Report generation method and device based on multi-agent cooperation, equipment and medium

Abstract

The invention relates to the technical field of knowledge maps, can be applied to business scenes of financial science and technology, medical health and the like, and discloses a report generation method, device, equipment and medium based on multi-agent collaboration.The method comprises the steps that a domain knowledge map is constructed and written into a shared work memory; determining a target document fragment based on the map structure and the reading history; semantic coding is conducted on the target document fragment, and a structured abstract is generated and written into a shared memory; performing semantic aggregation based on the updated shared memory to generate an initial report; and performing logic consistency and compliance verification on the initial report, and outputting a final report after the verification is passed. According to the method, the document understanding depth is improved through a graph-driven agent cooperation mechanism, so that the report generation process has a structured basis and global consistency; the reading and writing processes are linked through shared memory, the accuracy and traceability of report content are improved, and therefore the knowledge generation efficiency and reliability in a multi-source document scene are improved.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to a report generation method, apparatus, device and medium based on multi-agent collaboration. Background Technology

[0002] In the fintech sector, insurance institutions typically need to synthesize a large number of cross-source policy documents, industry reports, historical claims records, and operational documents when processing product descriptions, risk assessments, claims rule updates, and regulatory filings. Existing technologies mostly employ traditional document retrieval and manual analysis. While some systems incorporate large language models, they often only perform summary processing based on local fragments, failing to systematically model the structural relationships between thousands of insurance clauses, regulatory notices, and business process documents. This deficiency makes it difficult for systems to capture cross-document citation chains, regulatory logical connections, and dependencies between clauses, leading to issues such as missing content, logical jumps, and unclear basis during report generation. Furthermore, the lack of a unified, accessible, and updatable working memory structure prevents multiple models from sharing parsed contextual information, resulting in industry knowledge reports that are often one-sided and lack traceability, impacting analytical and compliance applications.

[0003] In the healthcare sector, medical institutions commonly face similar challenges in knowledge-intensive tasks such as compiling disease knowledge, interpreting medical guidelines, organizing drug usage standards, and summarizing case studies. Medical data is characterized by its multidisciplinary, multi-format, multi-temporal, and multi-source nature, while existing knowledge processing technologies typically only process single documents or partial segments, making it difficult to establish entity relationships and dependency structures across literature, departments, and disease domains. Due to the lack of modeling of causal relationships, diagnostic and treatment process connections, and standard reference chains among medical literature, existing systems often generate medical knowledge summaries or reports that suffer from contextual fragmentation, incomplete argumentation chains, and a lack of source evidence. Furthermore, medical knowledge processing scenarios often require collaborative semantic understanding involving multiple people and multiple models, but traditional systems generally employ a single-model linear processing approach, making it difficult to dynamically accumulate and update domain knowledge during continuous reading. This results in output reports that fail to meet the completeness and consistency requirements of clinical research or medical management. Summary of the Invention

[0004] The main objective of this invention is to provide a report generation method, apparatus, device, and storage medium based on multi-agent collaboration, aiming to solve the technical problem that existing technologies cannot perform structured association modeling of large-scale cross-source domain documents and continuously update the shared context under multi-agent collaboration, resulting in a lack of logical consistency and traceability of the generated knowledge reports.

[0005] To achieve the above objectives, the present invention provides a report generation method based on multi-agent cooperation, comprising: Receive multi-source domain data, perform entity recognition and relation extraction on the multi-source domain data, construct a domain knowledge graph based on the entity recognition and relation extraction results, and store the domain knowledge graph in shared working memory; The target document fragment is determined by selecting an intelligent agent based on the topological features of the domain knowledge graph and the historical reading state in the shared working memory; The target document fragment is semantically encoded using a reading agent to generate a structured summary, and the structured summary is written into the shared working memory to update the shared working memory; The writing agent performs semantic aggregation based on the updated shared working memory to generate an initial domain knowledge report; The initial domain knowledge report is subjected to logical consistency and constraint compliance checks. When the logical consistency and constraint compliance checks pass, the final domain knowledge report is output.

[0006] Furthermore, to achieve the above objectives, the present invention provides a report generation apparatus based on multi-agent collaboration, comprising: The domain knowledge construction module is used to receive multi-source domain data, perform entity recognition and relation extraction on the multi-source domain data, construct a domain knowledge graph based on the entity recognition and relation extraction results, and store the domain knowledge graph in shared working memory; The agent selection module is used to determine the target document fragment by utilizing the agent selection based on the topological features of the domain knowledge graph and the historical reading state in the shared working memory; The reading agent module is used to perform semantic encoding on the target document fragment using a reading agent, generate a structured summary, and write the structured summary into the shared working memory to update the shared working memory; The writing agent module is used to generate an initial domain knowledge report by leveraging the writing agent to perform semantic aggregation based on the updated shared working memory. The report verification module is used to perform logical consistency and constraint compliance verification on the initial domain knowledge report, and outputs the final domain knowledge report when the logical consistency and constraint compliance verification passes.

[0007] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a multi-agent cooperative report generation program stored in the memory and executable on the processor, wherein when the multi-agent cooperative report generation program is executed by the processor, it implements the steps of the multi-agent cooperative report generation method as described above.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a non-volatile computer-readable storage medium storing a report generation program based on multi-agent cooperation, wherein when the report generation program based on multi-agent cooperation is executed by a processor, it implements the steps of the report generation method based on multi-agent cooperation as described above.

[0009] Beneficial Effects: This invention relates to the field of knowledge graph technology and can be applied to business scenarios such as fintech and healthcare. It discloses a report generation method, apparatus, device, and medium based on multi-agent collaboration, comprising: constructing a domain knowledge graph and writing it into shared working memory; determining target document fragments based on the graph structure and reading history; semantically encoding the target document fragments and generating a structured summary, which is then written into shared memory; semantically aggregating the updated shared memory to generate an initial report; performing logical consistency and compliance checks on the initial report, and outputting the final report upon successful verification. This invention enhances document understanding depth through a graph-driven agent collaboration mechanism, enabling the report generation process to have a structured basis and global consistency; and improves the accuracy and traceability of report content by linking reading and writing processes through shared memory, thereby improving the efficiency and reliability of knowledge generation in multi-source document scenarios. Attached Figure Description

[0010] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for a report generation method based on multi-agent collaboration in one embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the report generation method based on multi-agent collaboration of the present invention; Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the report generation device based on multi-agent collaboration of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0011] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0012] The report generation method based on multi-agent cooperation provided in this invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server can construct a domain knowledge graph through the client and write it into shared working memory; determine target document fragments based on the graph structure and reading history; semantically encode the target document fragments and generate a structured summary, which is then written into shared memory; perform semantic aggregation based on the updated shared memory to generate an initial report; perform logical consistency and compliance checks on the initial report, and output the final report upon successful verification. This invention improves the depth of document understanding through a graph-driven intelligent agent collaboration mechanism, enabling the report generation process to have a structured basis and global consistency; and improves the accuracy and traceability of report content by linking reading and writing processes through shared memory, thereby improving the efficiency and reliability of knowledge generation in multi-source document scenarios. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0013] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the report generation method based on multi-agent collaboration provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0014] like Figure 2 As shown, the report generation method based on multi-agent cooperation proposed in this invention includes the following steps: S10, receive multi-source domain data, perform entity recognition and relation extraction on the multi-source domain data, construct a domain knowledge graph based on the entity recognition and relation extraction results, and store the domain knowledge graph in shared working memory; In this embodiment, the operation of receiving multi-source domain data is used to introduce text, structured records, or semi-structured content from different sources, which may include business archives, knowledge bases, public materials, or regulatory texts. Multi-source data differs in format, encoding, and expression, thus requiring unified standardization to ensure effective processing by subsequent identification modules. When performing entity identification on multi-source domain data, models or rules are used to identify concept names, indicator fields, condition descriptions, or key events with independent semantics, giving important parts of the data clear semantic boundaries. Subsequently, relation extraction is performed, identifying the relationships between these entities through cross-sentence inference, dependency structure analysis, or referential resolution, including logical references, trigger relationships, upstream and downstream dependencies, or topic consistency. These relationships may span different documents, therefore, the extraction process also requires fusion judgment of document source, temporal order, and semantic context.

[0015] When constructing a domain knowledge graph based on entity recognition and relation extraction results, entities are used as nodes and relations as edges. Through deduplication, synonym alignment, and attribute supplementation, the content is organized into a scalable graph structure. This graph structure not only expresses the semantics of a single document but also reflects implicit logical relationships across documents, enabling content from different sources to be understood within a unified structure. The constructed domain knowledge graph is then written into shared working memory, providing a unified semantic foundation for subsequent intelligent agents. Shared working memory can be implemented using graph databases, key-value indexes, or vector retrieval structures. During writing, retrieval indexes and node identifiers need to be established for rapid access and expansion by subsequent tasks.

[0016] In one implementation, multi-source domain data is introduced through batch file loading and API data aggregation. A format unification module handles differences in encoding, punctuation, and fields, enabling continuous text processing. Entity recognition uses a deep sequence model to scan the text segment by segment, generating entity labels. Relation extraction employs syntactic dependency analysis combined with cross-segment semantic judgment to ultimately establish relationships between entities. When constructing the graph structure, a distributed graph database is used to write to shared working memory, ensuring fast access to large-scale data in subsequent stages.

[0017] In another implementation, multi-source data can come from log streams or continuously updated databases, and real-time text normalization is performed through streaming preprocessing. Entity recognition employs a hybrid engine combining rules and models, using rules to identify standardized terms and models to identify unfamiliar concepts. Relationship extraction can identify conditional relationships between regulatory clauses through template matching. When writing graph structures to shared working memory, a vectorized node structure can be used, enabling agents to locate similar content through semantic retrieval, thereby enhancing cross-document understanding capabilities.

[0018] This embodiment receives multi-source domain data and, through entity recognition, relation extraction, and graph structure construction, organizes a large amount of scattered information into a scalable knowledge structure. After being written into shared working memory, multiple agents share a unified semantic context in subsequent processing stages, improving the accuracy of document screening, summary generation, and content integration. This results in a deeper understanding of cross-source materials, better structural consistency, and higher processing efficiency.

[0019] S20, using the selected agent to determine the target document fragment based on the topological features of the domain knowledge graph and the historical reading state in the shared working memory; In this embodiment, when selecting an agent to determine a target document fragment based on the topological features of the domain knowledge graph and the historical reading state in shared working memory, it is necessary to simultaneously process the semantic associations of the graph structure and the behavioral trajectories in the memory content. The topological features of the domain knowledge graph reflect the semantic and referential value between texts through the connection density, connection direction, and dependency links of nodes in the global structure. The sources of topological features include node centrality, edge weights, cross-level association, or information flow paths in the graph structure, used to represent the scope and importance of different content in the graph structure. Node centrality reflects the connectivity of an entity in the entire knowledge network, originating from network analysis, and in text graphs, it can be calculated through entity frequency, cross-document reference counts, and association link length. Edge weights describe the strength of associations between different entities, and their values ​​may come from semantic similarity, co-occurrence probability, or confidence of cross-document logical links.

[0020] Historical reading states in shared working memory reflect the system's understanding of content by recording document fragments, reading order, reading time intervals, or content paths related to the query task accessed by the selection agent and the reading agent during previous processing. Sources of historical reading states include structured summaries written in the previous stage, task context, reading sequence indexes, or reading preference weights. These serve to avoid reprocessing already read content and enable the agent to adjust its current decision-making logic based on historical behavior.

[0021] When determining target document fragments, the selection agent needs to read topological features from the graph structure and establish centrality, relevance, or weight scores for each document fragment, enabling them to be ranked within a unified feature space. Subsequently, the selection agent jointly analyzes this result with historical reading status, filtering out already read fragments and then adjusting priorities based on features provided by the graph structure, such as dependency links, upstream references, and downstream inference paths, redistributing unread fragments according to requirements within the scoring system. The selection agent then sorts the fragments based on the adjusted priorities, prioritizing those with high centrality, high relevance, and high task relevance, ultimately generating the target document fragments to be processed.

[0022] This embodiment leverages the topological features of the domain knowledge graph and the historical reading states in shared working memory for joint analysis. This expands the document fragment selection process from simple textual relevance judgment to semantic association judgment at the structural level. The document fragments selected by the agent not only have higher semantic value but also form logically continuous reading paths along knowledge links. This improves the accuracy and completeness of subsequent summary generation and content aggregation, reduces the overhead of repetitive processing, and increases overall processing efficiency.

[0023] S30, the reading agent performs semantic encoding on the target document fragment to generate a structured summary, and writes the structured summary into the shared working memory to update the shared working memory; In this embodiment, when using a reading agent to semantically encode target document fragments, it is necessary to uniformly model the expressive structure, syntactic construction, entity referential relationships, and contextual dependencies of the target document, so that the text content can be converted into a vector representation with machine-understandable capabilities. The sources of semantic encoding include the contextual modeling capabilities of language models, the feature aggregation capabilities of attention mechanisms, and the ability to abstract common entity relationships in domain texts. The semantic vectors generated by semantic encoding are used to represent the content distribution, logical structure, and conceptual associations of document fragments, enabling subsequent information extraction to identify structured information from the vector space.

[0024] When generating structured summaries, key information needs to be extracted from semantic vectors, including entity information, logical constraint information, and topic information. Entity information originates from previous knowledge graph construction processes and covers specific objects, concepts, or variables. Logical constraint information reflects causal relationships, scope limitations, or conditional inferences in the text and is an important component of structured representation. Topic information describes the core content of text fragments and is typically obtained in healthcare and financial texts through keyword terms, domain terminology, or topic vector clustering. The reading agent integrates the above information into a structured summary through processes such as feature selection, semantic compression, and sentence generation, making the text content structurally indexable, alignable, and reasonable.

[0025] When writing structured summaries into shared working memory, a persistent writing mechanism is required to ensure that the summaries can be accessed by other agents. The writing operation involves not only storing data but also updating the working memory content, enabling the new summaries to form a continuous semantic link with previously written segments within the memory system. This updating action needs to be accomplished through a gating update mechanism to control how old and new information are integrated. The sources of gating update mechanisms include memory network structures, attentional memory control methods, or weight selection strategies, ensuring that the shared working memory maintains the freshness and semantic stability of the information.

[0026] This embodiment uses semantic encoding and structured summary generation, combined with a shared working memory update mechanism, to enable target document fragments to enter the intelligent agent collaboration system in a structured and reusable form. This improves the accuracy of information extraction, reduces repetitive reading, enhances the logical consistency of content aggregation, and makes the subsequently generated knowledge reports more closely aligned with the original semantic structure and reference links of the document, thereby improving the overall content quality.

[0027] S40 utilizes the writing agent to perform semantic aggregation based on the updated shared working memory to generate an initial domain knowledge report; In this embodiment, when the writing agent performs semantic aggregation based on the updated shared working memory, it needs to read the structured summaries, semantic vectors, and association information written by the preceding agent from the shared working memory and organize these contents into an aggregatable semantic input. The updated shared working memory originates from the summary content continuously written by the reading agent and the memory adjustments generated by the gating update mechanism, making it contain multi-dimensional features across document fragments, such as topic elements, entity links, logical frameworks, and citation links. The writing agent needs to identify the semantic importance of different memory units through an attention reading mechanism, so that summaries with higher relevance receive higher feature weights during the aggregation process.

[0028] Semantic aggregation involves vector-level fusion of cross-summary representations, transforming the semantic structures of multiple memorized contents into a unified semantic representation. This process stems from the inter-sequence association computation capability of cross-attention mechanisms, enabling the writing agent to dynamically compute semantic correspondences between different summaries. When processing cross-document structures, cross-attention mechanisms can utilize attention matrices to construct association graphs, allowing the writing agent to identify logical coherence, citation dependencies, and topic reinforcement relationships between semantic fragments, and transform these relationships into aggregated semantic representations.

[0029] When generating the initial domain knowledge report, the writing agent needs to transform the aggregated semantic representation into a complete text description. The text generation process leverages the decoding capabilities of a language model, progressively generating multiple paragraphs from the aggregated semantic representation. This ensures the content presents a logical order, concept definitions, and argumentation structure in accordance with industry document expression conventions. The generated text content is then rearranged according to a pre-defined report structure, ensuring the report has a unified heading hierarchy, paragraph structure, and citation layout, making it directly applicable to tasks such as knowledge representation, risk analysis, or policy interpretation.

[0030] This embodiment aggregates the updated shared working memory and generates an initial domain knowledge report, enabling the system to integrate semantic content across document fragments to form a logically complete, semantically continuous textual expression with a reference chain. This provides a stable input foundation for subsequent logical consistency verification and compliance assessment, improves the quality and interpretability of report generation, and reduces the time cost of manual compilation.

[0031] S50, perform logical consistency and constraint compliance checks on the initial domain knowledge report, and output the final domain knowledge report when the logical consistency and constraint compliance checks pass.

[0032] In this embodiment, verifying the logical consistency and constraint compliance of the initial domain knowledge report is a crucial step in verifying the quality of the aggregated results. The initial domain knowledge report originates from the text generation process of the writing agent and includes structured expressions, citation chains, topic reasoning, and cross-summary integration. Logical consistency verification is used to determine the coherence of the internal reasoning process of the text, including the consistency of factual descriptions, the relevance of conceptual references, the logical connection between paragraphs, and the rationality of citation chains. Logical consistency verification is performed by a semantic parsing module, which can identify content with contradictions, broken links, or semantic jumps, classifying them as logical errors and generating tags.

[0033] Constraint compliance verification is used to determine whether the initial domain knowledge report meets external constraints or industry standards, such as report structure specifications, source citation rules, terminology usage rules, or compliance constraints of the business domain. This verification process is based on the collaborative execution of rule matching, semantic constraint detection, and citation relationship tracing mechanisms. By comparing the report content with established standards, it identifies parts that violate structural rules, have missing citations, or semantic deviations, and generates compliance violation tags.

[0034] Both logical consistency and constraint compliance checks use a unified intermediate representation format, parsing the report text into a logical graph structure composed of semantic nodes and reference nodes, making the error detection process interpretable. When the check passes, the intermediate representation structure is mapped back to the report text, becoming the basis for the final domain knowledge report output, thereby ensuring that the generated content meets directly applicable standards in terms of logical integrity and compliance.

[0035] In one implementation, the system parses the initial domain knowledge report using a semantic analysis module, breaking the text down into semantic units and establishing a reference link for each unit. Subsequently, a consistency determination model detects inconsistencies or content jumps between semantic units; if an issue is found, a logical error flag is generated. In this implementation, logical consistency checks and constraint compliance checks are performed sequentially. Only if the logical consistency check is error-free does the constraint compliance check proceed, creating a hierarchical verification process.

[0036] In another implementation, logical consistency verification and constraint compliance verification are performed in parallel. The system maps the initial domain knowledge report to a dual-path verification structure: one path uses a semantic consistency network to detect the inference link, and the other path uses a compliance detection network to match external specification templates. The results of the two verifications are ultimately combined and judged in the result fusion module, and a result of passing or failing verification is generated based on a threshold. When verification passes, the initial domain knowledge report is directly output as the final domain knowledge report, and the report structure and reference link are recorded in the subsequent storage module for traceability.

[0037] In another implementation, constraint compliance verification can be adapted to different industry standards by loading different types of rule sets to achieve cross-industry adaptation. For example, medical guideline structure rules or regulatory material citation standards can be added to the rule set to ensure that the verification maintains consistent processing logic across application scenarios.

[0038] This embodiment, by performing logical consistency checks and constraint compliance checks, can promptly detect inference contradictions, citation errors, or normative deviations after the report is generated, and ensure that the final output content meets reliable standards in terms of logical structure and industry norms. This reduces the impact of content errors on subsequent decisions and enhances the credibility and professionalism of automatically generated reports.

[0039] In one embodiment, step S10 includes: S101 receives multi-source domain data from different data sources; S102, perform data preprocessing on the multi-source domain data to obtain a preprocessed set of document fragments; S103, The preprocessed document fragment set is subjected to entity recognition using a named entity recognition model to form an entity set; S104, Use a relation extraction model to extract relations from the entity set to obtain a set of relations between entities; S105, construct a domain knowledge graph based on the entity set and the entity relationship set, so that the entities are the nodes of the domain knowledge graph and the entity relationships are the edges of the domain knowledge graph; S106, Store the domain knowledge graph in shared working memory.

[0040] In this embodiment, when receiving multi-source domain data from different data sources, the data source can be explicitly understood as a combination of structured storage systems, semi-structured document libraries, and unstructured text collections, such as record files exported from business databases, policy document collections, medical process description text collections, and question-and-answer interaction logs. The term "multi-source domain data" refers to a collection of raw information from multiple sources, in multiple formats, and covering multiple thematic ranges, providing material for subsequent identification of domain entities and relationships. During the receiving process, connection parameters and access protocols can be configured for different data sources through the data access adaptation module, and newly added or updated multi-source domain data can be uniformly imported into the processing pipeline according to a preset retrieval cycle or event-triggered mechanism.

[0041] Data preprocessing transforms multi-source domain data into a unified representation that is easily adapted for subsequent text analysis and structure extraction. Format standardization unifies different file encodings, tag formats, and structural labels. For example, it converts hypertext tags, rich text tags, and table field content into unified text fields and auxiliary tag fields, eliminating structural breaks caused by format differences. Noise removal removes content that significantly interferes with entity recognition and relation extraction, such as meaningless symbol sequences in logs, template headers and footers, duplicate advertising fragments, blank lines, and excessively long punctuation sequences. Redundant information is removed through regular expression matching, threshold filtering, and template recognition. Document segmentation divides long texts into appropriately sized and semantically complete units based on natural paragraphs, heading levels, or semantic breakpoints, forming a preprocessed set of document segments. This allows subsequent models to identify entities and relationships within a stable context window, avoiding semantic dilution caused by excessively large document scopes.

[0042] When the preprocessed set of document fragments enters the named entity recognition model, each document fragment is encoded into an input representation adapted to the model. This is achieved, for example, by generating a token sequence through word segmentation, sub-word splitting, or character encoding. This token sequence is then mapped to a vector space via an embedding layer, allowing the named entity recognition model to capture semantic features within the vector space. The goal of the named entity recognition model is to identify entity boundaries and entity categories related to domain concepts. This entity set includes, but is not limited to, product names, customer categories, and risk types in business domains, as well as drug names, symptom descriptions, and examination item names used in healthcare. The model can employ sequence labeling, pointer structures, or span classification structures. Through training, it obtains the output distribution of each label within the context, thereby labeling the position and type of each entity in the document fragment. The entity set aggregates entity labels from all fragments, mapping identical entity strings or semantically similar entities to a unified entity identifier. This is achieved, for example, by eliminating morphological differences through string normalization and alias merging mechanisms.

[0043] When inputting an entity set into a relation extraction model, features such as the context of entities within a document fragment and the distance and dependencies between entity pairs need to be encoded as input to the relation discrimination model. The relation extraction model is responsible for identifying semantic relationships between entities, such as "membership," "causality," "constraint," and "scope of application." The entity relation set consists of entity pairs and corresponding relation categories. In implementation, this can be achieved by constructing a context vector representation for each entity pair, concatenating the entity's start and end positions with the fragment's context, and then inputting it into a classification network. Alternatively, graph structure encoding can be used, treating words and entities in a sentence as nodes in a graph, propagating contextual features through a graph neural network, and then predicting relation categories between entity node pairs. The entity relation set retains entity identifiers and records relation direction, relation type, and relation occurrence location, facilitating the construction of subsequent graph structures.

[0044] When constructing a domain knowledge graph based on entity sets and sets of relationships between entities, entities can be mapped to nodes in the graph, and relationships between entities can be mapped to edges. Nodes in the domain knowledge graph not only contain entity identifiers but can also be appended with attributes such as entity category, source document identifier, frequency of occurrence, and domain label, for subsequent filtering and focusing across different business domains. Edges in the graph, in addition to relationship type, can also contain information such as relationship confidence, occurrence location statistics, and triggering document fragment identifiers, enabling the graph to not only express static connections between concepts but also reflect the strength of support for these connections in the corpus. The construction process involves traversing the set of relationships between entities, adding each entity pair to its corresponding two entity nodes in the graph structure. If a node already exists, its attributes are updated; otherwise, a new node is created, and an edge of the corresponding relationship type is added between the two nodes, forming a domain knowledge graph with a topological structure.

[0045] When storing domain knowledge graphs in shared working memory, a data representation format adapted for multi-agent access needs to be designed for the graph structure. Shared working memory can be understood as a state storage space supporting multiple access modes, capable of preserving the complete graph structure while providing interfaces for retrieving related information by node, edge, local subgraph, or document fragment. During the storage process, the domain knowledge graph can be indexed, such as creating inverted indexes for entity identifiers, filtering indexes for relation types, and aggregation indexes for local topologies. This facilitates subsequent access to the same graph content from different perspectives by agents—reading agents, writing agents, and so on. The domain knowledge graph in shared working memory can also be associated with subsequently written structured summaries and reading path histories, creating a unified contextual environment where the graph structure, text summaries, and agent behaviors support subsequent graph-aware selection, semantic aggregation, and report generation.

[0046] This embodiment receives multi-source domain data from different data sources, performs data preprocessing including format standardization, noise removal, and document segmentation, and then uses named entity recognition and relation extraction models to construct a domain knowledge graph composed of nodes and edges and stores it in shared working memory. This can form a well-structured and topologically queryable knowledge representation in a large-scale heterogeneous text environment, providing a unified, complete, and traceable data foundation for subsequent graph-aware selection and semantic aggregation by multiple agents in the shared working memory. This improves the accuracy and efficiency of cross-document knowledge extraction and provides structured support for subsequent report generation and compliance verification.

[0047] In one embodiment, step S20 above includes: S201, using a selected agent to extract topological features from the domain knowledge graph, the topological features including node centrality features and edge weight features; S202, the selected agent obtains the historical reading status from the shared working memory, the historical reading status including the record of read document fragments and the reading time sequence; S203, Based on the node centrality feature, determine the centrality value of each document fragment in the document fragment set; S204, Based on the query task and document fragment content, determine the relevance score of each document fragment in the document fragment set; S205, Based on the centrality value and the relevance score, determine the initial reading priority of each document fragment in the document fragment set; S206, Based on the read document fragment records, filter the read document fragments in the document fragment set to obtain unread document fragments; S207, Based on the dependency relationship between the edge weight features and the document fragments, adjust the initial reading priority of the unread document fragments to obtain the adjusted reading priority; S208, based on the adjusted reading priority, select the document segment with the highest reading priority from the unread document segments as the target document segment.

[0048] In this embodiment, when using the selection agent to determine target document fragments based on the topological features of the domain knowledge graph and the historical reading status in shared working memory, it is first necessary to establish an access interface to the domain knowledge graph within the selection agent. This allows the selection agent to obtain topological information from the graph structure that can be used to filter document fragments. The domain knowledge graph consists of entity nodes and edges representing relationships between entities. Nodes represent entities extracted from multi-source domain data, and edges represent the relationships between entities. In this structure, topological features encompass the connection patterns of nodes in the graph, the distribution of their neighbors, and their position within the overall structure. Node centrality features characterize the importance of a node in the graph. This can be determined by statistically analyzing the number of connections, connection levels, or path coverage between each entity node and other entity nodes, thus reflecting the influence range of document fragments related to that entity in the overall semantic network. Edge weight features characterize the strength of relationships between entities. They can be weighted based on the frequency of co-occurrence of entities, the semantic relevance of the context, the authority of the source document, or manually labeled trust markers, allowing the relationships between different entities to reflect differentiated strengths in the graph structure.

[0049] When selecting an agent to retrieve historical reading states from shared working memory, it needs to locate records related to document reading activities. Historical reading states include records of read document segments and reading time series. Records of read document segments can be maintained as a list of document segment identifiers, with each entry associated with a unique identifier for the document segment, the reading round, and the query task identifier that triggered the reading action. The reading time series describes the distribution of reading activities over time, recording the moment each reading action occurred using a timestamp sequence, thus reflecting the frequency and age of document segment access. When accessing shared working memory, the agent loads these historical reading states into an internal state vector or internal table structure for subsequent filtering of unread document segments and control of reading frequency.

[0050] When determining the centrality value of each document fragment in a document fragment set, the selection agent needs to establish a mapping relationship between document fragments and nodes in the domain knowledge graph. Each document fragment in the document fragment set is typically associated with several entity identifiers. The selection agent can search for the node centrality features of these entity nodes in the domain knowledge graph based on the set of entity nodes involved in the document fragment, and generate the corresponding centrality value of the document fragment through aggregation. For example, the node centrality features of multiple entity nodes involved in a document fragment can be weighted and accumulated or sorted for selection, so that document fragments associated with more entities with high centrality obtain higher centrality values. In this way, each document fragment corresponds to a centrality value that reflects structural importance, providing a structural basis for subsequent priority evaluation.

[0051] When determining the relevance score of each document fragment in a document fragment set based on the query task and document fragment content, it is necessary to consider both the current query requirement and the textual information within the document fragment. The query task can be a user-input question, a system-initiated analysis request, or an information requirement expressed by a downstream application, such as analyzing a specific regulatory topic, a specific disease type, or a specific product family. Document fragment content refers to the text content, structural tags, and entity annotations extracted from the preprocessed document fragment set. The agent selection can use text matching models, semantic encoding models, or vector retrieval engines to map the query task and document fragment content into vector representations, and determine the relevance score based on the similarity between vectors. Alternatively, it can combine entity overlap, domain keyword coverage, and syntactic dependency structures to assign a quantified relevance evaluation value to each document fragment. In this way, each document fragment in the document fragment set has a relevance score in the current query context, which measures its necessity for selection in the current task.

[0052] When determining the initial reading priority of each document fragment in a document fragment set based on centrality and relevance scores, the agent selection needs to integrate the two dimensions of structural importance and task relevance into a single ranking metric. Centrality represents the structural position of a document fragment within the overall knowledge network, while relevance score represents the degree of matching of a document fragment in the current query. The agent selection can combine these two dimensions into an initial reading priority through weighted fusion, ranking merging, or multidimensional ranking functions, prioritizing document fragments that are important in the graph structure and closely related to the query task. In this way, each document fragment is assigned an initial reading priority for subsequent filtering and adjustment.

[0053] When filtering read document fragments from the document fragment set based on already read document fragment records, and identifying unread document fragments, the agent needs to retrieve read records from its historical reading history and match the identifiers in the document fragment set with the record list. Document fragments that successfully match are considered processed and removed from the current candidate set; their corresponding initial reading priority is not included in the next round of evaluation. Document fragments not appearing in the read records are aggregated into an unread document fragment set, retaining their corresponding initial reading priority for subsequent dependency adjustments. This filtering process avoids repeatedly selecting the same document fragments, improving the coverage of new information during reading.

[0054] When adjusting the initial reading priority of unread document fragments based on edge weight features and dependencies between document fragments to obtain the adjusted reading priority, it is necessary to first construct the dependencies between document fragments. Document fragment dependencies can come from connections between entities in the domain knowledge graph, or from reference relationships, succession relationships between clauses, or cross-document reference relationships. For example, when an entity in a document fragment has a strong edge weight feature with another entity in the graph, and the document fragment corresponding to the latter has been marked as a key reference source, then the two fragments can be marked as having a dependency relationship. The agent can construct such a dependency structure within the set of unread document fragments, using edge weight features as a dependency strength reference, and then adjust the initial reading priority according to the dependency strength. Unread document fragments that are on key dependency chains or related to multiple read key fragments are given higher priority, while fragments with lower dependency are given lower priority, thereby generating an adjusted reading priority that better conforms to the overall understanding order.

[0055] When selecting the highest-priority document fragment from unread document fragments based on adjusted reading priorities as the target document fragment, the selection agent sorts the unread document fragments according to the adjusted reading priorities and outputs the highest-priority fragment to the reading agent. The target document fragment then becomes the input for the next round of semantic encoding and summarization, allowing the reading behavior to extend along a more reasonable path under the dual constraints of graph structure and historical reading states, covering key structural content while minimizing redundant information.

[0056] This embodiment utilizes a selected agent to extract topological features from a domain knowledge graph and combines them with historical reading states in shared working memory. It then performs centrality evaluation, task relevance evaluation, reading history filtering, and dependency-driven priority adjustment on a set of document fragments. This allows for the selection of structurally important, task-related, and unread target document fragments from numerous candidate fragments. This guides subsequent reading agents to gradually expand their understanding along more logical semantic paths, improving the efficiency and coherence of cross-document knowledge acquisition. Simultaneously, it provides a well-structured and clearly defined document input sequence for subsequent report generation.

[0057] In one embodiment, step S30 above includes: S301, use the language model in the reading agent to perform semantic encoding on the target document fragment to obtain a semantic vector; S302, Based on the semantic vector, extract key information from the target document fragment; S303, Generate a structured summary based on the key information; S304, Write the structured summary into the shared working memory.

[0058] In this embodiment, after receiving a target document fragment, the reading agent uses a semantic encoding unit centered on a language model to perform hierarchical representation modeling of the text content of the fragment. The semantic encoding unit first breaks down the target document fragment into a sequence of basic semantic units, maps each semantic unit to an initial vector using a word embedding matrix, and captures semantic relationships across sentences and paragraphs using a multi-head attention mechanism. A context-dependent semantic vector representation is generated through a self-attention stacked structure. This semantic vector contains not only local representations but also global dependency features across paragraphs, enabling subsequent extraction processes to analyze key information within the fragment in a unified representation space.

[0059] After obtaining the semantic vector, the key information extraction unit of the reading agent performs structured parsing of the semantic vector. The key information extraction unit uses an entity recognition component to identify entity information in the target document fragment, including domain-specific entities, event-type entities, and constraint variables; it uses a logical constraint analysis component to extract logical dependencies, conditional constraints, and inference links from the semantic vector; and it uses a topic analysis component to identify the topic information of the fragment based on the clustering density features of the semantic vector. These three types of key information come from the same semantic vector processing flow and are directly extracted from the semantic encoding results, maintaining consistency in the information source.

[0060] After obtaining entity information, logical constraint information, and topic information, the structured summary generation unit reorganizes the three types of information according to a preset structured format. Entities are used as nodes in the structured summary, logical constraints are used as limiting conditions between nodes, and topic information is used as the main semantics of the summary. This ensures that the generated structured summary maintains a unified data format, making it easier to align and merge with other summary entries in shared working memory.

[0061] During the writing of structured summaries into shared working memory, a gated update mechanism controls the writing behavior. This mechanism consists of an input gate, a forget gate, and an update gate. By detecting the existing summary content, semantic density, and conflict level in the shared working memory, it determines the writing mode of the new summary. When the summary semantically overlaps with existing content, the update gate performs an overwrite or enhancement write; when the summary is independent of existing content, the input gate performs an append write; and when the summary is redundant or conflicts with existing content, the forget gate performs filtering and replacement operations. This gated update mechanism maintains a unified storage structure within the shared working memory, enabling subsequent agents to directly read consistent data.

[0062] This embodiment enables the reading agent to complete the entire processing path from text understanding to structured writing within a unified semantic representation space through the continuous collaboration of semantic encoding, key information extraction, structured summary generation, and a gating update mechanism. Key information extraction based on a unified semantic vector ensures that entity information, logical constraint information, and topic information remain semantically aligned. The structured summary generation method ensures that the data in the shared working memory has a standardized format. The gating update mechanism dynamically adjusts the memory content during the writing phase, avoiding redundant accumulation and semantic conflicts, and maintaining the compactness and effectiveness of the shared working memory through an adaptive writing strategy.

[0063] In one embodiment, step S40 above includes: S401, using a writing agent to read memory content from the updated shared working memory; S402, use a cross-attention mechanism to perform semantic aggregation on the memory content to obtain an aggregated semantic representation; S403, Based on the aggregated semantic representation, generate text content through a generation model; S404, Based on the text content, generate an initial domain knowledge report according to a preset report structure.

[0064] In this embodiment, before generating the initial domain knowledge report, the writing agent first initiates a read request to the updated shared working memory. It matches the internal query vector with the index structure in the shared working memory to locate memory content associated with the current task objective. The shared working memory stores structured summaries written by the reading agent, domain knowledge graph information, and historical analysis results in a unified data format. During the reading phase, the writing agent combines memory content related to the same topic, business object, or regulatory clause into a memory content set to ensure that subsequent semantic aggregation covers multi-source information on the same knowledge chain. Simultaneously, the writing agent filters the memory content set according to the task context, removing entries with outdated timestamps or mismatched task tags, concentrating the memory content within the scope of information directly related to the current report generation requirements.

[0065] After retrieving the memory content, the cross-attention mechanism within the writing intelligence uses the memory content as keys and values, and the current task representation and decoding state as queries, to achieve semantic aggregation through attention weight calculation. The cross-attention mechanism maps each memory content to a vector space, calculates the similarity between the query vector and each memory vector to obtain attention weights, and then uses these attention weights to weightedly combine the memory vectors. In this process, for memory content that appears multiple times in shared working memory and has a higher relevance to the current task, the attention weight is significantly increased, thus contributing a larger proportion to the aggregated semantic representation; for memory content with weak relevance or inconsistent themes, the attention weight is compressed, and its impact on the aggregation result is significantly weakened. The cross-attention mechanism not only weights different memory content in a single aggregation but can also perform aggregation operations multiple times in a multi-layered structure, allowing memory content at different levels to form hierarchical expressions in layer-by-layer aggregation, thereby obtaining an aggregated semantic representation with both a global perspective and local details.

[0066] After the aggregate semantic representation is formed, the writing agent uses it as input to the generative model, driving the model to generate text content. The generative model contains encoding and decoding units. The encoding unit receives the aggregate semantic representation and maps it to an internal context state vector. The decoding unit outputs the text content word by word or phrase by phrase at each generation point based on the context state vector, the generated text, and task control signals. During the generation process, the decoding strategy of the generative model can be controlled, such as limiting repeated segments, limiting length ranges, or introducing domain constraint labels, so that the generated text content can both cover the knowledge points embodied in the aggregate semantic representation and maintain coherence and professionalism in language expression. The generative model can support output configurations for different languages ​​or styles, thus adapting to the text expression requirements of different usage scenarios without changing the aggregate semantic representation.

[0067] After the text content is generated, the writing agent maps it into an initial domain knowledge report based on a pre-defined report structure. This pre-defined report structure pre-defines the chapter order, paragraph hierarchy, and field layout, and may include structural elements such as an overview, topic analysis, clause interpretation, risk and opportunity analysis, and constraint and compliance explanations. Each structural element is associated with a set of semantic or topic tags. The writing agent performs content segmentation and tag matching on the generated text content, assigning sentences or paragraphs with a high degree of matching to a specific topic or semantic tag to their corresponding chapter positions, and arranging them sequentially according to the order set in the report structure. For multiple paragraphs under the same topic, the writing agent can sort these paragraphs based on their importance scores in the aggregate semantic representation, placing more important conclusive content at the beginning of the chapters, while supplementary explanations are placed in the latter half. Through this process of aligning the text content with the pre-defined report structure, an initial domain knowledge report is ultimately formed that structurally conforms to the expected template and closely revolves around the aggregate semantic representation in its content.

[0068] This embodiment utilizes a writing agent to read memory content from updated shared working memory, introduces a cross-attention mechanism to complete semantic aggregation, and generates text content through a generative model based on the aggregated semantic representation. Then, it organizes the text content according to a preset report structure to form an initial domain knowledge report. This achieves unified integration and controllable expression output of multi-source structured summaries and knowledge items, enabling the generated report to cover key information and relationships in memory while maintaining a stable chapter layout and content hierarchy in terms of chapter structure. This improves the completeness, consistency, and readability of the report content, providing a clear semantic and structural foundation for subsequent logical and compliance verification.

[0069] In one embodiment, step S50 above includes: S501, Perform a logical consistency check on the initial domain knowledge report and identify logical errors in the initial domain knowledge report; S502, Perform constraint compliance verification on the initial domain knowledge report and identify compliance violations in the initial domain knowledge report; S503, when a logical error or compliance violation is identified, a verification result containing error markers is generated, the verification result is sent to the writing agent, and the writing agent is triggered to regenerate the initial domain knowledge report. S504, when no logical errors or compliance violations are identified, the initial domain knowledge report is output as the final domain knowledge report.

[0070] In this embodiment, during the logical consistency verification of the initial domain knowledge report, the sentence structure, concept associations, and reasoning chains in the report are first formally reconstructed based on a parsing model, mapping the report content into an internally analyzable semantic graph structure. The semantic graph structure uses key concepts as nodes and causal, conditional, and inferential relationships between concepts as edges, representing the logical dependency paths contained in the text content. By detecting issues such as path closure, dependency loops, missing premises, and condition conflicts within the semantic graph, parts of the logical chain that do not satisfy the inferential relationships can be identified, such as undefined premises, contradictions between conclusions and facts, and incorrect order of condition judgments. Logical consistency verification relies on a set of rules or constraints, such as reasoning rules, chronological order rules, and entity attribute consistency rules, allowing logical errors to be accurately marked through rule matching and counterexample comparison.

[0071] After completing the logical consistency check, a constraint compliance check is performed. This check is based on a set of compliance constraints used for domain information extraction and judgment. These constraints may originate from industry standards, institutional rule sets, professional literature, standard entry descriptions, etc., and are used to define permitted description methods, data usage scope, inference boundaries, and sensitive information processing requirements. During the check, the entity and behavior descriptions and inference conclusions in the initial domain knowledge report are compared item by item with the compliance constraints. For example, it checks for entries that conflict with industry constraints, inappropriate extensions of sensitive information, content exceeding the scope of inference authorization, or descriptions that violate professional expression standards. For matched violations, a flag is generated, including the violation type, scope, and location.

[0072] When logical consistency checks or constraint compliance checks identify errors, a check result containing error markers is generated. This result includes the error category, the rule number that triggered the error, the corresponding text fragment index, and references to relevant logical paths or constraint entries, precisely indicating the location and nature of the error in the initial domain knowledge report. The check result is sent to the writing agent, which, upon receiving it, rereads the shared working memory and adjusts the text content based on the error markers during the generation process. This includes correcting inference chains, removing non-compliant descriptions, filling in missing premises, or adjusting expression. The mechanism that triggers the writing agent's regeneration ensures that the correction process for the initial domain knowledge report is not random but rather targeted based on the check results, ensuring that the output content meets all constraints at both the logical and compliance levels.

[0073] If neither logical consistency nor constraint compliance checks detect any errors, the system outputs the initial domain knowledge report directly as the final domain knowledge report. In this case, the report's semantic structure, fact chain, inference expressions, and domain constraints all satisfy all validation rules, thus it can be stored as reliable output or used further. The output of the final domain knowledge report may also trigger downstream processes, such as recording historical versions of the report, updating report citation information in shared working memory, or submitting the report for other agents to make decisions or inferences.

[0074] This embodiment performs logical consistency and constraint compliance checks on the initial domain knowledge report, and returns the check results containing error tags to the writing agent when errors are identified. This forms a cyclical text correction chain, achieving dual stability of the report content in terms of semantic logic and professional constraints. Logical checks ensure the reliability of the inference chain, compliance checks prevent content from exceeding the professional scope or violating domain rules, and error tags drive the writing agent to regenerate text, improving the accuracy, standardization, and consistency of the generated content. This ensures that the final domain knowledge report has a verifiable, traceable, and professionally applicable quality level.

[0075] In one embodiment, after step S50 above, the method further includes: S601, compare and analyze the references in the final domain knowledge report with the historical reading status in the shared working memory to obtain the comparison and analysis results; S602, Generate a feedback signal based on the comparative analysis results; S603, update the reading strategy parameters of the selected agent using the feedback signal; S604, the updated reading strategy parameters are stored in the shared working memory.

[0076] In this embodiment, after completing logical consistency and constraint compliance checks and outputting the final domain knowledge report, the system begins a comparative analysis process based on citations and historical reading status. Citations originate from citation information units within the final domain knowledge report. These citations are embedded in the report structure in the form of citation markers within the text, fact source indexes, cross-document association pointers, or citation path records, reflecting the external document fragments, knowledge nodes, or reasoning sources upon which the report content relied during its generation. Historical reading status is stored in shared working memory, containing records of document fragments previously selected by the chosen agent, reading order trajectories, time-series patterns, and coverage of read content. These two data sources represent two independent information chains: "which citations the model used" and "what content the system actually read."

[0077] The comparative analysis process first involves structured extraction of references and historical reading states, converting references into a set of reference nodes and historical reading states into a set of reading nodes. These two sets are mapped using a unified index structure, enabling comparison within the semantic and document spaces. Comparison methods can be based on various discrepancy calculations, such as set intersection and union, reference frequency difference, reading path coverage, the degree of matching between the reference chain and the reading chain, whether the reference source originates from unread segments, and whether the reading order matches the reference logic. These discrepancy calculations generate comparative analysis results, which typically include indicators such as uncovered references, low-frequency reading but high-frequency references, conflicts between reading and reference orders, and missing reference dependencies.

[0078] The process of generating feedback signals based on the comparative analysis results maps the aforementioned indicators into adjustments that allow selectable agents to update their reading behavior. Feedback signals can take the form of numerical vectors, differential weight distributions, or policy gradient signals, allowing the selected agents to directly use them for parameter updates. For example, when a high citation-low reading coverage scenario occurs, the feedback signal increases the base weight of the corresponding segment in priority calculation; when there is a conflict between reading order and citation order, the feedback signal adjusts the order preference factor in the policy parameters; when the citation chain contains unread nodes, the feedback signal increases the model's sensitivity to the continuity of dependency chains. The generation process of feedback signals ensures that the comparative analysis results can be encoded into a quantitative expression that can back-drive policy adjustments.

[0079] After receiving feedback signals, the selection agent updates its policy based on its internal policy parameter structure. Policy parameters typically include a set of parameters controlling reading selection behavior, such as reading priority calculation factors, dependency weight factors, centrality weight factors, reading coverage adjustment factors, and order preference factors. The update process replaces or shifts parameters according to the gradient or correction rules provided by the feedback signals, ensuring that the policy parameters better match the actual structure of the citation chain in the next round of target document selection, thus improving the matching degree between the reading content and the cited content. In this way, the system gradually optimizes its understanding of document knowledge chains during the reading cycle, enabling the selection agent to more accurately determine which document fragments have higher structural significance and usability.

[0080] The operation of storing the updated reading policy parameters into shared working memory creates a persistent global state, enabling all subsequent reading actions of selected agents to be executed based on a unified policy configuration. Shared working memory stores not only the reading state but also the policy parameters, giving policy updates both cross-round accumulation and cross-agent consistency. As the cycle of reading and report generation repeats, shared working memory gradually develops optimal reading path preferences across the entire knowledge graph, enabling the system to continuously learn and self-correct.

[0081] For example, the system as a whole includes five core modules: document preprocessing and knowledge graph construction module, multi-agent collaborative understanding module, graph perception selection strategy module, shared memory and semantic fusion module, and intelligent report generation and compliance verification module.

[0082] The document preprocessing and knowledge graph construction module includes: The system first receives insurance document inputs from multiple sources, including: regulatory policy documents, insurance terms and conditions, claims cases, market analysis reports, and internal operating documents.

[0083] Information extraction and entity recognition: The Named Entity Recognition (NER) model is used to identify entities such as "insurance product name", "risk type", "claims conditions" and "regulatory clauses".

[0084]

[0085] in This is a document fragment. The entity being extracted.

[0086] Relation Extraction and Knowledge Graph Generation: Using Relation Extraction Models Identify semantic relationships between entities (e.g., "constraints", "dependencies", "mutual exclusion risks"):

[0087] in , This indicates that two entities (or their vector representations) in a knowledge graph originate from different document fragments or different entities within the same fragment. This represents a relation extraction model used to determine the type of semantic association between two entities; Representing entities With entity The relationships between these elements, such as "constraints," "dependencies," and "mutually exclusive risks," can be either relationship labels or relationship vectors. This ultimately forms the domain knowledge graph. This provides structured support for subsequent graph-aware selection strategies. G represents the overall structure of the domain knowledge graph; E represents the set of entities in the graph, including those obtained through previous steps. All extracted entities, such as product names, risk types, etc.; R represents the set of relations in the graph, containing those obtained through... Extract all relationships between entities, such as constraints, dependencies, mutual exclusions, etc.

[0088] Among them, the multi-agent collaborative understanding module: The system employs three agents with different functions to collaborate on document parsing and generation tasks: A. Selector agent This agent is responsible for identifying the documents or document segments that should be prioritized for further reading. It is based on a graph structure. With reading history status Calculate reading priority:

[0089] in, This represents the i-th candidate document or document fragment; This indicates that the current reading history status is... When the system selects a document The probability of a document being selected for further reading. This value is output by softmax, and the sum of the probabilities of multiple candidate documents is 1. The reading history state vector at time t represents which documents have been read, the reading order, and the stage of understanding. It can be obtained by encoding RNN, Transformer state, or memory pool. This indicates that in the domain knowledge graph, it is related to candidate documents. Related subgraphs, such as local structures containing entity nodes and their relational edges involved in the document; Representing a pair of graphs The structural feature vector after graph feature extraction can contain metrics such as node centrality, node degree, edge weight, and clustering coefficient, which are used to measure the importance and connection patterns of the document in the overall knowledge graph. Indicates the structural features of the graph With reading history state vector The joint features obtained by concatenating vectors include both the current document's position in the knowledge graph and the current reading progress. The parameter matrix of the selector agent output layer maps the joint features into a scalar score or a multidimensional score vector for subsequent softmax normalization. softmax() represents the normalization function, which performs an exponential transformation and normalization on the scores of all candidate documents, transforming the original scores into a probability distribution for decision-making on "what to read next".

[0090] The selector's decision objective is to maximize coverage and information gain, enabling subsequent understanding and generation modules to grasp cross-document relationships.

[0091] B. Reader Intelligent Agent It is responsible for understanding the content of the currently selected document segment and extracting its core information. It employs a language model finely tuned for the financial field. Semantic encoding of documents:

[0092] in, This indicates the currently selected target document or document segment; The semantic encoding model used by the reading agent can be a pre-trained language model fine-tuned for fields such as finance and insurance, used to map the original text into a vector representation; Document The semantic vector representation, through the model The encoded vector compresses the semantic content of the document fragment, which can be used for subsequent key information extraction, structured summary generation, and writing to shared working memory. This is used to generate a structured summary (containing information such as topic, clause logic, and regulatory constraints), which is then written to the shared memory module.

[0093] C. Writer's Intelligent Agent After all document fragments have been fully read, the agent generates a structured report based on shared memory. The writer then utilizes a generative model. Generate target documents under the guidance of reinforced reward signals:

[0094] in, The aggregated result representing shared working memory is used to comprehensively represent the system's cumulative understanding throughout the reading process. This result can be calculated by multiple structured summaries, graph retrieval results, and contextual states through an aggregation network or attention mechanism. R represents the generative model used by the writing agent to generate complete natural language text based on the aggregated representation of shared working memory. It can be a decoding language model or an encoder-decoder structure. R represents the final report text output by the writing agent. In the original description, it includes insurance analysis reports, compliance statements, or market summaries, which can be regarded as a specific instance of "target domain knowledge report".

[0095] Among them, the graph-aware selection strategy module: To optimize the document selection order, two graph-aware strategies are proposed: Centrality-based priority strategy: Based on the centrality of document nodes in the knowledge graph. Adjust its reading priority:

[0096] in, This represents the current candidate document unit or document fragment. The system needs to select the object to be read next from multiple document fragments. Each object is denoted as . ; Indicates selecting a document fragment The priority or selection probability score. The symbol "∝" means "proportional to". In actual implementation, it is usually normalized over all candidate segments to become the true selection probability; Represents a document fragment In a domain knowledge graph, centrality is a numerical value that reflects the importance of the node corresponding to the segment in the graph structure, such as how many other entity nodes it is connected to, or whether it is located at the intersection of multiple paths. The larger the value, the more critical the structural position. Represents a document fragment The relevance score to the task query q is typically calculated using semantic matching models, vector similarity, or keyword matching. A higher score indicates that the content of that segment is closer to the current task requirements. q represents the task query content, describing the problem to be solved or the analysis topic, such as "solvency regulatory analysis" or "pension insurance risk comparison" as shown in the diagram. α represents the weighting coefficient, typically ranging from 0 to 1, used to balance "structural importance" and "task relevance." When α is large, segments with important structural positions in the diagram are prioritized; when α is small, segments highly relevant to the task query are prioritized.

[0097] Among them, the shared memory and semantic fusion module: Core information among multiple agents is stored and iteratively updated through shared working memory.

[0098] Memory unit update: Adopting a gated loop structure:

[0099] in, It represents the current state of the shared working memory, which corresponds to the latest content of the "shared working memory". It has already incorporated previously stored structured summaries, semantic representations and other information, which can be used for subsequent selection of agents and writing agents. It represents the shared working memory state at the previous moment, indicating the memory content before the current new information was accessed, which is equivalent to "old memory"; The vector representing the new information written to the shared working memory at the current moment can be understood as the encoded vector obtained by encoding the semantic representation such as the "structured summary" generated and organized by the reading agent in this round, and is used to update the shared working memory; Indicates to and The concatenation operation yields a joint vector, the purpose of which is to allow the updating unit to see both old memories and new content in a single computation, thereby learning "how much old to retain and how much new to absorb"; This represents updating the weight matrix, which consists of trainable parameters that act on the concatenated vector. The above is used to calculate the update gate signal. Intuitively, This determines how the system balances new and old information in different scenarios. For example, areas with frequent changes in regulatory policies may favor new information, while long-term stable rules tend to retain old information. () represents the Sigmoid activation function, which maps the input to a range of 0 to 1. Here, it is used to generate an "update gate" coefficient that controls the new information. and old memories The proportion in final memory ensures a smooth and continuous memory update process, preventing a single new input from completely overwriting the entire memory; The term "retention gate" is complementary to the update gate. When the update gate value is large, new information has a higher weight and old memory has a lower weight. When the update gate value is small, it tends to retain old memory content and reduce the disturbance caused by noisy summarization.

[0100] Semantic fusion: During the generation phase, the Writer Agent aggregates memory units using a cross-attention mechanism:

[0101] Where Q represents Query (query vector), which is given by the current generation state of the writing agent, such as the hidden state of the decoder when generating the next paragraph or the next sentence. Q represents "what information you want to retrieve" at the current writing position. This represents the shared working memory matrix at the current time t, continuously written to by the preceding reading agent. Rows or cells correspond to different memory entries, such as structured summaries of different document fragments, entity constraints, and key points of regulatory clauses. In this formula... , This means using shared working memory as both the key and value for attention; K represents the Key (set of key vectors), which here equals This indicates that the vector representations of each memory unit in the shared working memory are used as matching objects to calculate the relevance to the query vector Q, thereby obtaining the attention weight; V represents Value (the set of value vectors), which here is also equal to , represents the content itself that is weighted and aggregated. After the attention weights are calculated, these value vectors are weighted and summed to obtain the aggregated memory representation; Attention(·) represents the cross-attention calculation function, which can generally be understood as scaled dot product attention or an equivalent form: first calculate the similarity between Q and each K, then normalize to obtain the weight distribution, and then use the weights to perform a weighted summation on V; z represents the aggregated semantic representation, the output vector of Attention calculation, which is the comprehensive information "read" by WriterAgent from the shared working memory at the current generation time, used to guide the generation of the current sentence or paragraph, ensuring that the content is logically consistent with the existing memory and that the source of the reference is traceable.

[0102] Among them, the intelligent report generation and compliance verification module: Generation phase: The Writer Agent adds compliance constraint tags during decoding to ensure that the generated text complies with the insurance regulatory framework, such as the Insurance Law and the Solvency II Rules.

[0103] Compliance verification mechanism: The system automatically reviews the text after it is generated: Using rule matching + LLM compliance assessor Detect sensitive expressions and misleading inferences; Calculate the semantic consistency score between the nodes and the knowledge graph:

[0104] in, The semantic consistency score measures the semantic fit between the generated text and the reference nodes in the knowledge graph. It is calculated based on cosine similarity, and the value is usually normalized to fall in the range of [0,1]. The larger the value, the more consistent the generated text is with the reference knowledge. When the score is lower than the preset threshold, the action of "backtracking the corresponding memory node and regenerating" is triggered. cos(,) represents the cosine similarity operator, which measures the directional similarity between two vectors by the cosine value of the angle between them. Here, it is used to compare the closeness of two semantic vectors in the high-dimensional semantic space, thereby reflecting semantic consistency. To generate a semantic vector representation of the text, after the WriterAgent outputs text such as insurance analysis reports and compliance summaries, the system calls a semantic encoding model to map the currently generated text into a vector space, resulting in... This vector integrates the terms, factual relationships, and logical structure in the generated text, and is used for subsequent comparison with the reference vector of the knowledge graph; As a reference semantic vector, derived from domain knowledge graphs or predefined compliance benchmarks, the system selects knowledge graph nodes or normative entries (such as insurance clauses, regulatory requirements, risk definitions, etc.) corresponding to the currently generated content, and encodes them into vectors using the same semantic encoding model, forming... This can ensure and Cosine similarity is only comparable when the two entities are in the same semantic space.

[0105] Output content structure: The system can automatically generate three output types: "Comparative Analysis Report of Insurance Products", "Summary of Risk Factors and Compliance", and "Explanation of Relevance to Regulatory Policies".

[0106] This embodiment compares and analyzes the references in the final domain knowledge report with the historical reading states in shared working memory, and transforms the differences into feedback signals to update the reading strategy parameters of the selected agent. This allows the reading behavior to proactively converge towards the true structure of the reference chain, improving the matching degree between the reading path and the content references. Through this continuous strategy update mechanism, the system gradually optimizes the selection order and coverage of document fragments when repeatedly executing reading and report generation tasks, thereby improving the completeness of information collection, reference consistency, and the reliability of content generation. This ensures that the final output knowledge report forms a stable internal alignment relationship between the reference chain, reading path, and semantic structure.

[0107] In one embodiment, a report generation apparatus based on multi-agent cooperation is provided, which corresponds one-to-one with the report generation method based on multi-agent cooperation in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the report generation device based on multi-agent collaboration of the present invention. The modules include a domain knowledge construction module 10, an agent selection module 20, a reading agent module 30, a writing agent module 40, and a report verification module 50. Detailed descriptions of each functional module are as follows: The domain knowledge construction module 10 is used to receive multi-source domain data, perform entity recognition and relation extraction on the multi-source domain data, construct a domain knowledge graph based on the entity recognition and relation extraction results, and store the domain knowledge graph in shared working memory. The agent selection module 20 is used to determine the target document fragment by utilizing the agent selection based on the topological features of the domain knowledge graph and the historical reading state in the shared working memory; The reading agent module 30 is used to perform semantic encoding on the target document fragment using the reading agent, generate a structured summary, and write the structured summary into the shared working memory to update the shared working memory; The writing agent module 40 is used to generate an initial domain knowledge report by utilizing the writing agent to perform semantic aggregation based on the updated shared working memory. The report verification module 50 is used to perform logical consistency and constraint compliance verification on the initial domain knowledge report, and outputs the final domain knowledge report when the logical consistency and constraint compliance verification passes.

[0108] In one embodiment, the domain knowledge construction module 10 is specifically used for: Receive multi-source domain data from different data sources; The multi-source domain data is preprocessed to obtain a set of preprocessed document fragments; The preprocessed document fragment set is used to perform entity recognition using a named entity recognition model to form an entity set; The entity set is subjected to a relation extraction model to extract relations, resulting in a set of relations between entities. A domain knowledge graph is constructed based on the entity set and the set of relationships between entities, with entities as nodes and relationships between entities as edges. The domain knowledge graph is stored in shared working memory.

[0109] In one embodiment, the selected agent module 20 is specifically used for: The topological features of the domain knowledge graph are extracted by selecting an intelligent agent. The topological features include node centrality features and edge weight features. The selected agent retrieves historical reading status from the shared working memory, the historical reading status including records of read document segments and reading time series; Based on the node centrality feature, determine the centrality value of each document fragment in the document fragment set; Based on the query task and document fragment content, determine the relevance score of each document fragment in the document fragment set; Based on the centrality value and the relevance score, the initial reading priority of each document fragment in the document fragment set is determined; Based on the read document fragment records, the read document fragments in the document fragment set are filtered to obtain the unread document fragments; Based on the dependency relationship between the edge weight features and document fragments, the initial reading priority of the unread document fragments is adjusted to obtain the adjusted reading priority; Based on the adjusted reading priority, the document segment with the highest reading priority is selected from the unread document segments as the target document segment.

[0110] In one embodiment, the reading agent module 30 is specifically used for: The target document fragment is semantically encoded using a language model in the reading agent to obtain a semantic vector; Based on the semantic vector, key information is extracted from the target document fragment; Generate a structured summary based on the key information; The structured summary is written into the shared working memory.

[0111] In one embodiment, the writing agent module 40 is specifically used for: The writing agent retrieves memory content from the updated shared working memory. The memory content is semantically aggregated using a cross-attention mechanism to obtain an aggregated semantic representation. Based on the aggregated semantic representation, text content is generated through a generative model; Based on the text content, an initial domain knowledge report is generated according to a preset report structure.

[0112] In one embodiment, the report verification module 50 is specifically used for: Perform a logical consistency check on the initial domain knowledge report to identify logical errors in the initial domain knowledge report; Perform constraint compliance verification on the initial domain knowledge report to identify compliance violations in the initial domain knowledge report; When a logical error or compliance violation is detected, a verification result containing error markers is generated, the verification result is sent to the writing agent, and the writing agent is triggered to regenerate the initial domain knowledge report. If no logical errors or compliance violations are identified, the initial domain knowledge report is output as the final domain knowledge report.

[0113] In one embodiment, the report verification module 50 is specifically used for: The references in the final domain knowledge report are compared and analyzed with the historical reading status in the shared working memory to obtain the comparison and analysis results; A feedback signal is generated based on the comparative analysis results; The reading strategy parameters of the selected agent are updated using the feedback signal; The updated reading strategy parameters are stored in the shared working memory.

[0114] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When executed by the processor, the computer program implements the functions or steps of a server-side report generation method based on multi-agent cooperation.

[0115] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps of a report generation method based on multi-agent cooperation.

[0116] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Receive multi-source domain data, perform entity recognition and relation extraction on the multi-source domain data, construct a domain knowledge graph based on the entity recognition and relation extraction results, and store the domain knowledge graph in shared working memory; The target document fragment is determined by selecting an intelligent agent based on the topological features of the domain knowledge graph and the historical reading state in the shared working memory; The target document fragment is semantically encoded using a reading agent to generate a structured summary, and the structured summary is written into the shared working memory to update the shared working memory; The writing agent performs semantic aggregation based on the updated shared working memory to generate an initial domain knowledge report; The initial domain knowledge report is subjected to logical consistency and constraint compliance checks. When the logical consistency and constraint compliance checks pass, the final domain knowledge report is output.

[0117] In one embodiment, a non-volatile computer-readable storage medium is provided, which may be non-volatile or volatile, and stores a computer program thereon. When the computer program is executed by a processor, it performs the following steps: Receive multi-source domain data, perform entity recognition and relation extraction on the multi-source domain data, construct a domain knowledge graph based on the entity recognition and relation extraction results, and store the domain knowledge graph in shared working memory; The target document fragment is determined by selecting an intelligent agent based on the topological features of the domain knowledge graph and the historical reading state in the shared working memory; The target document fragment is semantically encoded using a reading agent to generate a structured summary, and the structured summary is written into the shared working memory to update the shared working memory; The writing agent performs semantic aggregation based on the updated shared working memory to generate an initial domain knowledge report; The initial domain knowledge report is subjected to logical consistency and constraint compliance checks. When the logical consistency and constraint compliance checks pass, the final domain knowledge report is output.

[0118] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0119] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0120] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

[0121] The user personal information involved in this application embodiment is all authorized (knowing and consenting) by the relevant parties or fully authorized by all parties, and the executing entity can obtain it through various open, legal and compliant means. The collection, storage, use, processing, transmission, provision and disclosure of the information, data and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals.

Claims

1. A report generation method based on multi-agent collaboration, characterized in that, Includes the following steps: Receive multi-source domain data, perform entity recognition and relation extraction on the multi-source domain data, construct a domain knowledge graph based on the entity recognition and relation extraction results, and store the domain knowledge graph in shared working memory; The target document fragment is determined by selecting an intelligent agent based on the topological features of the domain knowledge graph and the historical reading state in the shared working memory; The target document fragment is semantically encoded using a reading agent to generate a structured summary, and the structured summary is written into the shared working memory to update the shared working memory; The writing agent performs semantic aggregation based on the updated shared working memory to generate an initial domain knowledge report; The initial domain knowledge report is subjected to logical consistency and constraint compliance checks. When the logical consistency and constraint compliance checks pass, the final domain knowledge report is output.

2. The report generation method based on multi-agent cooperation as described in claim 1, characterized in that, The process includes receiving multi-source domain data, performing entity recognition and relation extraction on the multi-source domain data, constructing a domain knowledge graph based on the entity recognition and relation extraction results, and storing the domain knowledge graph in shared working memory, including: Receive multi-source domain data from different data sources; The multi-source domain data is preprocessed to obtain a set of preprocessed document fragments; The preprocessed document fragment set is used to perform entity recognition using a named entity recognition model to form an entity set; The entity set is subjected to a relation extraction model to extract relations, resulting in a set of relations between entities. A domain knowledge graph is constructed based on the entity set and the set of relationships between entities, with entities as nodes and relationships between entities as edges. The domain knowledge graph is stored in shared working memory.

3. The report generation method based on multi-agent cooperation as described in claim 1, characterized in that, By utilizing the topological features of the domain knowledge graph and the historical reading states in the shared working memory, the selected agent determines the target document fragment, including: The topological features of the domain knowledge graph are extracted by selecting an intelligent agent. The topological features include node centrality features and edge weight features. The selected agent retrieves historical reading status from the shared working memory, the historical reading status including records of read document segments and reading time series; Based on the node centrality feature, determine the centrality value of each document fragment in the document fragment set; Based on the query task and document fragment content, determine the relevance score of each document fragment in the document fragment set; Based on the centrality value and the relevance score, the initial reading priority of each document fragment in the document fragment set is determined; Based on the read document fragment records, the read document fragments in the document fragment set are filtered to obtain the unread document fragments; Based on the dependency relationship between the edge weight features and document fragments, the initial reading priority of the unread document fragments is adjusted to obtain the adjusted reading priority; Based on the adjusted reading priority, the document segment with the highest reading priority is selected from the unread document segments as the target document segment.

4. The report generation method based on multi-agent cooperation as described in claim 1, characterized in that, The target document fragment is semantically encoded using a reading agent to generate a structured summary, and the structured summary is written into the shared working memory to update the shared working memory, including: The target document fragment is semantically encoded using a language model in the reading agent to obtain a semantic vector; Based on the semantic vector, key information is extracted from the target document fragment; Generate a structured summary based on the key information; The structured summary is written into the shared working memory.

5. The report generation method based on multi-agent cooperation as described in claim 1, characterized in that, The writing agent performs semantic aggregation based on updated shared working memory to generate an initial domain knowledge report, including: The writing agent retrieves memory content from the updated shared working memory. The memory content is semantically aggregated using a cross-attention mechanism to obtain an aggregated semantic representation. Based on the aggregated semantic representation, text content is generated through a generative model; Based on the text content, an initial domain knowledge report is generated according to a preset report structure.

6. The report generation method based on multi-agent cooperation as described in claim 1, characterized in that, The initial domain knowledge report undergoes logical consistency and constraint compliance checks. Upon successful logical consistency and constraint compliance checks, a final domain knowledge report is output, including: Perform a logical consistency check on the initial domain knowledge report to identify logical errors in the initial domain knowledge report; Perform constraint compliance verification on the initial domain knowledge report to identify compliance violations in the initial domain knowledge report; When a logical error or compliance violation is detected, a verification result containing error markers is generated, the verification result is sent to the writing agent, and the writing agent is triggered to regenerate the initial domain knowledge report. If no logical errors or compliance violations are identified, the initial domain knowledge report is output as the final domain knowledge report.

7. The report generation method based on multi-agent cooperation as described in claim 1, characterized in that, After performing logical consistency and constraint compliance checks on the initial domain knowledge report, and outputting the final domain knowledge report upon successful logical consistency and constraint compliance checks, the process further includes: The references in the final domain knowledge report are compared and analyzed with the historical reading status in the shared working memory to obtain the comparison and analysis results; A feedback signal is generated based on the comparative analysis results; The reading strategy parameters of the selected agent are updated using the feedback signal; The updated reading strategy parameters are stored in the shared working memory.

8. A report generation device based on multi-agent collaboration, characterized in that, The report generation device based on multi-agent cooperation includes: The domain knowledge construction module is used to receive multi-source domain data, perform entity recognition and relation extraction on the multi-source domain data, construct a domain knowledge graph based on the entity recognition and relation extraction results, and store the domain knowledge graph in shared working memory; The agent selection module is used to determine the target document fragment by utilizing the agent selection based on the topological features of the domain knowledge graph and the historical reading state in the shared working memory; The reading agent module is used to perform semantic encoding on the target document fragment using a reading agent, generate a structured summary, and write the structured summary into the shared working memory to update the shared working memory; The writing agent module is used to generate an initial domain knowledge report by leveraging the writing agent to perform semantic aggregation based on the updated shared working memory. The report verification module is used to perform logical consistency and constraint compliance verification on the initial domain knowledge report, and outputs the final domain knowledge report when the logical consistency and constraint compliance verification passes.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a multi-agent cooperative report generation program stored in the memory and executable on the processor, wherein the multi-agent cooperative report generation program, when executed by the processor, implements the steps of the multi-agent cooperative report generation method as described in any one of claims 1-7.

10. A non-volatile computer-readable storage medium, characterized in that, The storage medium stores a report generation program based on multi-agent cooperation, which, when executed by a processor, implements the steps of the report generation method based on multi-agent cooperation as described in any one of claims 1-7.

Images