Document query analysis agent construction method and system based on large language model

Abstract

The invention provides a document query analysis agent construction method and system based on a large language model, and the method comprises the steps: carrying out the processing of a target PDF document through a large language model, and obtaining a PDF content index; matching the keyword with the synonym in the PDF content index to obtain first related page content; vectorizing the retrieval problem and the description text of each page, and calculating the similarity to obtain second related page content; and based on the first related page content and the second related page content, obtaining a preliminary answer through a large language model, and then performing verification and correction to obtain a final answer of the retrieval question. According to the method, the multi-source heterogeneous information of the text needing to be analyzed is integrally described, the related content of the proposed analysis problem is retrieved in a multi-dimensional matching mode, and then answer generation and verification are carried out, so that the precision of document query analysis is improved, and the consumption of required computing resources is reduced.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and system for constructing a document query and analysis intelligent agent based on a large language model. Background Technology

[0002] Document query analysis aims to transform scattered and implicit high-value information in various unstructured documents into explicit, structured, and directly usable knowledge outcomes. The emergence of Large Language Models (LLM) has upgraded document query analysis methods from traditional "character-level operations" to "semantic-level understanding." However, due to the lack of collaborative support from external retrieval and knowledge bases, document query analysis driven solely by LLM still suffers from technical limitations such as insufficient information extraction accuracy and poor adaptability to multi-source heterogeneous documents.

[0003] Currently, mainstream LLM-based document query and analysis methods attempt to assist LLM in document understanding by introducing external knowledge bases. Lewis et al. (Lewis, Patrick, et al. "Retrieval-augmentedgeneration for knowledge-intensive NLP tasks." Advances in Neural Information Processing Systems 33 (2020): 9459-9474.) proposed a collaborative architecture of "retrieval + generation," which combines the semantic understanding and logical reasoning capabilities of LLM with the accurate retrieval capabilities of external knowledge bases, completely breaking through the technical bottleneck of document query and analysis methods that rely solely on LLM. Darren Edge et al. (Darren Edge, Ha Trinh, et al. "From Local to Global: A Graph RAG Approach to Query-Focused Summarization." 2024. arXiv:2404.16130) proposed a graph index-based retrieval enhancement generation method. By constructing an entity knowledge graph and pre-generating community summaries, it effectively addresses the limitations of traditional RAG in global corpus analysis tasks and significantly improves the comprehensiveness and diversity of responses in large-scale text processing. Junde Wu et al. ("Medical Graph RAG: Evidence-based Medical Large Language Model via Graph Retrieval-Augmented Generation." In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics. 2025: 28443-28467) proposed a graph-based enhanced generative framework for medical retrieval. Through ternary graph construction and a U-shaped retrieval mechanism, they achieved high-quality evidential responses for reliable tracing in multiple medical benchmark tests.In the area of ​​Agent-RAG, Chia-Yuan Chang et al. (Chia-Yuan Chang, Zhimeng Jiang, Vineeth Rakesh, Menghai Pan, Chin-Chia Michael Yeh, Guanchu Wang, Mingzhi Hu, Zhichao Xu, Yan Zheng, Mahashweta Das, and Na Zou. "MAIN-RAG: Multi-Agent Filtering Retrieval-Augmented Generation." In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics. 2025: 2607-2622) significantly improved answer accuracy and effectively reduced irrelevant document interference in multiple benchmark tests through multi-agent collaborative filtering and adaptive thresholding mechanisms.

[0004] While these methods have made breakthroughs in their respective directions, document query and analysis methods based on traditional RAG still have room for improvement in dynamic knowledge adaptation and global association modeling; document query and analysis methods based on multi-agent collaboration have solved some accuracy problems, but will face the limitation of large resource requirements. Summary of the Invention

[0005] In view of the above, the main objective of this invention is to propose a method and system for constructing a document query and analysis intelligent agent based on a large language model, so as to solve the above-mentioned technical problems.

[0006] This invention proposes a method for constructing a document query and analysis intelligent agent based on a large language model, the method comprising the following steps: Step 1: Obtain the target PDF document, and extract features from each page of the target PDF document based on the OCR algorithm to obtain multi-source heterogeneous information for each page; Step 2: Use a large language model to semantically describe the multi-source heterogeneous information on each page, generating descriptive text for each page; construct a PDF content index based on the descriptive text of all pages. Step 3: Use templates to standardize the query and analysis questions required for the target PDF document, so as to construct a structured retrieval question that includes ID, question type, answer format and answer unit; Step 4: Use a large language model to perform semantic intent analysis on the structured retrieval question and then refine it to obtain core keywords and generate a set of synonyms for the corresponding core keywords; match and filter the core keywords and the corresponding set of synonyms with the PDF content index to obtain the first relevant page content; Step 5: Use a pre-trained language model to vectorize and encode the structured retrieval question and the description text of each page, then calculate the similarity, and filter according to a preset similarity threshold to obtain the content of the second relevant page. Step 6: Merge and filter the content of the first and second relevant pages in sequence, and use a large language model for logical association and reasoning analysis to generate a preliminary answer; Step 7: Using a large language model and combining the answer format and answer unit in the structured retrieval question, verify and correct the preliminary answer for formal compliance, unit accuracy, and expression standardization to obtain the final answer to the retrieval question.

[0007] This invention also proposes a document query and analysis intelligent agent construction system based on a large language model. The system adopts the document query and analysis intelligent agent construction method based on a large language model as described above. The system includes a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, it can implement the method steps described above.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the limitations of traditional RAG document query and analysis methods in terms of dynamic knowledge adaptation and global association modeling, as well as the limitations of multi-agent collaborative document query and analysis methods, which, while partially improving accuracy, suffer from high resource requirements. It proposes a holistic description of the multi-source heterogeneous information in the text to be analyzed, and the retrieval of relevant content for the proposed analysis question through multi-dimensional matching. Subsequently, answers are generated and verified. This approach effectively improves the accuracy and dynamic knowledge adaptation of document query and analysis while significantly reducing computational resource consumption during processing.

[0009] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the steps of a document query and analysis intelligent agent construction method based on a large language model proposed in this invention. Figure 2This is a general framework diagram of a document query and analysis intelligent agent construction method based on a large language model proposed in this invention. Detailed Implementation

[0011] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0012] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0013] Please see Figure 1 This embodiment provides a method for constructing a document query and analysis intelligent agent based on a large language model. The method includes the following steps: Step 1: Obtain the target PDF document. Based on the OCR algorithm, extract features from the information of each page of the target PDF document to obtain multi-source heterogeneous information for each page.

[0014] In step 1, the multi-source heterogeneous information on each page is structured Markdown text containing one or more of the following: text, tables, formulas, and charts. The OCR algorithm uses Nanonet-OCR-s, ​​which can parse the input document image into structured Markdown text. This not only recognizes the text content but also deeply understands and extracts complex semantic elements in the document, including tables, formulas, signatures, and checkboxes, ensuring the completeness and positional accuracy of the extracted multi-source heterogeneous information.

[0015] Step 2: Use a large language model to semantically describe the multi-source heterogeneous information on each page, generating descriptive text for each page; construct a PDF content index based on the descriptive text of all pages.

[0016] In step 2, when generating descriptive text, the large language model will summarize the core content expressed by the multi-source heterogeneous information of each page, the key information in the page, and the logical relationships between the page content.

[0017] Step 3: Use templates to standardize the query and analysis questions required for the target PDF document, so as to construct a structured retrieval question that includes ID, question type, answer format and answer unit.

[0018] In step 3, the template is a structured and standardized template, and the ID of the constructed structured retrieval question is unique; the question types are clearly divided into two categories: "query questions" and "analysis questions". "Query questions" refer to questions that require obtaining specific factual information from the target PDF document, while "analysis questions" refer to questions that require logical reasoning, comparative analysis, summarization, or trend judgment based on multi-source heterogeneous information on each page; the template also forcibly associates and configures "answer format" and "answer unit"; for "analysis questions", the "query questions" can be used as the material needed to analyze specific questions by configuring IDs.

[0019] Step 4: Use a large language model to perform semantic intent analysis on the structured retrieval question and then refine it to obtain core keywords and generate a set of synonyms for the corresponding core keywords; match and filter the core keywords and the corresponding set of synonyms with the PDF content index to obtain the first relevant page content.

[0020] In step 4, the extraction process includes: removing redundant words, including modal particles and conjunctions, and retaining nouns, verbs, and limiting adjectives as core keywords; when generating the synonym set, semantic adaptation is performed in combination with the domain scenario to which the structured retrieval question belongs, to ensure the consistency of domain expression between synonyms and core keywords.

[0021] Step 5: Use a pre-trained language model to vectorize and encode the structured retrieval question and the description text of each page, then calculate the similarity, and filter according to a preset similarity threshold to obtain the second relevant page content.

[0022] In step 5, a pre-trained language model (using Qwen3-Embedding) is used to vectorize and encode the structured retrieval question and the description text of each page, and then similarity calculation is performed. The results are then filtered according to a preset similarity threshold to obtain the second relevant page content. This process includes the following sub-steps: The structured retrieval question is vectorized using a pre-trained language model to obtain the semantic vector of the retrieval question; The description text of each page is vectorized and encoded using a pre-trained language model to obtain a set of semantic vectors for the page description text. The semantic vectors of each page description text in the set of page description text semantic vectors are compared with the semantic vectors of the retrieval question to obtain the set of semantic relevance between the page description text and the retrieval question. The following relationship exists in the correspondence process: ; in, Represents the semantic vector of the retrieval question. This represents the semantic vector of the page description text. express and Cosine similarity between them Indicates transpose. This indicates taking the 2-norm; Based on the semantic relevance set between the page description text and the search question, the semantic relevance between the page description text and the search question is compared with a preset similarity threshold, and page description texts with a semantic relevance greater than the preset similarity threshold are selected to obtain the second relevant page content.

[0023] Furthermore, the structured retrieval question is vectorized using a pre-trained language model to obtain the semantic vector of the retrieval question. This process includes the following sub-steps: Using a structured retrieval question as input text, a token segmenter is used to process the input text, resulting in a discrete token sequence. The following relationship exists in the process: ; in, This indicates that the data has been processed by a word segmenter. This indicates the input text. This represents the first token. This represents the second token. Indicates the first Each Token Indicates the index of the token. Indicates the first Each Token Indicates the length of the discrete token sequence; The discrete token sequence is one-hot encoded, and then word embeddings, position embeddings, and segment embeddings are successively fused to obtain a high-dimensional dense embedding matrix. The following relationship exists in the corresponding process: ; in, express The corresponding semantic vector, The word embedding weight matrix represents the word embedding weight matrix. This indicates that it has undergone one-hot encoding. The position embedding weight matrix represents the position. express The corresponding one-hot vector of position index, The segment embedding weight matrix represents the segment embedding weight matrix. express The corresponding segment index one-hot vector, Represents a high-dimensional dense embedding matrix. , and All represent elements in a high-dimensional densely embedded matrix; The high-dimensional dense embedding matrix is ​​processed sequentially through normalization, multi-head self-attention mechanism, feedforward network and linear layer to obtain the semantic vector of the retrieval question. The following relationship exists in the corresponding process: ; in, This represents the output after normalization. This indicates that the data has been normalized. This represents the output after processing by the multi-head attention mechanism. This indicates that the data has been processed using the softmax function. This represents the projection matrix corresponding to the query matrix. This represents the projection matrix corresponding to the key matrix. The dimension of the key vector. This represents the projection matrix corresponding to the value matrix. This represents the output after processing by the multi-head attention mechanism and adding it to the high-dimensional dense embedding matrix. Indicates intermediate features, This indicates that the data has been processed by the forward network. Represents the semantic vector of the retrieval question. This represents the projection matrix of the linear layer.

[0024] Step 6: Merge and filter the content of the first and second relevant pages in sequence, and use a large language model for logical association and reasoning analysis to generate a preliminary answer.

[0025] In step 6, the content of the first relevant page and the content of the second relevant page are merged and filtered sequentially, and logical association and reasoning analysis are performed using a large language model to generate a preliminary answer. This includes the following sub-steps: The content of the first relevant page and the second relevant page are merged, and duplicate pages are removed to obtain the merged page content; Based on the semantic relevance set between the page description text and the search question, each page in the merged page content is sorted in descending order according to the semantic relevance between the page description text and the search question. The first K (K is a preset positive integer) pages are retained and the remaining pages are removed to obtain the final relevant page content. A large language model is used to perform logical association and reasoning analysis on the content of the relevant pages to generate a preliminary answer.

[0026] Step 7: Using a large language model and combining the answer format and answer unit in the structured retrieval question, verify and correct the preliminary answer for formal compliance, unit accuracy, and expression standardization to obtain the final answer to the retrieval question.

[0027] In step 7, the specific process of verification and correction is as follows: First, the large language model compares whether the presentation format of the preliminary answer is consistent with the "answer format" preset in the template; then, for numerical answers in the template where the "answer unit" is configured to be non-empty, it verifies whether the numerical unit in the preliminary answer is consistent with the preset unit; finally, after checking the expression standard of the preliminary answer, the final answer is obtained.

[0028] This embodiment also provides a document query and analysis intelligent agent construction system based on a large language model. The system applies the document query and analysis intelligent agent construction method based on a large language model as described above. The system includes a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, it can implement the method steps described above.

[0029] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0030] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0031] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0032] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for constructing a document query and analysis intelligent agent based on a large language model, characterized in that, The method includes the following steps: Step 1: Obtain the target PDF document, and extract features from each page of the target PDF document based on the OCR algorithm to obtain multi-source heterogeneous information for each page; Step 2: Use a large language model to semantically describe the multi-source heterogeneous information on each page, generating descriptive text for each page; construct a PDF content index based on the descriptive text of all pages. Step 3: Use templates to standardize the query and analysis questions required for the target PDF document, so as to construct a structured retrieval question that includes ID, question type, answer format and answer unit; Step 4: Use a large language model to perform semantic intent analysis on the structured retrieval question and then refine it to obtain core keywords and generate a set of synonyms for the corresponding core keywords; match and filter the core keywords and the corresponding set of synonyms with the PDF content index to obtain the first relevant page content; Step 5: Use a pre-trained language model to vectorize and encode the structured retrieval question and the description text of each page, then calculate the similarity, and filter according to a preset similarity threshold to obtain the content of the second relevant page. Step 6: Merge and filter the content of the first and second relevant pages in sequence, and use a large language model for logical association and reasoning analysis to generate a preliminary answer; Step 7: Using a large language model and combining the answer format and answer unit in the structured retrieval question, verify and correct the preliminary answer for formal compliance, unit accuracy, and expression standardization to obtain the final answer to the retrieval question.

2. The method for constructing a document query and analysis intelligent agent based on a large language model according to claim 1, characterized in that, In step 1, the multi-source heterogeneous information on each page is structured Markdown text containing one or more of the following: text, tables, formulas, and charts. The OCR algorithm uses Nanonet-OCR-s, ​​which can parse the input document image into structured Markdown text. This not only recognizes the text content but also deeply understands and extracts complex semantic elements in the document, including tables, formulas, signatures, and checkboxes, ensuring the completeness and positional accuracy of the extracted multi-source heterogeneous information.

3. The method for constructing a document query and analysis intelligent agent based on a large language model according to claim 2, characterized in that, In step 2, when generating descriptive text, the large language model will summarize the core content expressed by the multi-source heterogeneous information of each page, the key information in the page, and the logical relationships between the page content.

4. The method for constructing a document query and analysis intelligent agent based on a large language model according to claim 3, characterized in that, In step 3, the template is a structured and standardized template, and the ID of the constructed structured retrieval question is unique. The question types are clearly divided into two categories: "query questions" and "analysis questions." "Query questions" refer to questions that require obtaining specific factual information from a target PDF document, while "analysis questions" refer to questions that require logical reasoning, comparative analysis, summarization, or trend judgment based on multi-source heterogeneous information on each page. The template also forcibly associates and configures "answer format" and "answer unit." For "analysis questions," the "query questions" can be configured as materials needed to analyze specific questions by configuring IDs.

5. The method for constructing a document query and analysis intelligent agent based on a large language model according to claim 4, characterized in that, In step 4, the extraction process includes: removing redundant words, including modal particles and conjunctions, and retaining nouns, verbs, and limiting adjectives as core keywords; when generating the synonym set, semantic adaptation is performed in combination with the domain scenario to which the structured retrieval question belongs, to ensure the consistency of domain expression between synonyms and core keywords.

6. The method for constructing a document query and analysis intelligent agent based on a large language model according to claim 5, characterized in that, In step 5, the structured retrieval question and the description text of each page are vectorized and encoded using a pre-trained language model before similarity calculation is performed. Then, the content of the second relevant page is filtered according to a preset similarity threshold. This process includes the following sub-steps: The structured retrieval question is vectorized using a pre-trained language model to obtain the semantic vector of the retrieval question; The description text of each page is vectorized and encoded using a pre-trained language model to obtain a set of semantic vectors for the page description text. The semantic vectors of each page description text in the set of page description text semantic vectors are compared with the semantic vectors of the retrieval question to obtain the set of semantic relevance between the page description text and the retrieval question. The following relationship exists in the correspondence process: ； in, Represents the semantic vector of the retrieval question. This represents the semantic vector of the page description text. express and Cosine similarity between them Indicates transpose. This indicates taking the 2-norm; Based on the semantic relevance set between the page description text and the search question, the semantic relevance between the page description text and the search question is compared with a preset similarity threshold, and page description texts with a semantic relevance greater than the preset similarity threshold are selected to obtain the second relevant page content.

7. The method for constructing a document query and analysis intelligent agent based on a large language model according to claim 6, characterized in that, The step of using a pre-trained language model to vectorize the structured retrieval question to obtain a semantic vector for the retrieval question specifically includes the following sub-steps: Using a structured retrieval question as input text, a token segmenter is used to process the input text, resulting in a discrete token sequence. The following relationship exists in the process: ； in, This indicates that the data has been processed by a word segmenter. This indicates the input text. This represents the first token. This represents the second token. Indicates the first Each Token Indicates the index of the token. Indicates the first Each Token Indicates the length of the discrete token sequence; The discrete token sequence is one-hot encoded, and then word embeddings, position embeddings, and segment embeddings are successively fused to obtain a high-dimensional dense embedding matrix. The following relationship exists in the corresponding process: ； in, express The corresponding semantic vector, The word embedding weight matrix represents the word embedding weight matrix. This indicates that it has undergone one-hot encoding. The position embedding weight matrix represents the position. express The corresponding one-hot vector of position index, The segment embedding weight matrix represents the segment embedding weight matrix. express The corresponding segment index one-hot vector, Represents a high-dimensional dense embedding matrix. , and All represent elements in a high-dimensional densely embedded matrix; The high-dimensional dense embedding matrix is ​​processed sequentially through normalization, multi-head self-attention mechanism, feedforward network and linear layer to obtain the semantic vector of the retrieval question. The following relationship exists in the corresponding process: ； in, This represents the output after normalization. This indicates that the data has been normalized. This represents the output after processing by the multi-head attention mechanism. This indicates that the data has been processed using the softmax function. This represents the projection matrix corresponding to the query matrix. This represents the projection matrix corresponding to the key matrix. The dimension of the key vector. This represents the projection matrix corresponding to the value matrix. This represents the output after processing by the multi-head attention mechanism and adding it to the high-dimensional dense embedding matrix. Indicates intermediate features. This indicates that the data has been processed by the forward network. Represents the semantic vector of the retrieval question. This represents the projection matrix of the linear layer.

8. The method for constructing a document query and analysis intelligent agent based on a large language model according to claim 7, characterized in that, In step 6, the content of the first relevant page and the content of the second relevant page are sequentially merged and filtered, and logical association and reasoning analysis are performed using a large language model to generate a preliminary answer. This includes the following sub-steps: The content of the first relevant page and the second relevant page are merged, and duplicate pages are removed to obtain the merged page content; Based on the semantic relevance set between the page description text and the search question, each page in the merged page content is sorted in descending order according to the semantic relevance between the page description text and the search question. The top K pages are retained and the remaining pages are removed to obtain the final relevant page content. A large language model is used to perform logical association and reasoning analysis on the content of the relevant pages to generate a preliminary answer.

9. A document query and analysis intelligent agent construction system based on a large language model, characterized in that, The system employs the document query and analysis intelligent agent construction method based on any one of claims 1 to 8. The system includes a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, it can implement the steps of the method described in any one of claims 1 to 8.

Images