from typeset. In some embodiments, the one or more MMLMs are configured to accept input in a specific format or of a specific file type. The data extraction manager system 200 may convert a document to the accepted file type prior to sending the document to the one or more MMLMs. For example, a PDF may be converted to a portable network graphic (PNG) prior to communication to the one or more MMLMs. Additionally or alternatively, cloud servers or other devices of the one or more LLMs 108 (e.g., that perform processing and/or generate outputs) may include pre-processing that converts several different file types to the file type required by the one or more MMLMs.

[0142] In some embodiments, the documents processed by the data extraction and population system 100 include forms, applications, surveys, etc. for which the document or portion thereof (e.g., page, section, etc.) includes a request for information. The document or portion thereof may also include one or more predefined responses. For example, the document or portion thereof may include multiple-choice, multiple-select, and/or ranking type questions. The one or more MMLMs may be configured to recognize the selections of predefined responses from the respondent to the request for information. For example, the one or more MMLMs may recognize circles around text, check marks, filled in boxes or bubbles, as a selection of the related text. In some embodiments, the MMLM is configured (e.g., trained, fine-tuned, etc.) to determine the portion of the text that represents the request for information (e.g., the question, survey directions, etc.) and determine the text that represents the predefined responses. The one or more MMLMs may be configured or prompted to process (e.g., consider) this information separately when generating a response.

[0143] In some embodiments, the one or more MMLMs are used during document ingestion. The data extraction manager system 200 and/or the OCR system 106 may be configured to recognize that the document includes images, figures, layouts, tables, and/or other content that may benefit from image-based processing. For example, the data extraction manager system 200 may consider a trade-off between the added cost and computations of using the one or more MMLMs against the potential for improved retrieval (and therefore extraction) accuracy if the one or more MMLMs are used. In some embodiments, the data extraction manager system 200 may request the one or more MMLMs to create a vector embedding of the document or portion thereof (e.g., page, paragraph, section, etc.). Additionally or alternatively, the data extraction manager system 200 may request the one or more MMLMs to generate a summary (e.g., a text-based summary) of the document or portion thereof. After a summary of the document or portion thereof is generated the one or more LLMs 108 may be used to create a vector embedding for the index.

[0144] FIG. 7 shows a flow of operations 600 for coordinating data extraction and population using multiple types of language models (e.g., text-based LLMs and MMLMs of the one or more LLMs 108) according to some embodiments. The flow of operations 600 shows a text-based side (e.g. on the left) and an image-based side (e.g., on the right). The path (e.g., text-based or image-based) used to traverse the flow of operations 600 may be independently chosen for ingestion and/or extraction. The path may also be independently chosen for each document, each page, each file, each task (e.g., group of data to extract), or any other appropriate level of granularity. The flow of operations 600 may be performed by the data extraction manager system 200.

[0145] The flow of operations 600 may provide several advantages. Some documents may be difficult for a text-based LLM to extract information from. Several examples of such documents are described herein. One such type of document, for example, includes selections of multiple-choice questions that are responded to by hand (e.g., with pen or pencil). The visual information included in such documents (e.g., a selection of a response, a layout, etc.) may be properly identified and used by an MMLM (e.g., of the one or more MMLMs) to aid in the extraction process. The image-based path using the MMLM may greatly improve extraction accuracy for some documents. The image-based path, however, may use significantly more computations than the text-based path, due in part to the larger number of parameters and general additional complexity associated with the MMLM. In addition, using the MMLM may increase network traffic by communicating larger image-based files. Advantageously, the flow of operations 600 provides the capability for the path chosen to be based on the type of document, the processing request, etc. allowing for the executing system to use the more costly (e.g., computationally) image-based path when necessary or when the benefit of the additional accuracy outweighs the added cost. Additionally or alternatively, if text-based extraction fails (e.g., the response validator 270 determines the response was missing, incorrect, etc.) the flow of operations 600 may proceed to executing the image-based extraction as a backup method.

[0146] Advantages are also provided during the ingestion phase. An index may be generated for chunks (e.g., portions of the document) using a text-based approach and/or using an image-based approach. Surprisingly, indexes created using the text-based approach may provide similar accuracy to indexes created using the image-based approach for many scenarios. Thus, by using the text-based path for document ingestion (e.g., indexing), computational expense may be significantly reduced while providing similar accuracy. During ingestion, the flow of operations 600 also provides the ability for certain documents to be ingested using the image-based approach (e.g., for certain document types and/or responsive to a failure or error in the text-based path).

[0147] The flow of operations 600 may also provide advantages to a system that is upgraded from a text-based only approach. By executing the flow of operations 600, a system for which many documents have already been ingested may obtain the advantages of using one or more MMLMs without generating a new retrieval index. Instead, documents may be retrieved using text-based chunks, but an image associated with the text-based chunks may be provided to the one or more MMLMs. In some scenarios, data ingestion may have a very long processing time and re-embedding data (generating a new index) may have a high cost and/or be time consuming, especially if performed using the one or more MMLMs in the image-based ingestion path.

[0148] The flow of operations 600 may include receiving at least one document from internal and/or external systems in operation 602. For example, the data manager 220 may receive (e.g., obtain, acquire, get, etc.) a document from the one or more data sources 104. The document may be of any of the types described herein. For example, the document may be a file, record, report, article, form, data, application, questionnaire, etc.). The document may include text, tables, columns, rows, charts, graphics, images, and/or other content. The document may be image-based, include text encoded for computer readability (e.g., plain text), and/or a combination of image-based and plain text.

[0149] In some embodiments, the flow of operations 600 includes a decision 604 to determine if the desired ingestion type is text-based or image-based. For example, the ingestion manager 240 may determine the desired ingestion type. The desired ingestion type may be based on various criteria. In some embodiments, the ingestion type is based on the document type (e.g., image-based or text-based, file type, purpose of the document, etc.). Additionally or alternatively, the ingestion type may be based on the one or more data sources 104 from which the document was obtained. For documents indicating text-based ingestion, the flow of operations 600 may proceed to operations 606 and 608. For documents indicating image-based ingestion, the flow of operations 600 may proceed to operations 610 and 612.

[0150] The flow of operations 600 may include providing the document to the OCR and receiving the response payload including document text in the operation 606. For example, the OCR manager 242 may communicate the document to the OCR system 106 and receive the response payload from the OCR system 106. In some embodiments, the response payload may include document text and table text (e.g., using a markup language as described herein). Other indications and/or markups may be provided by the OCR system 106. For example, the payload may include an indication of handwritten characters and/or typeset. Additionally or alternatively, the payload may include an indication of the text layout and/or where figures occur within the text.

[0151] Text-based ingestion in the flow of operations 600 may include generating one or more chunks from the document text and storing a mapping to a corresponding portion of the document associated with the one or more chunks in operation 608. Chunks may refer to segments of the document text that was returned from the OCR system 106. Chunks may also include tabular data, for example, using a markup language. In some embodiments, the tabular data is separated from the document text. For example, each table may be stored in a corresponding single chunk or a number of chunks. The operation 608 may include dividing the document text into chunks of a fixed length (e.g., 500 characters, 100 words, 4 sentences, etc.). In some embodiments, the fixed length may vary by an amount to complete a portion of the text of a coarser granularity. For example, if the fixed length is 500 characters for a chunk, the operation 608 may choose a larger number of characters to complete the word with the 500^th character or choose a smaller number of characters, thus not including the word that would have the 500^th character. The decision may be fixed (e.g., the operation 608 may always choose a smaller number of characters) or the decision may be based on the particular situation for the current chunk being processed (e.g., it may choose the smaller or larger number of characters based on which would cause the resulting chunk to be closest to a 500 character target).

[0152] The operation 608 may also include storing a mapping between a corresponding portion of the document for the one or more chunks. For example, a document identifier and/or a page identifier may be associated with each chunk. The mapping may map a chunk to a specific portion of the document that included the chunk. The portion of the document may, for example, be a page, a section, a paragraph, a line, or any other appropriate division of the document that may be retrieved based on a chunk in other operations of the flow of operations 600. In some embodiments, the operation 608 stores a mapping between a chunk and a page of the document that included the chunk. During retrieval, the mapping may be used to retrieve the page having a relevant chunk, for example, to provide to an MMLM (e.g., of the one or more MMLMs). The mapping between chunk and a specific portion of the document can also be used as a reference, for example, to refer back to source material. The reference for any relevant chunk (e.g., a chunk provided to a language model) or any used chunk (e.g., a chunk identified by the language model as being of particular pertinence to the extraction of the data) can be displayed in the user interface. References may be displayed responsive to certain interactions with the user interface (e.g., attempting to edit an extracted value) or responsive to errors during the ingestion and/or extraction processes.

[0153] Text-based document ingestion has previously been described with the flows of operations 500 (e.g., operations 502-510) in FIG. 4 and the swimlane diagram 400 in steps (402-410) in FIG. 3. Such operations and/or other similar operations described herein (e.g., those performed by the ingestion manager 240) may replace similar operations within the flow of operations 600.

[0154] At any operation of the text-based data ingestion path (e.g., the operations 606 and 608) any error may occur (represented by error block 605). If it is determined that the error may be avoided by performing image-based processing, the flow of operations may switch paths to the image-based document ingestion including operations 610 and 612. The error represented by the error block 605 may occur in other operations of the flow of operations 600. For example, if indexing fails and the document has not undergone image-based processing, the flow of operations 600 may continue to the operation 610. Embedding failures may be detected by the text embedder 110. The indexer 252 and/or another component of the data extraction manager system 200 may request the text embedder 110 to output a comprehension score (e.g., a coherence score, etc.) to indicate whether the output of the OCR had a coherent and understandable semantic meaning. For example, a low coherence score may indicate the text from different sections of the document, from figures, etc. was included in a chunk and image-based ingestion may provide an improvement. In some embodiments, error detection is performed by the component of the data extraction manager system 200 that is executing the current operation. For example, error detection within operations 606, 608, and 614, may be performed by a corresponding instruction set of the ingestion manager 240.

[0155] For documents indicating image-based ingestion, the flow of operations 600 may proceed to operations 610 and 612. Image-based ingestion may include separating the document into one or more portions in the operation 610. A portion may refer to a page, a section, an area, and/or any portion of a document that can be individually provided to the one or more MMLMs. The flow of operations 600 may include prompting a multi-modal language model (MMLM) to summarize each of the one or more portions of the document during the operation 612. For example, the generative AI manager 260 may communicate a request for summarization and a portion of the document to the one or more MMLMs sequentially until each of the portions have been summarized. In some embodiments, the portion of the document is converted to an file type accepted by the one or more MMLMs (e.g., an image-based file type such as a PNG) prior to being sent for summarization. In some embodiments, the summaries received from the MMLM are stored as chunks so that downstream processing of the flow of operations 600 can be performed by the same process (e.g., using the same instructions, etc.) regardless of whether a document or portion thereof was processed using the image-based path or the text-based path.

[0156] Similar to text-based content ingestion, references (e.g., mappings, etc.) between the one or more portions of the document and a corresponding location within the source content can be stored. For example, a page, a section, a paragraph, a line, or any other appropriate division of the document or combination thereof sufficient information retrieve a specific portion of the document that may have been found relevant during search and/or specifically called out as used by the language model. In some embodiments, the references are provided to the language model with the relevant portions during extraction. For example, each relevant portion may be provided with a title, file name, etc. either in metadata provided with the relevant portion or on the relevant portion (e.g., added to an upper margin) that includes the reference (e.g., document identifier or name and page number). During extraction a request may be provided to the language model to identify (e.g., by the reference) the portion or portions of the document that were used to extract the information.

[0157] In some embodiments, the one or more MMLMs are not used to perform data ingestion. Using the one or more MMLMs may cause additional computations to be performed, for example, because of the larger network structure. The text-based path may be used for document ingestion in such embodiments.

[0158] The flow of operations 600 may include generating an index for the one or more chunks or the one or more portions of the document by converting the one or more chunks or summaries into vector text embeddings using a text embedding model in the operation 614. Index generation may be performed by the indexer 252. The operation 614 may include generating for each chunk a vector embedding for the text of the chunk. The vector embedding may represent the semantic meaning of the chunk. For example, the vector embedding may be generated by averaging a vector embedding for each word of the chunk. In some embodiments, context and word order may be considered when generating the vector embedding for a chunk. For example, the operation 614 may execute a network model using a transformer-based architecture to generate the embedding. Generating an index has previously been described with reference to the operation 508. The operation 508 and other similar operations described herein (e.g., those performed by the ingestion manager 240 including the indexer 252) may replace similar operations within the flow of operations 600.

[0159] The operations 602-614 of the flow of operations 600 may describe document ingestion. After documents have been ingested, the extraction portion of the flow of operations 600 may be used to extract data from the documents (e.g., that may have been converted to chunks during the ingestion process). Extraction may begin with retrieving relevant chunks and/or portions of the documents. Extraction may also be performed with either of two paths (e.g., a text-based path and an image-based path).

[0160] The flow of operations 600 may include identifying a relevant chunk or a relevant portion of the document based on a search criterion related to a prompt for a language model. In some embodiments, a prompt that is to be sent to an LM (e.g., the one or more LLMs 108 and/or the one or more MMLMs) is converted into a vector embedding. For example, the same embedding model used to generate the index during the operation 614 may be used during the operation 616. After the prompt has been embedded, the vector embedding of the prompt may be compared to the vector embeddings of the index (e.g., for the chunks and/or the portions of the document that were ingested). The chunks and/or portions of the document having embeddings that satisfy a matching criterion with the prompt embedding or those that have the highest matching score (e.g., the lowest distance metric) to the prompt embedding may be identified as relevant and used in later processing. Additionally or alternatively, the keywords (e.g., from the prompt, etc.) may be used to identify relevant chunks and/or portions of the document. For example, keyword and/or regular expression searches may be performed on the chunks and/or the summaries of the portions of the document. Those chunks (and/or summaries) having the highest keyword frequency or having a keyword frequency above a threshold may be identified as relevant in operation 616. Identifying relevant chunks has previously been described with reference to operations 512 of FIG. 4, the operation 524 of FIGS. 5A-C, and operation 568 in FIG. 6. Such operations and other similar operations described herein (e.g., those performed by the semantic searcher 264 and/or the keyword searcher 266) may replace similar operations within the flow of operations 600.

[0161] In some embodiments, the flow of operations 600 includes a decision 618 to determine if the desired extraction type is text-based or image-based. For example, the generative AI manager 260 may determine the operational path to perform data extraction. The desired extraction type may be based on various criteria. In some embodiments, the extraction type is based on the document type (e.g., image-based or text-based, file type, purpose of the document, etc.). Additionally or alternatively, the extraction type may be based on the one or more data sources 104 from which the document was obtained. In some embodiments, the extraction type is based on the chunk that is identified as relevant in the operation 616. For example, the ingestion manager 240 may label each chunk during document ingestion to indicate whether the chunk should be processed using the text-based path or the image-based path. For chunks or portions of the document indicating text-based extraction, the flow of operations 600 may proceed to operations 620 and 622. For documents indicating image-based ingestion, the flow of operations 600 may proceed to operations 624 and 626.

[0162] Text-based extraction in the flow of operations 600 may include retrieving (e.g., getting, obtaining, etc.) the relevant chunk (e.g., identified for retrieval in the operation 616) in the operation 620. In some embodiments, additional chunks associated with a same portion of the document as the relevant chunk are also retrieved in the operation 620. The additional chunks may be identified and/or retrieved using the mapping from the operation 608. For example, the generative AI manager 260 and/or the prompt manager 262 may perform the operation 620. In some embodiments, chunk identification and retrieval is performed in one step, for example, if the vector embedding is stored with the chunk.

[0163] Text-based extraction may also include prompting an LLM (e.g., the one or more LLMs 108) with a request to extract particular information from the relevant chunk and the additional chunks in the operation 622. Similar operations and have previously been described herein (e.g., those performed by the prompt manager 262 and/or the LM manager 268 and in the operations 514 or 526) and may replace the operation 622 within the flow of operations 600.

[0164] At any operation of the text-based data extraction path (e.g., the operations 620 and 622) any error may occur (represented by error block 619). If it is determined that the error may be avoided by performing image-based processing, the flow of operations may switch paths to the image-based data extraction including operations 624 and 626. Extraction failures may be detected by the text response validator 270 as described herein.

[0165] Image-based extraction in the flow of operations 600 may include retrieving (e.g., getting, obtaining, etc. to be passed to an LM) the relevant portion of the document identified in the operation 616 or a portion of the document corresponding to the relevant chunk (e.g., identified for retrieval in the operation 620) in the operation 624. The portion of the document corresponding to the relevant chunk may be used, for example, if the document having the relevant chunk was ingested using the text-based process. The portion of the document corresponding to the relevant chunk may be retrieved using the mapping. For example, the generative AI manager 260 and/or the prompt manager 262 may also perform the operation 620. The portion of the document retrieved may be appropriate for image-based extraction. The flow of operations 600 may also include prompting an MMLM with a request to extract particular information from the portion of the document retrieved in operation 626. For example, the prompt manager 262 and/or LM manager 268 may prompt the one or more MMLMs. In some embodiments, the portion of the document is converted to a file type accepted by the one or more MMLMs (e.g., an image-based file type such as a PNG). In some embodiments, both the image-based portion of the document (e.g., page) and the relevant chunk (e.g., text extracted from the document) are provided to the one or more MMLMs. The one or more MMLMs, for example, may allow simultaneous input (e.g., by the same prompt) by two modalities or a first prompt may request that the MMLM store the text from the relevant chunk for consideration when responding to a second prompt that also includes the prompt to extract information and the image-based portion of the document (e.g., a second modality) associated with the relevant chunk.

[0166] Performing extraction using the image-based path (e.g., operations 624 and 626) is advantageous because it allows the information to be extracted using context including location of the text, figures, images, and other visual information. In some embodiments, the documents processed by the flow of operations 600 include forms, applications, surveys, etc. for which the document or portion thereof (e.g., page, section, etc.) includes a request for information. The document or portion thereof may also include one or more predefined responses. For example, the document or portion thereof may include multiple-choice, multiple-select, and/or ranking type questions. The one or more MMLMs may be configured to recognize the selections of predefined responses from the respondent to the request for information. For example, the one or more MMLMs may recognize circles around text, check marks, filled in boxes or bubbles, as a selection of the related text. In some embodiments, the MMLM is configured (e.g., trained, fine-tuned, etc.) to determine the portion of the text that represents the request for information (e.g., the question, survey directions, etc.) and determine the text that represents the predefined responses. The one or more MMLMs may be configured or prompted to process (e.g., consider) this information separately when generating a response.

[0167] In some embodiments, the operation 626 includes prompting the MMLM to determine if a response was provided to the request for information in the document. If the MMLM determines that no response was provided, the flow of operations may generate a new request (e.g., an email, webform, etc.) for the respondent. The request may include the request for information and/or the request may be a reminder or an indication that no response was provided. Similar processing may be performed if the response is not appropriate of communicated using an incorrect method (e.g., circling text rather than filling in a bubble, etc.). For example, the operation 626 may also include prompting the MMLM to determine if an appropriate response was provided. The operation 626 may include generating a chain-of-thought prompt, first asking the MMLM to determine if a response was provided and, if a response was provided, asking the MMLM if the response was provided in an appropriate manner. After a new response is obtained from the respondent, the new response can be ingested (e.g., operations 602-614) and the extraction process (e.g., prompt) may be run again (e.g., the operations 616-626). The index and chunks for the new document (filled in request for information) may replace those created during ingestion of the incorrect or incomplete document.

[0168] In some embodiments, a whole page or other portion of a document is provided to the language model at operation 626. The chunks (even if ingestion was performed using the text-based approach) may not be provided; therefore, the language model may not be able to identify the chunks that were used extract the information. The operation 626 may include prompting the MMLM with a request to identify the portion of the document that was used to extract the information. The MMLM may identify the portion based on a reference associated with the portion during the operation 610 and provided to the language model during the operation 626 (e.g., as metadata for the portion or on the image of the portion of the document such as in a margin, etc.).

[0169] In some embodiments, the flow of operations 600 includes storing the result from the LLM or the MMLM of the prompt. For example, the LM manager 268 may receive a response from the one or more LLMs 108 or the one or more MMLMs, response validator 270 may validate the response, and/or the generative AI manager 260 may store the result in the response storage 272.

[0170] It is contemplated that systems performing the flow of operations 600 (e.g., the data extraction manager system 200) are not required to implement both text-based and image-based paths for both ingestion and extraction. At least one benefit of the disclosure herein is that the additional accuracy provided by image-based extraction using the one or more MMLMs can be provided without significant computational expense incurred if all documents were ingested using the image-based approach. For example, this benefit may be provided by first storing a mapping between a chunk and a portion of a document and retrieving the original document or portion thereof (e.g., image, PDF, etc.) during extraction to be provided to the one or more MMLMs. Thus, only portions of documents considered relevant are processed by the one or more MMLMs. Additionally or alternatively, a data extraction process that already implements a fully text-based approach may not require that documents be re-ingested or the index of chunks be rebuilt.

[0171] In some embodiments, the text-based ingestion of the documents has already been performed. To save development time and overall system complexity, image-based ingestion may not be implemented by the data extraction manager system 200 and/or be available when performing the flow of operations 600. The configuration of the flow of operations 600 allows for modular approaches. For example, the decision 604 may always direct the flow of operations 600 to text-based ingestion if image-based ingestion has not yet been implemented in the data extraction manager system 200. It is noted that the decision 604 may not perform an active step. If image-based ingestion is not implemented, operational flow may automatically flow from operation 602 to the operation 606.

[0172] The detailed description of various embodiments herein makes reference to the accompanying drawings and pictures, which show various embodiments by way of illustration. While these various embodiments are described in sufficient detail to enable those skilled in the art to practice the disclosure, it should be understood that other embodiments may be realized and that logical and mechanical changes may be made without departing from the spirit and scope of the disclosure. Thus, the detailed description herein is presented for purposes of illustration only and not for purposes of limitation.

[0173] FIG. 8 shows relationships between objects found within the data extraction and population system 100 according to some embodiments. Extraction may be performed on submission content, for example, to populate information (e.g., data fields, etc.) for an application, registrations, or other similar types of forms. During the extraction process, data extraction and population system 100 may be provided with submission content, including one or more documents 310 from the one or more data sources 104, such as email databases and other servers. Each document is shown to have one or more portions 312 (e.g., portion A 312a and portion B 312b).

[0174] In some embodiments, the one or more portions 312 are used to generate one or more chunks 320. For example, the portion A 312a may have a one-to-one correspondence with the chunk for portion A 322a. Similarly, the portion B 312b may have a one-to-one correspondence with the chunk for portion B 322b. The one or more portions 312 (and therefore the one or more chunks 320) can be generated by the data extraction manager system 200 using any of the techniques or methods described herein (e.g., based on a number of words, a number of characters, type of content such as tabular content or textual content, with or without overlap, etc.).

[0175] The one or more chunks 320 need not store the entirety of the text they represent. For example, each of the one or more chunks 320 may refer back to the one or more documents 310 and the portion of the one or more portions 312 to which they correspond. In FIG. 8, a chunk of the one or more chunks 320 (e.g., chunk for portion A 322a) is shown to include an embedding 323, a reference 324, a chunk type 325, and a creation timestamp 326. The embedding 323 represents an entry in a searchable index for the chunk and is created during ingestion as described above. The reference 324 includes any data that sufficiently describes the source content that the chunk represents, such that it can be retrieved (e.g., during retrieval of relevant chunks for extraction, for traceability of errors or questionable extraction results, for display of source content, generation of a citation or list thereof, etc.).

[0176] The chunk type 325 may refer to one or more types of source content from which the chunk was generated. For example, a chunk type may refer to the structure of the source content (e.g., text, table, plot, picture), the modality of the source content (e.g., text, image, video, audio, etc.), the file type of the source content (e.g., PDF, PNG, Microsoft Excel, etc.), the purpose or intent of the source content (e.g., instruction manual, research article, advertisement, survey response), the language and encoding used in the source content (e.g., English, Mandarin, UTF-8, ASCII), or any other classification of the source content that may be useful in debugging errors or describing results. The chunk type 325 can be implemented as a number of tags (e.g., from a closed set of options), allowing the chunk to have more than one type. Alternatively, the chunk type 325 can be implemented as an enumeration wherein each chunk has one type (or one type from each of a number of type classes).

[0177] The creation timestamp 326 may refer to the time the chunk was created from the submission content. The creation timestamp 326 can be generated and associated with the chunk at the time of chunk creation. The creation timestamp 326 facilitates tracing versions of a document that was used to create a chunk. For example, the creation timestamp 326 of a chunk can be compared to the timestamps associated with the version history of the document from which the chunk was generated. Comparison allows the data extraction manager system 200 to determine which of the versions of the document were used to make a chunk. Similarly, the creation timestamp 326 can be displayed on a user interface to allow for human-in-the-loop validation.

[0178] The chunk and the corresponding portion of the source content may be used to generate a citation 330 for that chunk and/or corresponding portion. The citation 330 may include a particular format of data from the chunk and/or the corresponding portion. In some embodiments, multiple citations 330 are generated, and each citation 330 may be used for a particular purpose. For example, a first format for a citation 330 may be used for report generation and a second format for a citation 330 may be used for the user interface. In some embodiments, the citation 330 may refer to the reference 324 itself. For example, the citation 330 may include a document identifier and a page number, which can be displayed in the user interface or within a generated report for the chunk. As another example, the citation may include the document title and section. The reference 324 can be used to refer back to the appropriate document and portion and determine information used for the citation. As yet another example, the citation 330 may include the entirety of the portion (or, based on the size of the portions, a sub-portion thereof). In some embodiments, the form of the citation 330 depends on the chunk from which it is generated. For example, the form of the citation 330 may depend on the size of the chunk, the chunk type 325, the embedding 323, or any other information included in the chunk or the corresponding portion.

[0179] The data extraction manager system 200 may store an extraction result 340 for each of the data fields that are to be or were extracted for a submission. An extraction result 340 is shown to include the data field name 342, the extracted value 344, relevant chunks 346, used chunks 348, and language model reasoning 349. Results from the extraction process for the data field indicated by the data field name 342 are stored in the extraction result 340.

[0180] The extracted value 344 may be configured as a property of the extraction result 340. The extracted value 344 can store the value for the data field associated with the extraction result 340 (e.g., by the data field name 342) that was received from the language model (e.g., in response to the extraction prompt or the request to extract a value for the data field).

[0181] In some embodiments, the extraction result 340 stores the chunks identified as relevant during the search or augmentation process in a relevant chunks property 346 (e.g., field, etc.). The relevant chunks property 346 can store chunks identified by any of the searching processes, including those chunks identified by the keyword searcher 266 and/or the semantic searcher 264 that were provided to the language model with the extraction prompt. The extraction result 340 may also store a subset of the relevant chunks found during search that are identified by the language model as being used (e.g., of particular use) in extracting the value for the data field in the used chunks property 348. The used chunks property 348 may be populated by the response from the language model to the extraction prompt or a request to identify those chunks used to extract the value for the data field.

[0182] The extraction result 340 may also be configured to store the language model reasoning 349 used to extract the value provided in its response. The language model reasoning 349 may also be populated by the response from the language model to the extraction prompt or a request to provide its reasoning for extracting the value for the data field.

[0183] Although the objects shown in FIG. 8, the traceability features with reference to the user interface shown in FIG. 9A, and the flows of operation that support the traceability operations described herein (e.g., FIGS. 4, 6, 7, 10, and 11) are described with reference to chunking of the documents from the submission content, it is understood that similar features can be performed without chunking during ingestion. In some embodiments, the chunks are generated based on portions of a document identified by the language model as used to extract the value for the data field. The one or more chunks 320 may not cover all documents; for example, the union of the portions of the documents used to extract all of the data fields may include all of the document text; instead, the chunks can be generated on-demand as a result of the language model (or a user providing feedback in the process described below) identifying a portion of the document as useful. In some embodiments, chunks are not stored; the extraction result 340 can include some or all of the information shown as stored in the chunk for portion A 322a. For example, the extraction result 340 may store the reference 324 based on (e.g., describing, locating, etc.) the portion of the one or more documents 310 used by the language model during extraction. The data extraction manager system 200 can also generate the chunk type 325 and the creation timestamp 326 at the time of extraction. For example, the chunk type 325 may be determined automatically based on the identified portion of the one or more documents 310 (and may be more appropriately given a name such as “reference type” or similar).

[0184] FIG. 9A is an illustrative example of a user interface 700 for the data extraction manager system 200 according to some embodiments. For example, instructions for the user interface 700 may be generated by the interface manager 280 and communicated to a client device of the one or more UI clients 102 where the user interface is generated and used. The user interface 700 is shown to have a first display element 702 that includes a name of a data field extracted from the submission content and the value of the submission content. The first display element 702 may be a tabular or similarly structured listing of the ontological model of data fields extracted from the submission content. In some embodiments, the first display element 702 includes techniques to make certain data fields and/or extracted values more conspicuous. For example, data fields may be accentuated to point to potential errors, data fields that did not successfully extract a value, data fields for which a value appears to be an outlier, data fields that the language model flagged as uncertain, etc. Data fields can be accentuated by drawing a box around the field, displaying the text in bold or another color, highlighting the data field, or otherwise drawing attention to a particular value.

[0185] The configuration of user interface 700 provides several advantages when compared to other interfaces for information extraction systems. The user interface 700 facilitates human-in-the-loop auditing and traceability. The user interface 700 includes interactive elements that allow users to rapidly trace the extracted data field back to precise supporting evidence in source content. Traceability enhances transparency and can provide a user-friendly workflow to eliminate many document interactions associated with auditing and review. In addition to evidence in the source content, justifications for the extracted value, including language model reasoning, are incorporated into a single location further supporting review of automatically extracted data. The user interface 700 also supports collection of feedback for future improvement of the extraction system. When extracted values are updated a training sample is automatically generated with the human entered and validated ground truth value. Training samples collected after updates to extracted values capture instances where the extraction system previously failed, providing valuable training data for addressing more the difficult extraction problems encountered. The feedback system is seamless and requires no additional input from the user beyond a user's standard workflow.

[0186] In some embodiments, the first display element 702 is an interactive user interface element. The first display element 702 may facilitate updating the values that were extracted for the data field and providing feedback related to whether the extraction was correct. In some embodiments, interacting with the first display element 702 causes another view to be generated, allowing for the value to be updated, feedback to be provided, and viewing of additional details related to the data field and its extracted value. Various interactions can be used to trigger the generation of the view. The view may be generated in response to a click interaction, a hover interaction, or a button press interaction. The interaction may be performed directly upon the text of the data field or corresponding value, or the interaction may be performed through a button or similar interactive element that is disposed on the UI proximate to the text of the data field or corresponding value (e.g., horizontally aligned).

[0187] In some embodiments, interacting with a data field or value thereof in the first display element 702 causes a second view (e.g., data field view 710) that provides additional details to the 710 and the ability to update the value of the data field. Advantageously, the data field view 710 includes much of the information a human would use in order to make decisions related to whether or not an extracted value is correct. In addition, the data field view 710 allows a user to provide corrections and/or feedback when the value is incorrect. The data field view 710 is shown to include a title or field display 712, a correction text entry area 714, a current value display 716, a source display 718, a source type display 720, a source type entry area 722, an additional feedback entry area 724, and update reason entry area 726. The title or field display 712, current value display 716, source display 718, and source type display 720 may be read-only elements and display current values for the extracted value and/or the reference or citation from one or more chunks used by the language model to extract the current value. In some embodiments, the data field view 710 is a pop-up window or pop-up view that appears in response to the interaction.

[0188] In some embodiments, the data field view 710 facilitates review and correction of extracted values. For example, the interface elements 712 and 716-720 can be used to review information currently stored for the data field (e.g., the extracted value, etc.), whereas the interface elements 714, and 722-726 facilitate user update of the current values.

[0189] The data field view 710 is shown to include a display area including the current value display 716, the source display 718, and the source type display 720. The interface elements 716-720 may be configured to display the latest (e.g., current) value for the data field, the source, and the source type, respectively. The current value may be a value extracted by the data extraction manager system 200 or an updated value entered by the user. In some embodiments, the source display 718 displays the reference for the extracted value or a citation generated from the reference. In some embodiments, the source display 718 includes a hyperlink to the source document, to a view of language model reasoning, or is otherwise interactive. Responsive to interacting with the source display 718, reasoning link 728, the data field view 710, or another element of the user interface 700, the user interface 700 may generate an extraction reasoning view 750.

[0190] The source type display 720 may display the type of source from which the current value was extracted. For example, a source type may refer to a chunk type or be the same as the chunk type for a chunk used by the language model to extract the data value. In some embodiments, chunks are not used and the source type may refer to the source content or portion thereof used to extract the value. The source type may refer to the structure of the source content (e.g., text, table, plot, picture), the modality of the source content (e.g., text, image, video, audio, etc.), the file type of the source content (e.g., PDF, PNG, Microsoft Excel, etc.), the purpose or intent of the source content (e.g., instruction manual, research article, advertisement, survey response), the language and encoding used in the source content (e.g., English, Mandarin, UTF-8, ASCII), or any other classification of the source content that may be useful in debugging errors or describing results.

[0191] The extraction reasoning view 750 is shown to include relevant display information for the extracted value of the data field by including the title or field display 712, the current value display 716, source display 718, and the source type display 720. The extraction reasoning view 750 can facilitate review and/or audit of extracted data by providing information related to the reasoning the language model used to extract the value for the data field. For example, the extraction reasoning view 750 may display source content used by the language models (e.g., the used chunks, etc.), highlight or otherwise indicate words, phrases, or sentences that were particularly relevant to the extraction, and display a text based explanation of the extraction reasoning (e.g., provided by the language model in response to a prompt). The extraction reasoning view 750 may include one or more source material displays 752. A source material display 752 displays at least a portion of the contents of the source information used to extract the value. For example, the one or more source material displays 752 may include a citation such as a quote or other text, figures, table sections, etc., referred to by the reference. In some embodiments, the one or more source material displays 752 also include an indicator of particular relevance 754. The indicator of particular relevance 754 can highlight, underline, bold, or otherwise accentuate portions of the source material that were found to be of particular relevance (e.g., flagged by the language model at the request of the data extraction manager system 200). In some embodiments, the extraction reasoning view 750 is a pop-up window or pop-up view that appears in response to the interaction with the source display 718.

[0192] As shown in FIG. 9A, the user interface 700, for example in the extraction reasoning view 750, may include a display of the language model's reasoning for the extracted value (e.g., shown as reasoning display element 756). For example, the extraction reasoning view 750 may include text explaining how the contents of the reference (e.g., shown in the one or more source material displays 752) or the information highlighted by an indicator of particular relevance 754 supports the finding of the value that was extracted. The reasoning may be provided in a natural language format. In some embodiments, the request for reasoning is provided to the language model during extraction prompting (e.g., as part of the extraction prompt or as an additional prompt following the extraction prompt).

[0193] Referring now to FIG. 9B, the user interface 700 may generate a data source view 760 responsive to interacting with the source display 718, the data field view 710, or another element of the user interface 700. The data source view 760 is shown to display the source document from which the data value for a data field was extracted. The data source view 760 is shown to include relevant display information for the extracted value of the data field by including the title or field display 712, the current value display 716, source display 718, and the source type display 720. The data source view 760 is also shown to include source link 762. The data source view 760 may be configured to display the source content in response to an interaction with the source link 762 (e.g., a click, etc.). The 760 may include more than one 762, for example, if the language models used content (e.g., chunks, portions, etc.) from more than one document to extract the information.

[0194] As shown in FIG. 9B, the data source view 760 may include an integrated document viewer 766 with window 768. The integrated document viewer 766 may be configured to display documents of any type (e.g., PDF, tables, email, images, drawings, etc.). In some embodiments, the data source view 760 also includes information related to reasoning the language model used for extraction. For example, the integrated document viewer 766 may highlight or otherwise indicate portions of the document that were used by the language model to perform extraction. The integrated document viewer 766 may highlight used chunks and/or phrases, sentences, etc. identified by the language model as having particular relevance or importance.

[0195] The data source view 760 may include download button 764. The download button 764 may be configured to initiate a download of a source document. For example, the download button may be horizontally aligned with the corresponding source. Upon interaction with the download button 764, the user interface 700 may initiate the download, facilitating viewing documents in the users preferred application.

[0196] The data field view 710 is shown to include correction text entry area 714. The correction text entry area 714 is an editable text field that, upon clicking or another suitable interaction, allows the user to enter a new value for the data field. The correction text entry area 714, for example, can be used if no value was extracted, the user believes the current value is incorrect (e.g., shown in the current value display 716), or the current value is otherwise questionable (e.g., flagged by the language model, etc.). Upon entry of a new value into the correction text entry area 714, the user can use the “OK” button to send the new value entered to the underlying database of extracted results. Using the OK button can activate an API that uploads the new value entered from the client device displaying the user interface 700 to the data extraction manager system 200 storing the extraction results. In some embodiments, the original extracted value (and other information that is updated by way of the user interface 700) is stored in a separate property or field for the extraction result, thereby allowing for changes to be reverted. The entire history of changes for the value of a data field can be logged for future review and audit. The change history may include the previous value, the new value, the source of the new value, a timestamp for when the change was made, and/or any other information that may be useful for later review.

[0197] In some embodiments, the data field view 710 includes instructions that cause an updated value in the title or field display 712 to be committed (e.g., uploaded, etc.) by way of the “OK” button only after certain additional information related to the update has also been entered. For example, the OK button may remain greyed out or otherwise inactive until a new source type is selected by way of source type entry area 722, feedback is entered such as possible reasons the language mode may have extracted the wrong information in the additional feedback entry area 724, a reason for the update to the value is provided in the update reason entry area 726 and/or a new reference has been created.

[0198] After a new value is entered, additional information may be accentuated (e.g., highlighted, subjected to a change of color, etc.) indicating the need to enter additional information. The source type entry area 722 is shown as a drop-down list, allowing the user to quickly select acceptable source types (e.g., that are included with the submission type for this extraction result). The additional feedback entry area 724 allows the user to enter or additional information related to the change and the update reason entry area 726 allows the user to provide a reason why the change was necessary. Entry of additional information in the additional feedback entry area 724 may be optional (e.g., depending on the data field, the configuration of the data extraction manager system 200, etc.).

[0199] In some embodiments, selection of a source type by way of the source type entry area 722 (or other interaction with an element of the data field view 710) causes an additional view to open to facilitate user selection of a reference (e.g., document and page, etc.) to the source of the value entered. The additional view may include a listing of the documents of the type selected for the source. The user can select and view a document from the list. The user interface also includes instructions that allow the user to indicate a particular portion of the document from which the newly entered value can be found. For example, the user interface 700 may be configured to allow a user to highlight text, images, portions of a table, etc., in the additional view. Upon leaving the additional view, a reference may be generated by the user interface 700 based on the user document selection and the portion indicated as having the newly entered value. The reference generated can be uploaded from the user interface 700 on a client device of the one or more UI clients 102 to the data extraction manager system 200. Saving the updated source data, reference, and/or human reasoning for the updated value may facilitate generating reports or auditing data in the same way, whether information is extracted automatically or is corrected by a human-in-the-loop process.

[0200] Feedback provided by a user of the user interface 700 can be used to improve the extraction process of the data extraction manager system 200. For example, keywords used to search for relevant chunks, a prompt used to rank relevant chunks, and/or the extraction prompt can be improved using information provided via the user interface 700 (and the data field view 710). The extraction process may be improved by human-in-the-loop updates of the extraction process, or an automated improvement procedure may be executed on a periodic or on-demand basis. For example, the automated improvement procedure for the extraction process may execute every week, every month, or each time feedback was provided for at least 30 data fields, etc. The values entered in the correction text entry area 714 may be stored as ground truth values for a training sample (e.g., in training data) using the submission content and data field currently displayed by the user interface 700 and communicated to the appropriate system that performs the automated improvement procedure. In some embodiments, the feedback provided via the user interface 700 is used by the systems and/or methods for generating extractors (e.g., keywords for retrieval, a ranking prompt to order the relevance of retrieved chunks, and/or an extraction prompt).

[0201] In some embodiments, the feedback provided allows the determination of a cause of the incorrect data extraction. For example, the user interface 700, the data extraction manager system 200, or the system performing the improvement procedure can identify the cause by comparing the user-provided reference for an updated value for the data field to the relevant chunks property 346 and the used chunks property 348 of the related extraction result 340. The user-provided reference being found in the relevant chunks property 346 and the used chunks property 348 may be indicative of an issue with the extraction prompt. Alternatively, if the user-provided reference is absent from the relevant chunks property 346 and the used chunks property 348, a problem with the search may be indicated (e.g., an issue with keywords, a ranking prompt, etc.).

[0202] FIGS. 9A and 9B show illustrative views of the user interface 700 according to some embodiments. The layout of the interface elements and the description of the interactions between the interface elements should not be considered limiting and any layout including some or all of the elements described herein should be considered within the scope of the disclosure. For example, the download button 764 may be included in the data field view 710 and/or the extraction reasoning view 750, the extraction reasoning view 750 and the data source view 760 may be combined in a single view, and/or the extraction reasoning view 750 may be generated in response to an interaction with the 702 (e.g., rather than the data field view 710 as described).

[0203] FIGS. 10 and 11 are flows of operations that support generation of the user interface 700. In addition, previously described flows of operation in FIGS. 4, 6, and 7 also support traceability, review, and feedback features provided by the user interface 700.

[0204] FIG. 10 shows flow of operations 800 for generating a user interface with traceability features for a data extraction system according to some embodiments. For example, the flow of operations 800 may be performed by the data extraction manager system 200 to generate the objects and populate the properties shown in FIG. 8 to facilitate generation of the views and interface elements of the user interface 700.

[0205] In some embodiments, the flow of operations 800 begins with receiving submission content and a data field to be extracted from the content in operation 802. The submission content can be sourced from various sources (e.g., the one or more data sources 104) as described herein, including web servers, emails, articles, surveys, applications, etc. The data field may be one of several data fields related to a form, an application, or similar data set that is to be populated. For example, the data fields may be a portion of a data model (e.g., an ontology, etc.) used to represent the application or form. In operation 804, content relevant to the data field is identified. For example, the semantic searcher 264 and/or the keyword searcher 266 are used to identify relevant content that can be provided to the language model for data extraction. In some embodiments, the relevant information is further pruned by a ranking prompt or similar ranking procedure, wherein the most relevant content is provided to the language model (e.g., top five, ten, etc. portions of the document or chunks).

[0206] The operations 806-812 describe a series of prompts, a chain-of-thought prompt, or a prompt with several requests. Though described as four individual operations, it is understood that these may be combined into one operation of generating a single prompt (e.g., for the chain-of-thought prompt or the prompt with several requests). Similarly, several of the operations may not be performed. For example, if a chunking procedure is performed, it may not be necessary to request that the language model identify a reference for the portions of content used, as the chunks may already include that information.

[0207] The flow of operations 800 is shown to include prompting a language model with a request to extract a value for the data field from the relevant content in operation 806. The operation 806 may refer to the extraction prompt and can include the name of the data field to be extracted, a description of the data field, and/or any other information or metadata that provides additional context to the language model to facilitate identifying, calculating, or otherwise extracting a value for the data field from the relevant content.

[0208] The flow of operations 800 may include prompting a language model with a request to identify one or more portions of the relevant content used to extract the value in the operation 808. The form of the request to identify the relevant content may depend on the configuration of the data extraction manager system 200, the data field, and/or the language model used to extract the information. In some embodiments, the chunks are provided to the language model that are associated with reference to a particular portion of the submission content. The operation 808 may include requesting that the language model identify the chunks used to extract the value for the data field. The chunks may be of appropriate size to provide specific information to the user in the user interface 700, allowing them to quickly determine whether the correct information was extracted. No additional information may be needed from the language model other than the identification of the used chunks.

[0209] In addition to identifying the used chunks (e.g., for larger chunks such as whole pages or chunks over a threshold size that may lead to slower human review of the extracted values), operation 808 may include generating a request for the language model to identify a smaller portion of the chunk or content provided to the language model, thereby allowing the user interface 700 to provide specific information to the user for faster review. Identifying smaller portions within the relevant content can also be appropriate when the ingested documents are not chunked and multiple pages of a document are provided to the language model. For example, documents ingested and/or extracted using the image-based branch of the flow of operations 600 may not be subjected to chunking, or extraction performed by language models with large context windows may also not be subjected to chunking.

[0210] When additional (e.g., more fine-grained location information than the chunk alone) is requested in the operation 808, the flow of operations also provides for requesting the language model to identify a corresponding reference for each respective portion of the one or more portions of content used to extract the value in the operation 810. The prompt generated in the operation 810 may request identification of specific location information such as the document identifier (e.g., file name, etc.), page number, paragraph number, column number, line number, word number, and/or character number. The operation 810 may also include determining whether spatial coordinates on the page are more appropriate (e.g., for images, tables, etc.) than word- or text-based coordinates. Where determined appropriate, operation 810 may request a reference location such as the spatial coordinates of a corner of a bounding box and the size, or the spatial coordinates of two opposite corners of a bounding box, etc.

[0211] In some embodiments, the flow of operations 800 includes prompting a language model with a request to provide reasoning used to extract the value for the data field in the operation 812. The prompt may directly ask for a reasoning process with a prompt such as “explain your reasoning for selecting this value, referencing specific evidence or phrases from the text.” The prompt may elaborate on the information desired in the reasoning with a prompt such as “provide a brief justification of the extracted value, describing how you determined this value and which parts of the content influenced your choice,” or “After extracting the value, outline the steps and criteria you used in making your decision, including any contextual clues or supporting statements.” In some embodiments, the request to provide reasoning is a chain-of-thought prompt, for example, that requests multiple parts of the reasoning process as in “Step by step, explain your reasoning process. Describe how you evaluated the relevance of the content, specify how particularly relevant content (e.g., phrases, etc.) were used to justify the extracted value, and conclude with the justification for your final value.”

[0212] The responses to the requests from the language model are stored in the operation 814. For example, the information can be stored in the extraction result 340. In some embodiments, the data provided in the responses to the requests are stored in a tuple, for example, a tuple including the value extracted, the corresponding reference for each respective portion of the one or more portions of content used to extract the value, and the reasoning used to extract the value. As used herein, “tuple” refers to any data structure that allows for the contents of the tuple to be displayed, viewed, etc. together. A tuple does not necessarily refer to any specific coding or memory structure that uses the word tuple, array, list, or similar in a particular syntax. For example, a tuple may refer to a syntax-specific tuple, an array, a list, an object, a set, a collection of pointers, a set of properties of an object, etc. In the operation 816, the information stored in the tuple or pointed to by elements of the tuple is displayed. For example, the data extraction manager system 200 (e.g., by way of the interface manager 280) may cause a user interface to be generated including a name of the data field, the value extracted for the data field, the corresponding references, and/or the reasoning used to extract the value. The data extraction manager system 200 can generate instructions for the user interface 700 and transmit those instructions to a client device (e.g., of the one or more UI clients 102).

[0213] FIG. 11 shows flow of operations 900 for generating and managing an interactive user interface for the data extraction manager system 200 that facilitates traceability according to some embodiments. Though described as a flow that is partially performed by the client device displaying the user interface and executing the user interface instructions, it is understood that some steps of the flow of operations 900 also describe the instructions generated by the data extraction manager system 200 (e.g., by way of the interface manager 280) and transmitted to the client device for execution.

[0214] The flow of operations 900 is shown to include extracting, using a language model, a value for a data field from content of a submission in the operation 902. The operation 902 can include operations from any of the components and/or flows previously described for data extraction (e.g., the data extraction manager system 200, flow of operations 500, flow of operations 600, flow of operations 800, etc.). During the extraction process, the value is associated with a reference to a portion of the content used to extract the value. The reference association can be made via a listing of used chunks and/or portions of the submission content that the language model identified as having particular relevance (e.g., within a used chunk, page, etc.).

[0215] The flow of operations 900 can include generating a user interface with an interactive element including a name of the data field and a value extracted for the data field by a language model in the operation 904. For example, a tabular view such as the first display element 702 of the user interface 700 can be generated during the operation 904 that includes the name of the data field and the extracted value. The user interface may include a number of data fields extracted from the submission content. The interactive element can generate a view of the ontological model in any structured or unstructured format. In some embodiments, the operation 904 accentuates certain data fields and/or extracted values for improved conspicuity. For example, data fields may be accentuated to point to potential errors, data fields that did not successfully extract a value, data fields for which a value appears to be an outlier, data fields that the language model flagged as uncertain, etc. Data fields can be accentuated by drawing a box around the field, displaying the text in bold or another color, highlighting the data field, or otherwise drawing attention to a particular value.

[0216] Individual portions of the interactive element can have dedicated responses. For example, each of the data field names and/or extracted values can respond to an interaction by generating a new view specific to that data field. Responsive to an interaction with the interactive element, the user interface can display the reference to the portion of the content used to extract the value. For example, the user interface can generate an additional view as shown in the data field view 710. The additional view includes a reference (e.g., a file name, document identifier, etc., and a page number or other location within the document). In some embodiments, the reference itself is interactive (e.g., the reference may include a hyperlink) and facilitates further review by the user.

[0217] Responsive to an interaction with the reference, the portion of content to which the reference refers is displayed in operation 908. For example, if the user clicks the reference or another interactive element associated with the reference, another view may be generated. The extraction reasoning view 750 provided in FIG. 9A is an example of this view. The view can include source material display 752 for displaying at least a portion of the contents of the source information used to extract the value. The one or more source material displays 752 may include a citation such as a quote or other text, figures, table sections, etc., referred to by the reference. In some embodiments, the one or more source material displays 752 also include an indicator of particular relevance 754. The indicator of particular relevance 754 can highlight, underline, bold, or otherwise accentuate portions of the source material that were found to be of particular relevance (e.g., flagged by the language model in response to the request of the data extraction manager system 200).

[0218] FIG. 12 shows flow of operations 920 for generating and managing an interactive user interface for the data extraction manager system 200 that facilitates collecting feedback and generating training samples according to some embodiments. Though described as a flow that is partially performed by the client device displaying the user interface and executing the user interface instructions, it is understood that some steps of the flow of operations 920 also describe the instructions generated by the data extraction manager system 200 (e.g., by way of the interface manager 280) and transmitted to the client device for execution.

[0219] The flow of operations 920 is shown to begin similar to the flow of operations 900. For example, the flow of operations 920 includes extracting, using a language model, a value for a data field from content of a submission in operation 922 and generating a user interface with an interactive element including a name of the data field and a value extracted for the data field by a language model in operation 924.

[0220] The flow of operations 920 is shown to include displaying a text element for editing the value extracted for the data field in operation 926. In some embodiments, the text element for editing the value is generated in response to an interaction with the interactive element generated in the operation 924. For example, the data field view 710 may be generated in response to an interaction with the first display element 702. The text element provides an area (e.g., region, field, etc.) where the user can provide an updated value replacing the value that was automatically extracted in the operation 922. The user may provide an updated value when the extraction results are incorrect. Advantageously, the data extraction manager system 200, by way of the flow of operations 920, can generate training samples (e.g., training data) including at least a portion of the submission content and the updated value to facilitate real-time or near real-time learning as users make corrections to data that is extracted.

[0221] In the operation 928, an updated value may be received. For example, the user may enter a corrected value into the text element on the user interface and submit the entered value (e.g., by way of the correction text entry area 714). In some embodiments, the user can provide additional information related to the update. The user interface may facilitate entry of a new source type, a new source (e.g., a reference), a reason why the update was necessary, and/or additional feedback. The additional information may be included with the training sample to be used by an algorithm for improving the extraction process (e.g., algorithms for adjusting the RAG retrieval by improving keywords and/or ranking prompts and data extraction by improving extraction prompts). In some embodiments, the update of a value is interpreted as an indication that (i) the extracted value was incorrect and (ii) the entered value is correct and is the value that should have been extracted. A training sample may be generated using the original submission content or a portion thereof (e.g., relevant chunks, multiple documents, etc.) and the entered value.

[0222] During the operation 930, the request for the language model to extract the data field from content of the submission is adjusted using training data having the updated value for the data field and the content of the submission. Any part of the extraction process may be adjusted during the operation 930, for example, new keywords may be generated using the training data, ranking prompts may be adjusted using the training data, and/or extraction prompts may be adjusted using the training data. In some embodiments, a number of training samples (e.g., one training sample corresponding to one updated value and the corresponding submission content) are collected and used together to adjust the extraction process. For example, training samples may be collected for a period of time (e.g., a week, a month, etc.) before being used to adjust the extraction process. The adjustment may be performed periodically (e.g., every week, every month, etc.) using the new training samples that have been generated during the recent period. Alternatively, the adjustment may be performed after a certain number of training samples have been collected from user-entered updates (e.g., corrections). For example, the adjustment may be performed after 100 or 200 training samples are collected.

[0223] An embodiment of the present disclosure relates to a method for extracting particular information from a document. The method includes receiving, by one or more processors, a response payload that includes document text of the document and one or more tables of the document represented using markdown language. The response payload is generated from an optical character recognition tool. The method also includes separating, by the one or more processors using the markdown language, the response payload into a first portion having the one or more tables and a second portion having the document text. The method also includes forming, by the one or more processors using a chunking methodology, one or more table chunks from the first portion of the response payload and one or more text chunks from the second portion of the response payload. The method also includes identifying, by the one or more processors, a relevant table chunk of the one or more table chunks or a relevant text chunk of the one or more text chunks based on a search criterion related to a prompt for a large language model. The method also includes storing a response from the large language model to the prompt and the relevant table chunk or the relevant text chunk.

[0224] In some embodiments, the method also includes identifying, by the one or more processors, the first portion of the response payload having the one or more tables using a text-based search for sequences of characters used by the markdown language.

[0225] In some embodiments, the text-based search includes using regular expressions.

[0226] In some embodiments, the markdown language includes markdown symbols that separate the one or more tables from the document text.

[0227] In some embodiments, the markdown symbols indicate one or more boundaries of the one or more tables.

[0228] In some embodiments, the method also includes generating, by the one or more processors, an index for the one or more table chunks and the one or more text chunks. Generating the index includes converting the one or more table chunks and the one or more text chunks into vector text embeddings using a text embedding model. The search criterion includes a distance between the vector text embeddings and a prompt embedding of the prompt using the text embedding model.

[0229] In some embodiments, the search criterion includes a keyword search of row or column headers of the one or more table chunks.

[0230] In some embodiments, the method also includes associating, by the one or more processors, a document identifier for the document and a page identifier for a page of the document with each of the one or more table chunks and the one or more text chunks.

[0231] In some embodiments, the method also includes storing the document identifier and the page identifier associated with the relevant table chunk or the relevant text chunk with the prompt responsive to an error indicated by the large language model.

[0232] In some embodiments, the document is in a portable document format (PDF).

[0233] In some embodiments, the method also includes requesting, by the one or more processors, one or more data elements related to the particular information to be populated in a data store by sending the prompt to the large language model.

[0234] Another embodiment of the present disclosure relates to a method for preparing a document for retrieval augmentation. The method includes receiving, by one or more processors, a response payload that includes document text of the document and one or more tables of the document represented using markdown language, wherein the response payload is generated from an optical character recognition tool. The method also includes separating, by the one or more processors using the markdown language, the response payload into a first portion having the one or more tables and a second portion having the document text. The method also includes forming, by the one or more processors using a chunking methodology, one or more table chunks from the first portion of the response payload and one or more text chunks from the second portion of the response payload. The method also includes generating, by the one or more processors, an index for the one or more table chunks and the one or more text chunks. Generating the index includes converting the one or more table chunks and the one or more text chunks into vector text embeddings using a text embedding model and entries of the index that satisfy a distance criterion with a prompt for a large language model are relevant for the prompt.

[0235] In some embodiments, the method also includes identifying, by the one or more processors, the first portion of the response payload having the one or more tables using a text-based search for sequences of characters used by the markdown language.

[0236] In some embodiments, the text-based search includes using regular expressions.

[0237] In some embodiments, the markdown language includes markdown symbols that separate the one or more tables from the document text.

[0238] In some embodiments, the markdown symbols indicate one or more boundaries of the one or more tables.

[0239] In some embodiments, the method also includes identifying, by the one or more processors, a relevant table chunk of the one or more table chunks or a relevant text chunk of the one or more text chunks based on the distance criterion. The method also includes storing a result from the large language model in response to the prompt and the relevant table chunk.

[0240] In some embodiments, the method also includes associating, by the one or more processors, a document identifier for the document and a page identifier for a page of the document with each of the one or more table chunks and the one or more text chunks and storing the document identifier and the page identifier associated with the relevant table chunk or the relevant text chunk with the prompt responsive to an error indicated by the large language model.

[0241] In some embodiments, the method also includes identifying, by the one or more processors, a relevant table chunk of the one or more table chunks based on a keyword search of row or column headers of the one or more table chunks.

[0242] Another embodiment of the present disclosure relates to a system for extracting particular information from a document. The system includes one or more processors and one or more tangible, non-transitory memories configured to communicate with the one or more processors. The one or more tangible, non-transitory memories having instructions stored thereon that, in response to execution by the one or more processors, cause the one or more processors to perform operations. The operations include receiving a response payload that includes document text of the document and one or more tables of the document represented using markdown language, wherein the response payload is generated from an optical character recognition tool. The operations also include separating, using the markdown language, the response payload into a first portion having the one or more tables and a second portion having the document text. The operations also include forming, using a chunking methodology, one or more table chunks from the first portion of the response payload and one or more text chunks from the second portion of the response payload. The operations also include identifying a relevant table chunk of the one or more table chunks or a relevant text chunk of the one or more text chunks based on a search criterion related to a prompt for a large language model. The operations also include storing a response from the large language model to the prompt and the relevant table chunk or the relevant text chunk.

[0243] Another embodiment of the present disclosure relates to a method for document retrieval within retrieval augmented generation. The method includes acquiring, by one or more processors, an extraction prompt to cause a large language model to extract information from provided text. The method also includes identifying, by the one or more processors, one or more relevant chunks according to retrieval parameters associated with the extraction prompt. The one or more relevant chunks are identified from an index of one or more chunks from one or more documents, the index including vector text embeddings of the one or more chunks. The method also includes storing a response from the large language model to the extraction prompt and the one or more relevant chunks.

[0244] In some embodiments, the method also includes forming, by the one or more processors, one or more table chunks having tabular data of the one or more documents and one or more text chunks having text data of the one or more documents. The retrieval parameters include a chunk type designation indicating the one or more relevant chunks are to be retrieved from the one or more table chunks, the one or more text chunks, or both the one or more table chunks and the one or more text chunks.

[0245] In some embodiments, the method also includes receiving, by the one or more processors, a response payload that includes document text of a document of the one or more documents and one or more tables of the document represented using markdown language. The response payload is generated from an optical character recognition tool. The method also includes separating, by the one or more processors using the markdown language, the response payload into a first portion having the one or more tables and a second portion having the document text. Forming the one or more table chunks is based on the first portion and forming the one or more text chunks is based on the second portion.

[0246] In some embodiments, the retrieval parameters include a search type designation indicating the one or more relevant chunks are to be searched using a semantic search, a keyword search, or both the semantic search and the keyword search.

[0247] In some embodiments, the method also includes generating, by the one or more processors, distance metrics between the vector text embeddings and a vector text embedding of the extraction prompt and generating, by the one or more processors, keyword scores between the one or more chunks and a keyword associated by the extraction prompt. The retrieval parameters include weighting parameters for a weighted function of the keyword scores and the distance metrics and retrieving the one or more relevant chunks is based on the weighted function of the keyword scores and the distance metrics.

[0248] In some embodiments, the method also includes generating, by the one or more processors, match scores for the one or more chunks using a text-based search. The retrieval parameters include one or more regular expressions to perform the text-based search.

[0249] In some embodiments, the retrieval parameters include a document type designation indicating one or more document types from where the one or more relevant chunks are to originate.

[0250] In some embodiments, the retrieval parameters include a search reach criterion. One or more reached chunks that satisfy the search reach criterion with a relevant chunk of the one or more relevant chunks are provided to the large language model.

[0251] In some embodiments, the retrieval parameters include a hierarchy of sets of the retrieval parameters. Identifying the one or more relevant chunks according to the retrieval parameters includes identifying, by the one or more processors, the one or more relevant chunks according to a first set of retrieval parameters of the hierarchy; determining, by the one or more processors, whether the one or more relevant chunks satisfy a retrieval criterion; and responsive to determining that the one or more relevant chunks do not satisfy the retrieval criterion, identifying, by the one or more processors, the one or more relevant chunks according to a second set of retrieval parameters of the hierarchy.

[0252] Another embodiment of the present disclosure relates to a system for document retrieval within retrieval augmented generation. The system includes one or more processors and one or more memories storing instructions that, when executed by the one or more processors, cause the system to perform operations. The operations include acquiring an extraction prompt to cause a large language model to extract information from provided text. The operations also include identifying one or more relevant chunks according to retrieval parameters associated with the extraction prompt. The one or more relevant chunks are identified from an index of one or more chunks of one or more documents. The index includes vector text embeddings of the one or more chunks. The operations also include storing a response from the large language model to the extraction prompt and the one or more relevant chunks.

[0253] In some embodiments, the operations also include forming, by the one or more processors, one or more table chunks having tabular data of the one or more documents and one or more text chunks having text data of the one or more documents. The retrieval parameters include a chunk type designation indicating the one or more relevant chunks are to be retrieved from the one or more table chunks, the one or more text chunks, or both the one or more table chunks and the one or more text chunks.

[0254] In some embodiments, the operations also include receiving a response payload that includes document text of a document of the one or more documents and one or more tables of the document represented using markdown language. The response payload is generated from an optical character recognition tool. The operations also include separating, using the markdown language, the response payload into a first portion having the one or more tables and a second portion having the document text. Forming the one or more table chunks is based on the first portion and forming the one or more text chunks is based on the second portion.

[0255] In some embodiments, the retrieval parameters include a search type designation indicating the one or more relevant chunks are to be searched using a semantic search, a keyword search, or both the semantic search and the keyword search.

[0256] In some embodiments, the operations also include generating distance metrics between the vector text embeddings and a vector text embedding of the extraction prompt and generating keyword scores between the one or more chunks and a keyword associated with the extraction prompt. The retrieval parameters include weighting parameters for a weighted function of the keyword scores and the distance metrics and retrieving the one or more relevant chunks is based on the weighted function of the keyword scores and the distance metrics.

[0257] In some embodiments, the operations also include generating match scores for the one or more chunks using a text-based search, wherein the retrieval parameters include one or more regular expressions to perform the text-based search.

[0258] In some embodiments, the retrieval parameters include a document type designation indicating one or more document types from where the one or more relevant chunks are to originate.

[0259] In some embodiments, the retrieval parameters include a search reach criterion. One or more reached chunks that satisfy the search reach criterion with a relevant chunk of the one or more relevant chunks are provided to the large language model.

[0260] In some embodiments, the retrieval parameters include a hierarchy of sets of the retrieval parameters. Identifying the one or more relevant chunks according to the retrieval parameters includes identifying the one or more relevant chunks according to a first set of retrieval parameters of the hierarchy; determining whether the one or more relevant chunks satisfy a retrieval criterion; and responsive to determining that the one or more relevant chunks do not satisfy the retrieval criterion, identifying the one or more relevant chunks according to a second set of retrieval parameters of the hierarchy.

[0261] Another embodiment of the present disclosure relates to a non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to generate an index for one or more chunks of one or more documents. Generating the index includes converting the one or more chunks into vector text embeddings using a text embedding model. The instructions also cause the one or more processors to acquire an extraction prompt configured to cause a large language model to extract requested data from retrieved chunks of the one or more chunks. The instructions also cause the one or more processors to identify one or more relevant chunks according to a first set of retrieval parameters associated with the extraction prompt. The instructions also cause the one or more processors to determine whether the one or more relevant chunks satisfy a retrieval criterion. The instructions also cause the one or more processors to identify the one or more relevant chunks according to a second set of retrieval parameters associated with the extraction prompt responsive to determining that the one or more relevant chunks do not satisfy the retrieval criterion and to store a response from the large language model to the extraction prompt and the one or more relevant chunks.

[0262] In some embodiments, the second set of retrieval parameters includes a search reach criterion, wherein one or more reached chunks that satisfy the search reach criterion with a relevant chunk of the one or more relevant chunks are provided to the large language model.

[0263] An embodiment of the present disclosure relates to a method for providing traceability in large language model responses. The method includes generating, by one or more processors, a plurality of chunks from a document. The method also includes associating, by the one or more processors, (i) a document identifier for the document and a page identifier or (ii) a chunk identifier for each chunk of the plurality of chunks. The method also includes identifying, by the one or more processors, one or more relevant chunks from the plurality of chunks based on a search criterion. The method also includes transmitting, by the one or more processors to a large language model, a prompt including (i) a first request to extract particular information using the one or more relevant chunks, (ii) the one or more relevant chunks, and (iii) a second request for the large language model to identify used chunks of the one or more relevant chunks used to extract the particular information. The method also includes storing the particular information from a response to the prompt from the large language model with (i) the document identifiers associated with the used chunks and the page identifiers associated with the used chunks or (ii) the chunk identifiers for the used chunks.

[0264] In some embodiments, the method also includes generating by the one or more processors, a user interface including the particular information and a citation to the document generated from (i) the document identifiers associated with the used chunks and the page identifiers associated with the used chunks or (ii) the chunk identifiers for the used chunks.

[0265] In some embodiments, the method also includes detecting, by the one or more processors, an error condition in the response from the large language model. Generating the user interface including the particular information and the citation to the document is responsive to detecting the error condition.

[0266] In some embodiments, the method also includes generating, by the one or more processors, a report document including the particular information and a citation to the document generated (i) the document identifiers associated with the used chunks and the page identifiers associated with the used chunks or (ii) the chunk identifiers for the used chunks.

[0267] In some embodiments, the plurality of chunks includes table chunks and text chunks. The method also includes associating, by the one or more processors, a chunk type for each respective chunk of the plurality of chunks, the chunk type indicating that a respective chunk includes tabular data or text data.

[0268] In some embodiments, the method also includes generating, by the one or more processors, a citation list based on the document identifiers and page identifiers associated with the used chunks and storing the citation list with the particular information from the response.

[0269] In some embodiments, the method also includes recording a timestamp for each chunk used by the large language model and storing the timestamp with the particular information from the response.

[0270] In some embodiments, the prompt also includes a third request to output an error condition encountered by the large language model.

[0271] In some embodiments, the document is in a portable document format (PDF).

[0272] Another embodiment of the present disclosure relates to a system for maintaining traceability in document processing. The system includes one or more processors and one or more memories storing instructions that, when executed by the one or more processors, cause the system to implement an ingestion manager. The ingestion manager is configured to generate a plurality of chunks from documents received for processing, assign unique identifiers to each chunk of the plurality of chunks, and maintain a traceability database storing relationships between chunk identifiers; source document identifiers; page location identifiers; and chunk content. The instructions also cause the system to implement a retrieval manager configured to identify one or more relevant chunks from the plurality of chunks based on a search criterion and retrieve traceability information for the one or more relevant chunks. The instructions also cause the system to implement a generative artificial intelligence manager. The generative artificial intelligence manager is configured to transmit, to a large language model, a prompt including a first request to extract particular information using the one or more relevant chunks, the one or more relevant chunks, and a second request for the large language model to identify used chunks of the one or more relevant chunks used to extract the particular information, receive a response from the large language model, and store in a data store the response; the chunk identifiers of the used chunks used by the large language model to generate the response; the source document identifiers associated with the used chunks; and the page location identifiers associated with the used chunks.

[0273] In some embodiments, the generative artificial intelligence manager is further configured to receive, from the large language model, the chunk identifiers of the used chunks, wherein the used chunks are a subset of the one or more relevant chunks.

[0274] In some embodiments, the ingestion manager is further configured to separate table chunks from text chunks and maintain separate traceability records for the table chunks and the text chunks.

[0275] In some embodiments, the retrieval manager is further configured to record retrieval timestamps; track the one or more relevant chunks that are provided to the large language model; and maintain a usage history for each chunk of the plurality of chunks.

[0276] In some embodiments, the generative artificial intelligence manager is further configured to detect error conditions in the response; retrieve the chunk content from the used chunks for verification; and generate error reports including the source document identifiers associated with the used chunks and the page location identifiers associated with the used chunks.

[0277] In some embodiments, the ingestion manager is further configured to generate globally unique identifiers for each chunk of the plurality of chunks; maintain version history of each chunk of the plurality of chunks when the documents are updated; and link chunks from different versions of a same document.

[0278] In some embodiments, the generative artificial intelligence manager is further configured to generate formatted citations including the source document identifiers associated with the used chunks and the page location identifiers associated with the used chunks; and store the formatted citations with the response in the data store.

[0279] Another embodiments relates to a non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to generate a plurality of chunks from documents received for processing. The instructions also cause the one or more processors to create and store traceability records including chunk identifiers; source document identifiers; page location identifiers; chunk type designation; and chunk content, wherein the chunk type designation indicates whether a chunk of the plurality of chunks contains tabular data or text data. The instructions also cause the one or more processors to identify one or more relevant chunks from the plurality of chunks based on a search criterion. The instructions also cause the one or more processors to receive a response from a large language model to a prompt including the one or more relevant chunks. The response including used chunks of the one or more relevant chunks used by the large language model to extract particular information requested by the prompt. The instructions also cause the one or more processors to store in data store the response: the chunk identifiers of the used chunks used by the large language model to generate the response; the source document identifiers associated with the used chunks; and the page location identifiers associated with the used chunks.

[0280] In some embodiments, the instructions also cause the one or more processors to detect error conditions in the response; retrieve the chunk content from the used chunks for verification; and generate error reports including the source document identifiers associated with the used chunks and the page location identifiers associated with the used chunks.

[0281] In some embodiments, the instructions also cause the one or more processors to maintain a version history of each chunk of the plurality of chunks when the documents are updated and link chunks from different versions of a same document.

[0282] In some embodiments, the traceability records also include: creation timestamps; last access timestamps; usage counts; and error flags.

[0283] Another embodiment relates to a method for providing source content for information extracted by language models. The method includes generating, by one or more processors, a plurality of chunks from one or more documents, each chunk associated with an identifier referring to a portion of the one or more documents that corresponds to the chunk. The method also includes identifying, by the one or more processors, one or more relevant chunks from the plurality of chunks based on a search criterion. The method also includes transmitting, by the one or more processors to a language model, a prompt including (i) a first request to extract particular information from the one or more relevant chunks and (ii) a second request for the language model to identify one or more used chunks of the one or more relevant chunks used to extract the particular information. The method also includes generating, by the one or more processors, instructions for a user interface including the particular information extracted by the language model and a citation to the portion of the one or more documents from where the particular information is extracted based on the identifiers of the one or more used chunks.

[0284] In some embodiments, the identifier includes at least one of a document identifier and a page identifier or a chunk identifier.

[0285] In some embodiments, the method also includes detecting, by the one or more processors, an error condition in a response from the language model including the particular information extracted. Generating the user interface including the particular information and the citation to the portion of the one or more documents is responsive to detecting the error condition.

[0286] In some embodiments, the method also includes generating, by the one or more processors, a report document including the particular information and the citation to the portion of the one or more documents from where the particular information is extracted.

[0287] In some embodiments, the method also includes associating, by the one or more processors, a timestamp with each chunk of the plurality of chunks indicating a time the chunk was created, wherein the user interface also includes the timestamp associated with the one or more used chunks.

[0288] Another embodiment relates to a system for providing source content for information extracted by language models. The system includes one or more processing circuits configured to generate a plurality of chunks from one or more documents, where each respective chunk is associated with a corresponding reference to a portion of the one or more documents used to generate the respective chunk. The one or more processing circuits are also configured to identify one or more relevant chunks from the plurality of chunks based on a search criterion. The one or more processing circuits are also configured to transmit, to a language model, one or more prompts including (i) a first request to extract a value for a data field from the one or more relevant chunks and (ii) a second request for the language model to identify one or more used chunks of the one or more relevant chunks used to extract the value. The one or more processing circuits are also configured to generate instructions for a user interface including the value extracted and the corresponding reference for the one or more used chunks.

[0289] In some embodiments, the one or more prompts further include a third request to identify portions of the one or more used chunks used to extract the value for the data field.

[0290] In some embodiments, the one or more prompts also include a third request to provide reasoning used to extract the value for the data field and the user interface includes the reasoning.

[0291] In some embodiments, the instructions for the user interface are configured to generate a first interactive element including the value extracted and, responsive to an interaction with the first interactive element, generate a second interactive element including a text entry element for a user to enter an updated value for the data field and a reference to at least a subset of the one or more used chunks.

[0292] In some embodiments, the one or more processing circuits are also configured to display a used chunk of the subset of the one or more used chunks within the second interactive element or responsive to an interaction with the second interactive element.

[0293] In some embodiments, the one or more processing circuits are also configured to display reasoning used by the language model to extract the value for the data field within the second interactive element or responsive to an interaction with the second interactive element.

[0294] In some embodiments, the one or more processing circuits are also configured to generate a training sample including at least a portion of the one or more documents and the updated value entered by the user in the text entry element.

[0295] In some embodiments, the training sample is used to adjust a prompt or keywords used to extract values for the data field.

[0296] In some embodiments, the one or more processing circuits are also configured to identify a source type for each chunk based on a type of document for the portion of the one or more documents used to generate the chunk and the user interface includes the source type for the one or more used chunks.

[0297] Another embodiment relates to a system for providing source content for information extracted by one or more language models. The system includes one or more processing circuits configured to generate a first interactive element on a user interface, the first interactive element including a name of a data field and a value for the data field extracted by the one or more language models. The one or more processing circuits are also configured to, responsive to an interaction with the first interactive element, generate a second interactive element including a text entry element for a user to enter an updated value for the data field and a reference to at least a portion of the source content used by the one or more language models to extract the value for the data field. The reference is generated by the one or more language models in response to a request to identify the source content used by the one or more language models to extract the value from submission content provided to the one or more language models.

[0298] In some embodiments, the one or more processing circuits are also configured to display at least the portion of the source content used by the one or more language models to extract the value for the data field within the second interactive element or responsive to an interaction with the second interactive element.

[0299] In some embodiments, the one or more processing circuits are also configured to display reasoning used by the one or more language models to extract the value for the data field within the second interactive element or responsive to an interaction with the second interactive element.

[0300] In some embodiments, the one or more processing circuits are also configured to generate a training sample including at least a portion of the submission content and the updated value entered by the user in the text entry element.

[0301] In some embodiments, the training sample is used to adjust a prompt or keywords used to extract values for the data field.

[0302] In some embodiments, the one or more processing circuits are also configured to display a type of the source content.

[0303] These embodiments are illustrative only and should not be considered limiting.