Knowledge-intensive multi-document question and answer method and product based on multi-agent collaboration

Abstract

The invention discloses a knowledge-intensive multi-document question-answering method and product based on multi-agent cooperation, and the method comprises the steps: 1, dynamically generating a multi-dimensional complementary expert role set based on the multi-view agent cooperation of roles, forming a candidate strategy set based on all expert roles, and selecting an optimal strategy for integrating knowledge through a voting mechanism; 2, constructing a closed-loop process of refinement of the reflection knowledge, extracting atomic facts through a sliding window and an reflection construction mechanism, and performing a recursive process of task decomposition and knowledge distillation to obtain essence knowledge; 3, based on the optimal strategy in the step 1 and the essence knowledge obtained in the step 2, cross-document comprehensive reasoning is achieved, and a final answer is generated. According to the method, the limitation of a single role view angle is broken through, a high-quality knowledge basis is provided for subsequent reasoning, different types of knowledge-intensive cross-document tasks can be flexibly handled, dispersed and multi-source information is efficiently processed, and the method has wide application prospects and market value.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, and in particular relates to a knowledge-intensive multi-document question answering method and product based on multi-agent collaboration. Background Technology

[0002] Due to their excellent performance across various scenarios, large language models are widely expected to be applied to cross-document question-answering tasks, such as legal analysis and financial report analysis. However, due to a lack of in-depth domain-specific knowledge, large language models still struggle to generate factually accurate content (i.e., “illusions”) in tasks requiring precise factual answers. To address this limitation, the Retrieval-Enhanced Generation (RAG) framework has been proposed, which enhances the ability of large language models to answer questions by incorporating external knowledge. However, traditional RAG methods process knowledge through simple text chunking, which faces the problem that document-level knowledge may be too long to fit within the context window of the language model, causing large models to be unable to consider all knowledge, as longer documents often overlook key information. Furthermore, the relevant information required to solve these tasks is often scattered across multiple documents, making it difficult for models to effectively integrate and utilize scattered information for reasoning. For example, in fiscal year budget analysis, key data needs to be extracted and correlated from multiple historical budget documents, expenditure reports, and related policy documents, requiring models to have the ability to integrate and compare information across documents. However, these sample documents often contain a lot of noise, making it difficult for models to identify relationships between information and perform accurate knowledge reasoning.

[0003] Many studies have indicated that analogizing the working mechanism of large language models to human reasoning processes helps in better handling complex tasks. Similar to humans, large language models, when processing information, are not limited to direct reading but tend to refine information into a structured form to reduce cognitive load and improve judgment accuracy. This process is often accompanied by self-reflection and verification. Likewise, large language models possess reflective reasoning capabilities, able to identify and correct errors, simulating the deliberate judgment process of humans. Cognitive fit theory further points out that humans tend to employ different forms of knowledge organization when facing different types of tasks, such as using tables for statistical analysis or relying on graphics to assist in long-chain reasoning. Recent research shows that large language models also possess the ability to construct diverse knowledge structures.

[0004] However, these methods are limited by their reliance on a single strategy, which hinders their effectiveness in handling complex real-world information. This dependence on a single approach leads to difficulties in establishing logical connections and results in information loss across documents. This makes reasoning challenging, especially when dealing with fragmented content. While large language models can generate diverse knowledge structures and reflect their reasoning, their inability to dynamically combine multiple strategies limits their ability to effectively process and integrate scattered information, which is crucial for more complex tasks.

[0005] In summary, existing technologies for cross-document question answering tasks suffer from the following problems: First, traditional retrieval-enhanced generation frameworks, due to their simple text chunking for knowledge processing, struggle to adapt to the contextual windows of large language models, easily overlooking key information in long documents and failing to effectively integrate relevant information scattered across multiple documents, thus limiting inference performance. Second, existing knowledge processing methods relying on single strategies struggle to establish logical connections between complex information, easily leading to cross-document information loss, especially when processing fragmented content, resulting in significant inference difficulties. Third, while large language models possess the ability to generate diverse knowledge structures and reflective reasoning, they cannot dynamically combine multiple strategies, making it difficult to efficiently process and integrate scattered information, and thus failing to meet the needs of complex cross-document question answering tasks. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and to propose a knowledge-intensive multi-document question answering method and product based on multi-agent collaboration.

[0007] In view of this, the present invention proposes a knowledge-intensive multi-document question answering method based on multi-agent collaboration, comprising: Step 1: Based on role-based multi-perspective intelligent agent collaboration, dynamically generate a set of multi-dimensional complementary expert roles, form a set of candidate strategies based on each expert role, and select the optimal strategy that integrates knowledge through a voting mechanism; Step 2: Construct a closed-loop process for reflective knowledge refinement. Through a sliding window and a reflective construction mechanism, extract atomic-level facts, and carry out a recursive process of task decomposition and knowledge distillation to obtain essential knowledge. Step 3: Based on the optimal strategy in Step 1 and the essential knowledge obtained in Step 2, perform cross-document comprehensive reasoning to generate the final answer.

[0008] As an improvement to the above method, step 1 includes: Step 101: Construct a general role definition prompt template based on the task description to guide the large language model to dynamically generate a set of multi-dimensional complementary expert roles; each expert role is a structured intelligent agent, including role description, voting criteria and task consistency weight; Step 102: Generate multiple strategies in parallel based on each expert role to integrate long text information representations across documents, forming a candidate strategy set; Step 103: Each expert role votes on the set of candidate strategies and selects the strategy with the highest ranking as the optimal strategy for integrating knowledge, based on the task consistency weight.

[0009] As an improvement to the above method, the prompt template in step 101 includes the target task and related document titles.

[0010] As an improvement to the above method, in step 101, semantic constraints are used to ensure that there is minimal semantic overlap between the features of each expert role, thereby avoiding viewpoint redundancy.

[0011] As an improvement to the above method, in step 102, the candidate strategies include at least one of the following: a knowledge graph dynamically generated based on document content, a comparison table, a timeline, and a flowchart.

[0012] As an improvement to the above method, step 2 includes: Step 201: Process the document using a document parsing tool, extract atomic-level facts by combining sliding windows and self-reflection mechanisms, generate a local summary for each sliding window, update the task memory related to the current window, and then summarize them to form a document summary; Step 202: Traverse the document set based on the optimal strategy to initially construct coarse-grained cross-document knowledge; Step 203: Break the task into a set of subtasks, call the large language model to extract content related to the subtasks, further generate essential knowledge, and iteratively repeat this step until the set of subtasks and the essential knowledge are confirmed as the final form by the reflection mechanism.

[0013] As an improvement to the above method, the sliding window size in step 201 is a configurable parameter, and adjacent paragraphs are processed each time to generate local summaries and task memories.

[0014] As an improvement to the above method, in step 203, if the task cannot be split, the task splitting step is skipped.

[0015] As an improvement to the above method, step 3 includes: Taking the optimal strategy, document summary, final set of subtasks, and essential knowledge as input, the large language model is guided by prompt words to perform cross-document comprehensive reasoning and generate the final answer.

[0016] On the other hand, the present invention provides a computer program product including computer-readable instructions that, when executed on a computer device, cause the computer device to perform the above-described method.

[0017] Compared with the prior art, the advantages of the present invention are: 1. Propose a role-based multi-perspective agent collaboration: Unlike the fixed knowledge processing mode of traditional RAG (Research and Answering), this invention can dynamically generate suitable roles based on the specific needs of cross-document question answering tasks (such as fiscal year budget analysis), and integrate the information insights generated by agents from multiple role perspectives through a voting mechanism to determine the optimal strategy for knowledge integration. This mechanism breaks through the limitations of a single role perspective, accurately extracting core information from each document, and proactively revealing the inherent connections between fragmented content scattered across multiple documents (such as the correspondence between budget information from different years), solving the problem that traditional methods struggle to establish logical connections across documents.

[0018] 2. Constructing a Reflective Knowledge Refinement Closed-Loop Process: This innovative approach deeply integrates multi-role intelligent agent collaboration with sliding window technology and a reflective mechanism. After information extraction by multiple expert intelligent agents, a reflective knowledge construction phase is introduced. Through the self-reflection mechanism of a large-scale language model, coarse-grained knowledge undergoes secondary verification (e.g., verifying the compliance of budget data and the consistency of cross-year data comparisons), refined breakdown (e.g., breaking down complex budget preparation logic into sub-modules of "revenue forecasting - expenditure allocation - gap analysis"), and simplification and optimization (e.g., merging duplicate budget descriptions and unifying data standards across departments), forming a closed loop of "information extraction - multi-perspective integration - reflective refinement." This process effectively reduces the loss of long text information across documents, providing a high-quality knowledge foundation for subsequent reasoning and overcoming the shortcomings of existing methods that only focus on information extraction and lack dynamic refinement phases.

[0019] 3. Achieving a cross-document reasoning mode with dynamic strategy combination: Breaking through the limitations of existing technologies that rely on single knowledge processing strategies, this model, through a collaborative design of "dynamic role generation - multi-perspective integration - reflective refinement," allows large-scale language models to autonomously combine appropriate processing strategies based on task complexity and document information characteristics (such as task decomposition and effective knowledge decomposition when handling complex tasks). This dynamic combination capability enables the model to flexibly cope with different types of knowledge-intensive cross-document tasks, efficiently process scattered and multi-source information, and solve the problem that existing models are difficult to adapt to complex scenarios due to fixed strategies. It has broad application prospects and market value. Attached Figure Description

[0020] Figure 1 This is a flowchart of the role-based multi-perspective intelligent agent collaboration process of the present invention; Figure 2 This is a flowchart of the process for constructing a reflective knowledge refinement closed loop in this invention; Figure 3 This is a flowchart of the cross-document reasoning mode for implementing dynamic strategy combination according to the present invention. Detailed Implementation

[0021] Based on the above reasons, this invention proposes a knowledge-intensive multi-document question answering method based on multi-agent collaboration. The core of this method lies in simulating the collaboration and deliberation process of human expert teams when dealing with complex problems. By introducing a role-based multi-perspective collaborative dynamic strategy mechanism and constructing a closed-loop process for reflective knowledge refinement, it systematically solves the shortcomings of traditional methods in processing long documents, cross-document information fusion, and deep reasoning.

[0022] The method mainly includes: Step 1) Role-based multi-perspective agent collaboration: Roles are dynamically generated based on tasks, and each agent integrates insights from multiple perspectives through a voting mechanism to determine the optimal information processing strategy. This facilitates concise information extraction and reveals the relationships between fragmented content.

[0023] Specifically, it includes: Step 101) Construct a general role definition prompt template based on the task description. The prompt template includes the target task and related document titles, which are used to guide the large language model to dynamically generate a multi-dimensional and complementary set of expert roles. It should be noted that the large language model mentioned in this invention, such as the Qwen series, can be used.

[0024] Step 102) Drive the agent to generate multiple strategies for integrating long text information representation across documents using the generated expert perspective, forming a candidate strategy set for subsequent intuitive representation and information simplification.

[0025] Step 103) Then, each expert agent votes on each strategy in the candidate strategy set according to its own perspective, summarizes the scores and selects the strategy with the highest ranking as the optimal strategy.

[0026] Step 2) Construct a closed-loop process for reflective knowledge refinement; introduce a reflective knowledge construction mechanism, combining multiple perspectives to enhance the diversity of results. This process can further refine, segment, and simplify useful content, reduce information loss, and improve reasoning efficiency.

[0027] Specifically, it includes: Step 201) Process the parsed documents. Use the MinerU document parsing tool to obtain the document content and cut long documents into short documents and integrate them into the same folder. Utilize the sliding window technique and the self-reflection ability of the large model, starting from the aforementioned expert perspective, extract atomic-level facts from the document set, judge similarity to filter and retain short documents relevant to the task, and discard other parts to reduce the burden of subsequent processing. At the same time, atomic-level information can also serve as a document summary.

[0028] Step 202) Then, the optimal strategy is used to drive the large language model to traverse the preserved document set and initially construct coarse-grained cross-document knowledge.

[0029] (Step 203) Next, for the specific task, the expert agent is used to break the task down into a set of subtasks and vote to select the most suitable set (skip this step if the original task is simple enough or cannot be broken down). Using the generated summary and coarse-grained cross-document knowledge as background information, the large language model is invoked to extract content related to the subtasks and further generate concise knowledge. This step is repeated iteratively until the set of subtasks and the concise knowledge are both discussed and confirmed to be correct by experts.

[0030] Step 3) Implements a cross-document reasoning mode with dynamic strategy composition.

[0031] Specifically, it includes: Step 301) Using the optimal strategy determined in Step 1, the information summary generated in Step 2, the final set of subtasks, and the final simplified knowledge as input, the large model is guided to generate the final result of the task through prompt words.

[0032] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0033] Example 1 Embodiment 1 of the present invention proposes a knowledge-intensive multi-document question answering method based on multi-agent collaboration. It mainly includes three core stages: Step 1) Role-based multi-perspective agent collaboration; Step 2) Construct a closed-loop process for reflective knowledge refinement; Step 3) Implements a cross-document reasoning mode with dynamic strategy composition.

[0034] Step 1) specifically includes: Step 101) Generation and Initialization of the Dynamic Expert Role Set. This step is the cornerstone of building diverse problem-solving perspectives. First, based on the user-submitted task description (e.g., "Analyze a country's fiscal year budget"), the system activates a predefined, highly structured, general role definition prompt template. This template not only contains a precise description of the target problem but also automatically links and lists the titles of multiple documents highly relevant to the problem in the knowledge base (e.g., "A country's fiscal report for a certain year"), providing rich context for role generation. The core function of the prompt template is to guide the large language model to simulate an "expert committee," generating a multi-dimensional expert role set that is both professional and complementary in terms of domain knowledge, analytical methods, and focus.

[0035] Each generated expert role is a structured intelligent agent, consisting of three indispensable components: (1) Role description: This is the "personalization" setting of the role, serving as the core cue to guide the subsequent behavior of the large language model. It clearly defines the specific perspective, professional field, and thinking paradigm that the model must adopt when handling tasks. For example, the "data analyst" role would be told to "focus on extracting insights from quantitative data, trend charts, and statistical models, and remain skeptical of any qualitative descriptions"; (2) Voting criteria: This is the role's "judgment criteria," providing a quantitative basis for subsequent strategy evaluation. It defines the core dimensions and specific judgment rules that the role uses when examining a candidate strategy. For example, the voting criteria for the product manager role may include "clarity of user value (weight 40%)", "market feasibility (weight 30%)", and "synergy with company strategy (weight 30%)", with a scoring detail of 1-5 points designed for each dimension; (3) Task consistency: This is a dynamic weight coefficient used to measure the closeness of the correlation between the perspective of the specific role and the final goal of the current task. It determines the weight of the role's score in the final strategy summary. This consistency coefficient is not fixed, but is fine-tuned during the evaluation process in step 103 based on the semantic matching degree between the role description and the task objective.

[0036] To ensure efficiency and diversity of perspectives, the system embeds semantic constraints in the generated prompts. By requiring the model to maximize the cosine distance between the character description vectors or minimize semantic similarity, it ensures that the features of each expert character have minimal semantic overlap, thereby fundamentally avoiding perspective redundancy and "groupthink" problems. Simultaneously, semantic constraints are set through prompt templates to ensure minimal semantic overlap of the features of each expert character, avoiding perspective redundancy.

[0037] Step 102) Parallel Generation of Candidate Strategies. Guided by the set of expert roles generated in Step 101 (e.g., instantiated data analysts, technical experts, product managers, risk managers, legal advisors, etc.), the system enters the strategy conception stage. Each role acts as an independent "thinking thread," driving the large language model to generate a unique multimodal representation strategy for integrating long-text information across documents, based on the specific problem and the provided document set context. Here, "multimodal representation" refers not only to traditional text but also to structured data dynamically generated based on document content, including at least one of knowledge graphs, comparison tables, timelines, and flowcharts in one embodiment. For example, a data analyst might propose a strategy of "extracting all years' financial budget growth rate and market share data from various documents to construct a competitive landscape comparison table"; a technical expert might propose a strategy of "analyzing core technical parameters and patent documents to draw a technology evolution path map"; and a product manager might propose a strategy of "integrating user feedback and feature descriptions to create a feature-value mapping matrix." All these strategies are aggregated to form a rich set of candidate strategies with diverse perspectives.

[0038] Step 103) Strategy Optimization Based on Weighted Voting. This step simulates the decision-making process of an expert committee. The system presents the candidate strategy set generated in Step 102 to each expert role. Each role independently and thoroughly evaluates the effectiveness of each strategy in the set based on its own voting criteria defined in Step 101. The evaluation process includes two levels: first, qualitative ranking to determine the priority of strategies; second, quantitative scoring to assign a specific score to each strategy under its own criteria. Subsequently, the system initiates a weighted aggregation algorithm. This algorithm multiplies each role's score for a strategy by its task consistency weight, and then sums the weighted scores of all roles to obtain the final comprehensive score for the strategy. Finally, the system selects the strategy with the highest comprehensive score ranking and determines it as the optimal strategy for executing subsequent tasks. This process ensures that the final selected strategy is a comprehensive solution that has undergone multi-dimensional and quantitative evaluation and best balances the perspectives of all experts.

[0039] Step 2) specifically includes: Step 201) Iterative Self-Reflection and Memory Enhancement Mechanism. This step aims to endow the system with the ability to continuously learn and self-correct. First, the system constructs a fixed-size memory storage area to cache task memories generated by the large language model during the processing of the current task. These memories are not simple dialogue history, but structured analysis records, including but not limited to: error diagnosis and correction suggestions for previously generated results, identification and supplementation of missing information, confirmation and affirmation of existing content, and key intermediate conclusions in the reasoning process.

[0040] Before implementation, the original long documents need to be preprocessed. In one embodiment, the advanced MinerU document parsing tool is used, which can extract text, table, chart titles, and structural information from the document with high fidelity. Subsequently, each long document is cut into short document fragments that are easy to process according to semantic boundaries (such as chapters and paragraphs), and integrated into the same task folder to prepare for subsequent fine-grained processing.

[0041] The core of the reflective construction mechanism is an iterative cycle: 1. Input Construction: Design a corresponding task prompt template for the specific task scenario (such as summarizing, question answering, or inference). This template integrates three parts of information: the background information required for the task (from document fragments), the results generated by the previous model, and all relevant existing memories in the memory storage area.

[0042] 2. Multi-role model invocation: Using this constructed input, the large language model guided by the expert role set generated in step 101 is invoked. All roles share the same memory storage area to ensure knowledge consistency.

[0043] 3. Output and Evaluation: The expert agent generates the output for this task. Following this, the system triggers the large model's self-reflection capability, requiring the model to critically evaluate the output and generate a new memory of the current task. This memory contains evaluation opinions from the perspectives of all expert roles.

[0044] 4. Decision-Making and Updating: The system analyzes this newly generated task memory. If the memory determines that the current output "has no major problems and can be adopted," the loop terminates and the current output is retained. Conversely, if the memory points out specific problems (such as "incorrect data reference" or "broken logical chain"), the system will immediately correct the output based on these judgments and synchronously update the complete record of "error content - correction opinion - corrected result" to the memory storage area before entering the next iteration. This process ensures the accuracy and robustness of the output. Step 202) Then, the sliding window technique and the reflection construction mechanism described in step 201) are integrated to extract atomic-level facts from the document set, discard irrelevant content, and achieve effective information compression and continuous retention of contextual memory.

[0045] Specifically, each document in the document set is split into multiple paragraphs to form a paragraph set; then the sliding window size is set to 2 (i.e., two adjacent paragraphs are processed each time); next, a document reading task prompt template is designed with the current local window content and question as background information, combined with the local summary generated by the previous window and the corresponding memory, and the large model is called to generate the local summary of the current window and the current memory.

[0046] After performing the above processing on all local windows of each document, all local summaries are summarized to form the full-text summary of the document, thereby constructing a high-quality set of contextual information for all documents.

[0047] Step 202) Progressive knowledge refinement based on sliding window. To address the issues of information loss and model attention distraction in long document context understanding, this step deeply integrates the sliding window technique with the reflective construction mechanism of Step 201), enabling the efficient extraction of "atomic-level facts" from massive documents while discarding irrelevant content.

[0048] Specifically, the system splits each document in the document set into a sequence of paragraphs based on logical paragraphs. Then, a configurable sliding window is set, with a size of 2 in one embodiment, meaning that only two adjacent paragraphs are processed at a time. The system designs a document intensive reading task prompt template, whose input includes: the content of the two paragraphs in the current sliding window, the original question, the partial summary generated in the previous window, and related memories. Based on this information, the large model completes two tasks: first, generating a new partial summary representing the core information of the two current paragraphs; and second, updating the task memories related to the content of the current window.

[0049] Once all windows for a document have been processed sequentially, the system aggregates the partial summaries produced by all windows and generates a high-quality full-text summary of the document using an integrated prompt template. This process is repeated for all documents, ultimately constructing a high-quality set of contextual information representing the core information of the entire document set. This method achieves effective information compression and ensures the continuous transfer and retention of contextual memory within the document.

[0050] Step 203) Recursive process of task decomposition and knowledge distillation. Design a prompt template for the process of splitting a task into subtasks, using the high-quality contextual information set generated in step 202) and the original task as background information, and generate a set of sub-problems from an expert perspective (skip this step if the original task is simple enough or cannot be split). Next, we designed a knowledge simplification task prompt template, using the generated set of subtasks and coarse-grained knowledge as background information, and further extracted relevant content from it from an expert's perspective to generate simplified knowledge. Finally, repeat the iterative process of task decomposition and knowledge simplification described above until both the set of subtasks and the simplified knowledge are confirmed as the final form by the reflective construction mechanism in step 201).

[0051] Step 3) specifically includes: Step 301) Using the optimal strategy determined in Step 1), the high-quality contextual information set generated in Step 2), the final sub-problem set, and the final simplified knowledge as input, the large model is guided by prompt words to perform logical reasoning, cross-document comprehensive reasoning for the target problem, and the final answer is generated.

[0052] Example 2 Embodiments of the present invention may also provide a computer program product, including a computer program. When the computer program is executed by a processor, it can implement the various steps in the above method embodiments.

[0053] Overview: By proposing an intelligent agent question answering method that links multiple long documents to solve knowledge-intensive tasks, this invention can achieve concise cross-document information extraction, reveal the relationship between fragmented content, reduce the loss of information in long documents, and thus improve the reasoning efficiency and result accuracy of large language models in knowledge-intensive cross-document question answering tasks.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A knowledge-intensive multi-document question answering method based on multi-agent collaboration, comprising: Step 1: Based on role-based multi-perspective intelligent agent collaboration, dynamically generate a set of multi-dimensional complementary expert roles, form a set of candidate strategies based on each expert role, and select the optimal strategy that integrates knowledge through a voting mechanism; Step 2: Construct a closed-loop process for reflective knowledge refinement. Through a sliding window and a reflective construction mechanism, extract atomic-level facts, and carry out a recursive process of task decomposition and knowledge distillation to obtain essential knowledge. Step 3: Based on the optimal strategy in Step 1 and the essential knowledge obtained in Step 2, perform cross-document comprehensive reasoning to generate the final answer.

2. The knowledge-intensive multi-document question answering method based on multi-agent collaboration according to claim 1, characterized in that, Step 1 includes: Step 101: Construct a general role definition prompt template based on the task description to guide the large language model to dynamically generate a set of multi-dimensional complementary expert roles; each expert role is a structured intelligent agent, including role description, voting criteria and task consistency weight; Step 102: Generate multiple strategies in parallel based on each expert role to integrate long text information representations across documents, forming a candidate strategy set; Step 103: Each expert role votes on the set of candidate strategies and selects the strategy with the highest ranking as the optimal strategy for integrating knowledge, based on the task consistency weight.

3. The knowledge-intensive multi-document question answering method based on multi-agent collaboration according to claim 2, characterized in that, The prompt template for step 101 includes the target task and related document titles.

4. The knowledge-intensive multi-document question answering method based on multi-agent collaboration according to claim 2, characterized in that, In step 101, semantic constraints are used to ensure that there is minimal semantic overlap between the features of each expert role, thus avoiding viewpoint redundancy.

5. The knowledge-intensive multi-document question answering method based on multi-agent collaboration according to claim 2, characterized in that, In step 102, the candidate strategies include at least one of the following: a knowledge graph dynamically generated based on document content, a comparison table, a timeline, and a flowchart.

6. The knowledge-intensive multi-document question answering method based on multi-agent collaboration according to claim 1, characterized in that, Step 2 includes: Step 201: Process the document using a document parsing tool, extract atomic-level facts by combining sliding windows and self-reflection mechanisms, generate a local summary for each sliding window, update the task memory related to the current window, and then summarize them to form a document summary; Step 202: Traverse the document set based on the optimal strategy to initially construct coarse-grained cross-document knowledge; Step 203: Break the task into a set of subtasks, call the large language model to extract content related to the subtasks, further generate essential knowledge, and iteratively repeat this step until the set of subtasks and the essential knowledge are confirmed as the final form by the reflection mechanism.

7. The knowledge-intensive multi-document question answering method based on multi-agent collaboration according to claim 6, characterized in that, The sliding window size in step 201 is a configurable parameter, and each time adjacent paragraphs are processed, a local summary and task memory are generated.

8. The knowledge-intensive multi-document question answering method based on multi-agent collaboration according to claim 6, characterized in that, In step 203, if the task cannot be split, the task splitting step is skipped.

9. The knowledge-intensive multi-document question answering method based on multi-agent collaboration according to claim 6, characterized in that, Step 3 includes: Taking the optimal strategy, document summary, final set of subtasks, and essential knowledge as input, the large language model is guided by prompt words to perform cross-document comprehensive reasoning and generate the final answer.

10. A computer program product, characterized in that, Includes computer-readable instructions that, when executed on a computer device, cause the computer device to perform the method as described in any one of claims 1 to 9.

Images