An Enhanced Enterprise Internal Information Retrieval System and Method Based on Local Graph Structure

Abstract

This invention provides an enhanced enterprise internal information retrieval system and method based on a local graph structure, relating to the field of natural language processing technology. Through a dual-path hybrid retrieval mechanism, and addressing the differentiated characteristics of clause-based and comprehension-based multi-hop questions in enterprise policy documents, this invention designs a retrieval strategy that combines a global decomposition path and an incremental iteration path. The global decomposition path achieves broad coverage by pre-decomposing sub-problems, while the incremental iteration path employs a bootstrapping closed loop of "retrieval-judgment-re-retrieval" to achieve dynamic depth mining. The two approaches work together to balance comprehensiveness and relevance, solving the problem that a single retrieval strategy cannot simultaneously handle simple queries and complex reasoning.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an enhanced enterprise internal information retrieval generation system and method based on local graph structure. Background Technology

[0002] In recent years, artificial intelligence technology has made significant progress, especially in the field of natural language processing (NLP). Many domestic enterprises and research institutions have accumulated rich experience in the training, optimization, and application of large language models, and have launched a series of high-performance large language models and products. The application of large language models in enterprise internal information retrieval and question answering is gradually increasing. Many enterprises are beginning to build their own information databases and utilize large language models to provide intelligent information query or question-answering services to improve employee work efficiency and compliance.

[0003] Because the output of a large language model is closely related to the content of the enterprise's own information database, it is necessary to avoid the illusion problem as much as possible. The focus of the model's capabilities has shifted from emphasizing the breadth and comprehensibility of knowledge to the accuracy, analytical logic, and specific format of the model's output. Since large language models themselves have extremely strong contextual analysis and summarization capabilities, key issues affecting the effectiveness of implementation include how to provide the model with correct text segments, how to use the Chain-of-Thought (CoT) method to help the model handle the original questions, correct errors in output, and review formatting.

[0004] In terms of graph structure applications, Google's GraphRAG integrates graph structures and retrieval / question answering into a single framework. It segments raw text into chunks, and the large language model extracts entities and relationships from these chunks using pre-defined prompts. It also establishes communities between different chunks to connect them and stores summaries to improve retrieval speed. The core idea of ​​Google's GraphRAG framework is to transform raw text data into a structured knowledge graph for retrieval and reasoning, thereby enhancing the understanding and analytical capabilities of the large language model in complex problem scenarios. Regarding the RAG framework, Gao Y et al. proposed using knowledge graphs as an auxiliary tool, combined with the graph model DR.KNOWS to explain and summarize complex medical concepts. To further improve accuracy in question answering tasks, in addition to using RAG or knowledge graphs to acquire knowledge, CoT (Coding on Logic) is also needed to improve the logic. CoT can significantly enhance the model's reasoning ability.

[0005] However, the aforementioned graph-based retrieval enhancement methods still have certain limitations in practical applications. First, the method requires multiple rounds of information extraction and structured processing during the knowledge base construction phase, including text segmentation, entity relation extraction, knowledge graph construction, community detection, and community summary generation. Many of these steps typically rely on large language models, thus consuming significant computational resources and model call costs when processing large-scale text data. Second, because this method constructs a hierarchical knowledge graph structure, when knowledge base content is added or modified, the relevant text often needs to undergo entity relation extraction and other processing steps again. Simultaneously, the node connections, community partitioning results, and community summary content in the original graph structure may also change. Therefore, the system may need to re-execute part or even all of the index construction process, resulting in high knowledge base update costs.

[0006] Therefore, existing technologies suffer from fragmented retrieval, insufficient associative reasoning capabilities, and high knowledge update costs in complex multi-hop question-answering scenarios. Existing retrieval enhancement methods primarily rely on text vector retrieval, which struggles to effectively depict the structured relationships between knowledge points. In multi-hop queries, issues such as fragmented contextual information and incomplete reasoning chains easily arise. While graph-based retrieval enhancement methods (such as GraphRAG) can improve the analysis capabilities of complex problems by constructing a global knowledge graph, their index construction phase typically involves multiple processing steps, including entity relation extraction, community partitioning, and community summary generation. When the knowledge base is updated, the graph index often needs to be rebuilt, resulting in high computational costs and system maintenance overhead. Summary of the Invention

[0007] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing an enhanced enterprise internal information retrieval and generation system and method based on local graph structures. This system dynamically constructs or utilizes local graph structures related to the question during the retrieval phase to structure and organize multi-hop related information. Combined with the natural language understanding and generation capabilities of a large language model, it achieves efficient reasoning and accurate answers to complex queries. This method reduces the cost of global graph construction and maintenance while maintaining the reasoning capabilities of graph structures, improves knowledge base update efficiency and system response performance, and is therefore more suitable for intelligent question-answering scenarios requiring frequent data updates.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] On the one hand, the present invention provides an enhanced enterprise internal information retrieval generation system based on a local graph structure, including a problem decomposition module, a sub-problem queue module, a knowledge graph hybrid retrieval module, a graph database, a vector database, an evidence chain caching module, a verification module, a definition module, and a formatting module;

[0010] The problem decomposition module is used to analyze the sentence structure, keyword features, and entity mentions of the original input problem based on rule-based prompt templates and a lightweight classification model. Based on the analysis results, the problem is decomposed. If the problem contains a clear file name, chapter number, or clause number, it is determined to be a clause retrieval problem, the problem type is marked as SIMPLE, the complexity score is low, and no problem decomposition is performed. If the problem involves logical relationships, constraints, or requires cross-paragraph information integration, it is determined to be a comprehension-based multi-hop problem, the problem type is marked as COMPLEX, and a semantic decomposition operation is performed to decompose the problem into a series of sub-problems.

[0011] The sub-problem queue module is used to store the sequence of sub-problems decomposed by the problem decomposition module. Each sub-problem includes the original problem reference identifier, sub-problem text, expected retrieval entity type, and priority weight.

[0012] The knowledge graph hybrid retrieval module is used to receive question type identifiers and execute differentiated retrieval strategies for clause-based retrieval questions and comprehension-based multi-hop questions;

[0013] For clause retrieval questions, the key identifiers in the question, namely the document name, chapter number, and clause number, are used as query conditions. The graph state index is used for precise matching to locate the target clause node, extract the complete clause text, effective date, and revision record from the node attributes, and recall relevant context fragments through vector semantic matching to construct a single-hop evidence set.

[0014] For the understanding-based multi-hop problem, a dual-path parallel retrieval mechanism is initiated, with path A being the global decomposition path and path B being the incremental iteration path.

[0015] The global path decomposition process is as follows: traverse each sub-problem in the sub-problem queue, and perform text semantic retrieval and knowledge graph subgraph extraction and expansion respectively; then integrate all subgraphs, identify shared nodes and connecting edges across subgraphs, and construct a multi-hop candidate evidence set; if there is a logical relationship between subgraphs, mark the relationship in the evidence chain buffer module to form a preliminary evidence chain structure.

[0016] The incremental iteration path process is as follows:

[0017] Starting with the original input question, bootstrapping iterative reasoning and graph expansion are performed. In each iteration, the sub-question generated in the previous round is used as the query to perform semantic retrieval and recall new document fragments. New key entities are identified from the new document fragments, and neighborhood subgraphs are extracted from the knowledge graph using the new key entities as seeds. The newly acquired evidence in this round is accumulated and merged with historical evidence and stored in the evidence chain cache module. The definition module is called to determine whether the question has been sufficiently answered based on the merged evidence. If it is determined to be sufficient, the iteration terminates, the retrieval termination condition flag is set to "sufficiently resolved," and the formatting module generates the answer. If it is determined to be insufficient and the number of iterations has not reached the upper limit, the definition module is called to analyze the evidence gap, continue to generate the next sub-question, and store it in the sub-question queue module, and enter the next round of iteration. If the number of iterations reaches the upper limit, the retrieval termination condition flag is set to "upper limit reached," and the iteration terminates. If no new valid evidence is added in two consecutive iterations during the iteration process, the iteration terminates.

[0018] In the first iteration, the original question is used as the query, and semantic retrieval recalls the Top-K related document fragments. When merging evidence, the Top-K related document fragments are merged with the neighborhood subgraph.

[0019] The graph database is used to store the nodes and edge relationships of the enterprise's institutional knowledge graph, including entity nodes and relationship edges; entity nodes include institutional clauses, responsible entities, process nodes, and conditional constraints; relationship edges include reference relationships, preconditions, postconditions, and logical AND / OR relationships.

[0020] The vector database is used to store text blocks, which are sub-blocks formed by dividing the preprocessed text of the original rules and regulations documents.

[0021] The evidence chain caching module is used to store a set of candidate evidence in the multi-hop reasoning process, including textual evidence (original clause content), structural evidence (graph path), evidence source identifier, and evidence timeliness marker.

[0022] The verification module is used to evaluate the quality of candidate evidence in the evidence chain caching module.

[0023] The definition module is used to generate a natural language response draft by utilizing a large language model, combining retrieved textual evidence and graph paths, and to determine whether the question has been adequately answered.

[0024] The formatting module is used to structure the answers based on the scenario's specifications.

[0025] Furthermore, the method by which the problem decomposition module decomposes the comprehension-based multi-hop problem is as follows:

[0026] First, identify the core entity set E={e1,e2,...,e...} in the problem.n This includes the responsible party, the type of event, and the actions taken;

[0027] Secondly, analyze the semantic relations between entities R={r1,r2,...,r m} Identify logical structures, including conditional relationships, causal relationships, parallel relationships, and referential relationships;

[0028] Then, based on the entity-relationship structure, the original problem is decomposed into several subproblems Q={q1,q2,...,q k Each sub-problem corresponds to a specific reasoning sub-goal;

[0029] Finally, the sub-problems are stored in the sub-problem queue module according to their logical dependencies, and the priority weights of each sub-problem are set based on the semantic importance of the problem in the original sentence and the degree centrality of the nodes in the knowledge graph.

[0030] Furthermore, in the process of global decomposition path, the text semantic retrieval operation is as follows: encode the sub-question into a semantic vector, and recall Top-K relevant text fragments in the document vector library;

[0031] The subgraph extraction and expansion operation of the knowledge graph is as follows: using the key entities in the sub-problems as seed nodes, a breadth-first expansion with a limited depth is performed in the knowledge graph. The limited depth is dynamically adjusted according to the problem complexity, generating a local subgraph G. i =(V i E i ), where V i Let E be the set of adjacent nodes. i For the set of relation edges;

[0032] The iterative process of the incremental iteration path is as follows:

[0033] Round 1 iteration: Starting with the original problem Q o For the query, semantic vector retrieval is used to recall Top-K relevant document fragments D. (1) From D (1) Identify the key entity set E in the middle (1) In knowledge graphs, E (1) Extract a first-order neighborhood subgraph G from the seed node (1) Document evidence D (1) With subgraph G (1) Merge and construct the current evidence set. (1) =(D (1) G (1) The evidence is stored in the evidence chain cache module and marked as having accumulated round number 1; the definition module is called based on Evidence. (1) Determine whether the question has been adequately answered;

[0034] If the determination is sufficient, terminate the iteration, set the retrieval termination condition flag to "fully resolved", and enter the formatting module to generate an answer; if the determination is insufficient, the definition module analyzes the evidence gap and generates the next sub-question Sub-Q (1) , and store Sub-Q (1) in the sub-question queue module, and set the generation round mark to 2;

[0035] Iteration in the t-th (t≥2) round: Using the sub-question Sub-Q (t) generated in the (t - 1)-th round as the query, perform semantic retrieval to recall new document fragments D (t) ; Identify new key entities E (t) from D (t) , and use E (t) as the seed to extract the neighborhood subgraph G (t) in the knowledge graph; Cumulatively merge the evidence newly obtained in this round with the historical evidence, that is, Evidence (t) = Evidence (t-1) ∪ (D (t) , G (t) ), and store it in the evidence chain cache module;

[0036] Based on the merged evidence Evidence (t) , call the definition module to determine again whether the question has been fully answered;

[0037] If the determination is sufficient, set the retrieval termination condition flag to "fully resolved" and terminate the iteration; if the determination is insufficient and t < T_max, the definition module analyzes the evidence gap and continues to generate the next sub-question Sub-Q (t+1) , store it in the sub-question queue module, set the generation round mark to t + 1, and enter the next round of iteration; if the determination is insufficient and t = T_max, set the retrieval termination condition flag to "reached the upper limit" and terminate the iteration;

[0038] Among them, the evidence merging specifically includes:

[0039] Document evidence deduplication: Remove duplicates based on the text content hash value, and reorder according to relevance and timeliness;

[0040] Subgraph merging: Unify the node identifiers, integrate the newly added node and edge relationships, identify the entity associations across rounds, and expand the graph structure in the local subgraph cache;

[0041] Evidence chain update: Mark the source sub-question, the round to which it belongs, and the association relationship with the historical evidence of the newly added evidence in this round in the evidence chain cache module;

[0042] The method by which the definition module determines whether the question has been fully answered is:

[0043] Check whether the answer covers all sub-objectives of the question, whether it cites sufficient terms and conditions, and whether it clarifies the conditions, constraints, and exceptions.

[0044] If the evidence is deemed insufficient, the feedback is sent to the knowledge graph hybrid retrieval module, specifying the entities or relation types that need to be supplemented for retrieval, and a targeted supplementary retrieval is performed.

[0045] If the evidence is deemed sufficient, the final set of evidence and the draft response will be passed to the formatting module.

[0046] Furthermore, the specific method for the verification module to perform quality assessment is as follows:

[0047] First, calculate the semantic relevance score between the evidence and the question; encode the evidence text and the question text into vectors respectively, calculate the cosine similarity, set a similarity threshold θ1, and mark evidence below this threshold as low relevance;

[0048] Secondly, verify the integrity of the graph structure; check whether the chain of evidence forms a closed reasoning path, that is, whether there is a connected path from the question entity to the answer entity, and whether there are isolated evidence fragments, that is, text without structural association or structure without text support.

[0049] Then, the verification results are determined; if the semantic relevance scores are all higher than the threshold and the structure is complete, the verification is passed and the system enters the positioning module; if there is low relevance evidence or structural breakage, the re-retrieval mechanism is triggered to expand the depth of the knowledge graph path, increase the number of semantic recalls, or adjust the priority of sub-questions for re-retrieval, and return to the knowledge graph hybrid retrieval module to perform supplementary retrieval, with a maximum of N retries.

[0050] Furthermore, the structured processing of the formatting module includes:

[0051] Citation of clauses: Clearly indicate the name of the cited policy document, chapter number, and clause number in your answer;

[0052] Source Notes: List all content involved in the answer and explain the logical relationship between each part;

[0053] Additional logical explanation: For comprehension-based multi-hop problems, a brief explanation of the reasoning path is provided;

[0054] Timeliness reminder: Indicate the effective status of the cited clauses, i.e., currently in effect / revised / repealed.

[0055] On the other hand, the present invention also provides an enhanced generation method for enterprise internal information retrieval based on local graph structure, implemented through the aforementioned enhanced generation system for enterprise internal information retrieval based on local graph structure, comprising the following steps:

[0056] Step 1: After the user inputs a question, the question type is determined. Based on the rule-based prompt template and lightweight classification model, the sentence structure, keyword features, and entity mentions of the question are analyzed. If the question contains a clear document name, chapter number, or clause number, it is determined to be a clause retrieval question, the question type is set to SIMPLE, the complexity score is low, and the process proceeds to Step 3. If the question involves logical relationships, constraints, or requires cross-paragraph information integration, it is determined to be a comprehension-based multi-hop question, the question type is set to COMPLEX, and the process proceeds to Step 2.

[0057] Step 2: For complex questions identified as COMPLEX, perform semantic decomposition to break down the question into a series of sub-questions, and then proceed to Step 3; each sub-question contains the original question reference identifier, sub-question text, expected retrieval entity type, and priority weight;

[0058] Step 3: Knowledge Graph Hybrid Retrieval; Based on the question type identifier identified in Step 1 and the sub-questions decomposed in Step 2, execute differentiated retrieval strategies;

[0059] For clause retrieval questions, the key identifiers in the question, namely the document name, chapter number, and clause number, are used as query conditions. The graph state index is used for precise matching to locate the target clause node, extract the complete clause text, effective date, and revision record from the node attributes, and recall relevant context fragments through vector semantic matching to construct a single-hop evidence set.

[0060] For the understanding-based multi-hop problem, a dual-path parallel retrieval mechanism is initiated, with path A being the global decomposition path and path B being the incremental iteration path.

[0061] The global path decomposition process is as follows: traverse each sub-problem in the sub-problem queue, and perform text semantic retrieval and knowledge graph subgraph extraction and expansion respectively; then integrate all subgraphs, identify shared nodes and connecting edges across subgraphs, and construct a multi-hop candidate evidence set; if there is a logical relationship between subgraphs, mark the relationship in the evidence chain buffer module to form a preliminary evidence chain structure.

[0062] The incremental iteration path process is as follows:

[0063] Starting with the original input question, bootstrap iterative reasoning and graph expansion are performed. In each iteration, the sub-question generated in the previous round is used as the query to perform semantic retrieval to recall newly added document fragments. New key entities are identified from the newly added document fragments, and neighborhood subgraphs are extracted from the knowledge graph using the new key entities as seeds. The newly acquired evidence in this round is accumulated and merged with historical evidence and stored in the evidence chain cache module, and then step 4 is executed.

[0064] In the first iteration, the original question is used as the query, and semantic retrieval recalls the Top-K related document fragments. When merging evidence, the Top-K related document fragments are merged with the neighborhood subgraph.

[0065] Step 4: Verify the search results, i.e., evaluate the quality of candidate evidence in the evidence chain cache module; if the verification fails, increase the depth of the graph path, increase the number of semantic recalls, or adjust the priority of sub-questions, and return to Step 3 to search again; if the verification passes, proceed to the next step.

[0066] Step 5: Using a large language model, combine the retrieved textual evidence and graph paths to generate a natural language response draft. Based on the merged evidence, determine whether the question has been adequately answered. If adequately answered, terminate the iteration, set the retrieval termination condition flag to "adequately resolved," and proceed to Step 6. If inadequately answered and the iteration count has not reached the upper limit, analyze the evidence gap, generate the next sub-question, return to Step 3, and proceed to the next round of retrieval iteration. If the iteration count reaches the upper limit, set the retrieval termination condition flag to "upper limit reached," terminate the iteration, and proceed to Step 6. If no new valid evidence is added in two consecutive iterations, terminate the iteration and proceed to Step 6.

[0067] Step 6: Format the answer output; Based on the scenario's requirements, the answer is output after being structured.

[0068] Furthermore, the specific method for problem decomposition in step 2 is as follows:

[0069] First, identify the core entity set E={e1,e2,...,e...} in the problem. n This includes the responsible party, the type of event, and the actions taken;

[0070] Secondly, analyze the semantic relations between entities R={r1,r2,...,r m} Identify logical structures, including conditional relationships, causal relationships, parallel relationships, and referential relationships;

[0071] Then, based on the entity-relationship structure, the original problem is decomposed into several subproblems Q={q1,q2,...,q k Each sub-problem corresponds to a specific reasoning sub-goal;

[0072] Finally, the sub-problems are stored in the sub-problem queue module according to their logical dependencies, and the priority weights of each sub-problem are set based on the semantic importance of the problem in the original sentence and the degree centrality of the nodes in the knowledge graph.

[0073] Furthermore, in step 3, during the global path decomposition process, the text semantic retrieval operation is as follows: the sub-question is encoded into a semantic vector, and Top-K related text fragments are retrieved from the document vector library;

[0074] The subgraph extraction and expansion operation of the knowledge graph is as follows: using the key entities in the sub-problems as seed nodes, a breadth-first expansion with a limited depth is performed in the knowledge graph. The limited depth is dynamically adjusted according to the problem complexity, generating a local subgraph G. i =(V i E i ), where V i Let E be the set of adjacent nodes. i For the set of relation edges;

[0075] The iterative process of the incremental iteration path is as follows:

[0076] In the first iteration, the original problem Q was used. o For the query, semantic vector retrieval is used to recall Top-K relevant document fragments D. (1) From D (1) Identify the key entity set E in the middle (1) In knowledge graphs, E (1) Extract a first-order neighborhood subgraph G from the seed node (1) Document evidence D (1) With subgraph G (1) Merge and construct the current evidence set. (1) =(D (1) G (1) ), marking the cumulative round as 1;

[0077] In the t-th (t≥2) iteration, the sub-problem Sub-Q generated in the (t-1)-th iteration is used. (t) To retrieve the newly added document fragment D, perform semantic retrieval. (t) From D (t) Identifying new key entity E (t) , with E (t) Extracting neighborhood subgraph G from the knowledge graph for the seed (t) The newly acquired evidence in this round will be combined with historical evidence to form a comprehensive evidence pool. (t) =Evidence (t-1) ∪(D (t) G (t) );

[0078] Specifically, evidence consolidation includes:

[0079] Document evidence deduplication: Remove duplicates based on text content hash value, and reorder by relevance and timeliness;

[0080] Subgraph merging: unify node identifiers, integrate newly added nodes and edge relationships, identify entity associations across rounds, and expand the graph structure in the local subgraph cache;

[0081] Evidence chain update: In the evidence chain cache module, mark the source sub-problem, round to which the newly added evidence belongs, and its relationship with historical evidence;

[0082] The method for determining whether the question has been adequately answered in step 5 is as follows:

[0083] Check whether the answer covers all sub-objectives of the question, whether it cites sufficient terms and conditions, and whether it clarifies the conditions, constraints, and exceptions.

[0084] If the evidence is deemed insufficient, return to step 3, specify the entities or relation types that need to be supplemented for retrieval, and perform targeted supplementary retrieval.

[0085] If the evidence is deemed sufficient, the final evidence set and draft response are obtained, and step 6 is executed.

[0086] Furthermore, the specific method for quality assessment in step 4 is as follows:

[0087] First, calculate the semantic relevance score between the evidence and the question; encode the evidence text and the question text into vectors respectively, calculate the cosine similarity, set a similarity threshold θ1, and mark evidence below this threshold as low relevance;

[0088] Secondly, verify the integrity of the graph structure; check whether the chain of evidence forms a closed reasoning path, that is, whether there is a connected path from the question entity to the answer entity, and whether there are isolated evidence fragments, that is, text without structural association or structure without text support.

[0089] Then, the verification result is determined; the verification passes if the semantic relevance score is higher than the threshold and the structure is complete; if there is low relevance evidence or structural breakage, the verification fails; the retrieval can be retried up to N times.

[0090] Furthermore, the structuring process in step 6 includes:

[0091] Citation of clauses: Clearly indicate the name of the cited policy document, chapter number, and clause number in your answer;

[0092] Source Notes: List all content involved in the answer and explain the logical relationship between each part;

[0093] Additional logical explanation: For comprehension-based multi-hop problems, a brief explanation of the reasoning path is provided;

[0094] Timeliness reminder: Indicate the effective status of the cited clauses, i.e., currently in effect / revised / repealed.

[0095] The beneficial effects of adopting the above technical solution are as follows: The enterprise internal information retrieval enhancement generation system and method based on local graph structure provided by this invention, through a dual-path hybrid retrieval mechanism, designs a retrieval strategy that combines a global decomposition path and an incremental iteration path to address the differentiated characteristics of clause-based retrieval and comprehension-based multi-hop questions in enterprise policy documents. The global decomposition path achieves broad coverage by pre-decomposing sub-questions, while the incremental iteration path achieves dynamic depth mining through a bootstrapping closed loop of "retrieval-judgment-re-retrieval." The two work together to balance comprehensiveness and relevance of the retrieval, solving the problem that a single retrieval strategy cannot simultaneously handle simple queries and complex reasoning. For multi-hop questions, this invention introduces a dynamic sub-question generation mechanism in the incremental iteration path. The definition module autonomously determines the next hop retrieval direction based on the current evidence gap, rather than relying on a pre-set fixed list of sub-questions. This mechanism avoids invalid retrieval, ensures that each iteration accurately supplements missing evidence, and significantly improves the efficiency and evidence coverage of solving complex problems. In the mechanism of this invention, an evidence accumulation and merging strategy is designed. In each iteration, newly added document evidence and subgraph structures are deduplicated, correlated, and integrated with historical evidence to construct a cumulative evidence chain across iterations. Simultaneously, the sufficiency judgment of the Define module is used as the core termination condition, rather than simply relying on the number of iterations. This ensures that answer generation stops promptly when evidence is sufficient and actively expands the search when evidence is insufficient, achieving a dynamic balance between search depth and answer quality. Relevant fields are set during the search, including the current iteration count, maximum iteration threshold, and criteria for continuous lack of new evidence, to prevent infinite loop searches in evidence-sparse scenarios and ensure that the system returns valid answers or clearly marks information gaps within a predictable timeframe. Attached Figure Description

[0096] Figure 1 This is a schematic diagram of the structure of the enterprise internal information retrieval enhancement generation system based on local graph structure provided in the first embodiment of the present invention;

[0097] Figure 2 A flowchart of the enterprise internal information retrieval enhancement generation method based on local graph structure provided in the second embodiment of the present invention;

[0098] Figure 3 The flowchart illustrates the dual-path parallel retrieval process provided in the second embodiment of the present invention. Detailed Implementation

[0099] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0100] Example 1:

[0101] An enhanced enterprise internal information retrieval generation system based on local graph structure, such as Figure 1 As shown, it includes a problem decomposition module, a sub-problem queue module, a knowledge graph hybrid retrieval module, a graph database, a vector database, an evidence chain caching module, a verification module, a definition module, and a formatting module.

[0102] The question decomposition module analyzes the sentence structure, keyword features, and entity mentions of the original input question based on rule-based prompt templates and a lightweight classification model, and decomposes the question according to the analysis results. If the question contains a clear document name, chapter number, or clause number (such as "Article Twelve" or "Chapter Two"), it is identified as a clause retrieval question, the question type is set to SIMPLE, the complexity score is low, no question decomposition is performed, and it directly enters the knowledge graph hybrid retrieval module for precise retrieval. If the question involves logical relationships, constraints, or requires cross-paragraph information integration (such as "what obligations need to be fulfilled"), it is identified as a comprehension-based multi-hop question, the question type is set to COMPLEX, and semantic decomposition is performed to decompose the question into a series of sub-questions.

[0103] The method for decomposing the comprehension-based multi-hop problem is as follows:

[0104] First, identify the core entity set E={e1,e2,...,e...} in the problem. n This includes the responsible party (e.g., "department head"), the type of incident (e.g., "safety accident"), and the actions taken (e.g., "reporting obligation").

[0105] Secondly, analyze the semantic relations between entities R={r1,r2,...,r m}, identify logical structures such as conditional relationships, causal relationships, parallel relationships, and referential relationships;

[0106] Then, based on the entity-relationship structure, the original problem is decomposed into several subproblems Q={q1,q2,...,q k Each sub-question corresponds to a specific reasoning sub-goal, such as "What is the definition of a safety accident?" or "What are the responsibilities of a department head?"

[0107] Finally, the sub-problems are stored in the sub-problem queue module according to their logical dependencies, and the priority weights of each sub-problem are set based on the semantic importance of the problem in the original sentence and the degree centrality of the nodes in the knowledge graph.

[0108] The sub-problem queue module is used to store the sequence of sub-problems decomposed by the problem decomposition module. Each sub-problem contains the original problem reference identifier, sub-problem text, expected retrieval entity type, and priority weight.

[0109] The knowledge graph hybrid retrieval module is used to receive question type identifiers (including question type (clause-based / comprehension-based multi-hop), complexity score, and decomposition status flags) and execute differentiated retrieval strategies for clause-based questions and comprehension-based multi-hop questions.

[0110] For clause retrieval questions (SIMPLE), the key identifiers in the question, namely the document name, chapter number, and clause number, are used as query conditions. Precise matching is performed on the graph state index to locate the target clause node. The complete clause text, effective date, and revision record are extracted from the node attributes. At the same time, relevant context fragments (such as the clause's pre-conditions and subsequent exceptions) are recalled through vector semantic matching to construct a single-hop evidence set.

[0111] For the understanding multi-hop problem (COMPLEX), a dual-path parallel retrieval mechanism is initiated, with path A being the global decomposition path and path B being the incremental iteration path.

[0112] The process of global path decomposition is as follows:

[0113] Iterate through each subproblem q in the subproblem queue i The process involves performing text semantic retrieval and knowledge graph subgraph extraction and expansion, respectively. Text semantic retrieval encodes sub-questions into semantic vectors and retrieves Top-K relevant text fragments from the document vector library. Knowledge graph subgraph extraction and expansion uses key entities from the sub-questions as seed nodes and performs a limited-depth breadth-first expansion (depth d=3 in this embodiment) on the knowledge graph. The limited depth is dynamically adjusted according to the problem complexity, generating a local subgraph G. i =(V i E i ), where V i Let E be the set of adjacent nodes. i For the set of relation edges;

[0114] Then, all subgraphs are integrated to identify shared nodes and connecting edges across subgraphs, constructing a multi-hop candidate evidence set. If there are logical relationships between subgraphs (such as jointly referencing a clause or having conditional dependencies), the relationships are marked in the evidence chain buffer module to form a preliminary evidence chain structure.

[0115] The incremental iteration path process is as follows:

[0116] Starting with the original problem (not the subproblems after decomposition), perform bootstrap iterative reasoning and graph expansion.

[0117] Round 1 iteration: Starting with the original problem Q o For the query, semantic vector retrieval is used to recall Top-K relevant document fragments D. (1) From D (1) Identify the key entity set E in the middle(1) In knowledge graphs, E (1) Extract a first-order neighborhood subgraph G from the seed node (1) Document evidence D (1) With subgraph G (1) Merge and construct the current evidence set. (1) =(D (1) G (1) ), and store it in the evidence chain cache module, marking the cumulative round as 1.

[0118] Calling the definition module based on Evidence (1) Determine if the question has been adequately answered. If adequately answered, terminate the iteration, set the retrieval termination condition flag to "adequately resolved," and proceed to the formatting module to generate the answer. If inadequately answered, analyze the evidence gaps in the definition module and generate the next sub-question, Sub-Q. (2) (Clearly pointing to the missing information dimension, such as "what is the specific definition of X" or "what are the triggering conditions for Y"), Sub-Q (2) Store it in the subproblem queue module and set the generation round marker to 2.

[0119] Iteration round t (t≥2): The sub-problem Sub-Q generated in round t-1. (t) To retrieve the newly added document fragment D, perform semantic retrieval. (t) From D (t) Identifying new key entity E (t) , with E (t) Extracting neighborhood subgraph G from the knowledge graph for the seed (t) The newly acquired evidence in this round is combined with historical evidence to form Evidence. (t) =Evidence (t-1) ∪(D (t) G (t) The evidence is stored in the evidence chain cache module. Specifically, evidence merging includes:

[0120] Document evidence deduplication: Remove duplicates based on text content hash value, and reorder by relevance and timeliness;

[0121] Subgraph merging: unify node identifiers, integrate newly added nodes and edge relationships, identify entity associations across rounds, and expand the graph structure in the local subgraph cache;

[0122] Evidence Chain Update: In the evidence chain cache module, mark the source sub-problem, round, and relationship with historical evidence of the newly added evidence in this round.

[0123] Based on the merged evidence Evidence (t), call the definition module to determine again whether the question has been fully answered. If it is determined to be sufficient, set the retrieval termination condition flag to "fully resolved" and terminate the iteration. If it is determined to be insufficient and t < T_max (the preset maximum number of iterations, for example, set to 3 - 5 times, which can be specifically configured according to the complexity of the question), the definition module analyzes the evidence gap and continues to generate the next sub-question Sub-Q (t+1) , store it in the sub-question queue module, set the generation round mark to t + 1, and enter the next round of iteration. If it is determined to be insufficient and t = T_max, set the retrieval termination condition flag to "reached the upper limit" and terminate the iteration.

[0124] The graph database is used to store the node and edge relationships of the enterprise institution knowledge graph, including entity nodes and relationship edges; entity nodes include institution clauses, responsible parties, process nodes, and conditional constraints; relationship edges include reference relationships, preconditions, post-results, and logical AND / OR relationships.

[0125] The vector database is used to store text blocks, that is, the regular institution text obtained by preprocessing the original regular institution documents. For example, if the text is 100 tokens and each 50 tokens is set as a text block, then it is divided into two blocks.

[0126] The evidence chain cache module is used to store the candidate evidence sets in the multi-hop reasoning process, including text evidence (i.e., the original clause content), structural evidence (i.e., the graph path), evidence source identifier, and evidence timeliness mark.

[0127] The verification module is used to evaluate the quality of the candidate evidence in the evidence chain cache module. The specific method is as follows:

[0128] First, calculate the semantic relevance score between the evidence and the question; encode the evidence text and the question text into vectors respectively, calculate the cosine similarity, set the similarity threshold θ1 (such as 0.75), and mark the evidence with a similarity lower than this threshold as low relevant;

[0129] Second, verify the integrity of the graph structure; check whether the evidence chain forms a closed reasoning path (whether there is a connected path from the question entity to the answer entity), and whether there are isolated evidence fragments (text without structural association or structure without text support);

[0130] Then, determine the verification result; if the semantic relevance scores are all higher than the threshold and the structure is complete, the verification passes and enters the definition module; if there are low relevant evidences or structural breaks, trigger the re-retrieval mechanism, increase the depth of the graph path (d + 1), increase the semantic recall quantity (K × 1.5), or adjust the sub-question priority to re-retrieve, return to the knowledge graph hybrid retrieval module to perform supplementary retrieval, and retry at most N times (such as N = 2).

[0131] The definition module is used to leverage a large language model, combining retrieved textual evidence and graph paths, to generate a draft natural language response and determine whether the question has been adequately answered. The method for determining whether a question has been adequately answered is as follows:

[0132] Check whether the answer covers all sub-goals of the question (for multi-hop questions), whether it cites sufficient clauses (at least one directly relevant clause), and whether it clarifies the conditions, constraints, and exceptions;

[0133] If the evidence is deemed insufficient (e.g., there is no corresponding evidence for the key sub-targets or the conditional relationships are unclear), the feedback is sent to the knowledge graph hybrid retrieval module, specifying the entities or relationship types that need to be supplemented for retrieval, and a targeted supplementary retrieval is performed.

[0134] If the evidence is deemed sufficient, the final set of evidence and the draft response will be passed to the formatting module.

[0135] The formatting module is used to structure answers based on scenario-based specification requirements, including:

[0136] Citation of clauses: Clearly indicate the name of the cited document, chapter number, and clause number in your answer, in the format of "Article 12 of the Safety Management System (2024 Revised Edition)";

[0137] Source Notes: List all content involved in the answer and explain the logical relationship between each part (e.g., "Comprehensive analysis based on Chapter 2 Responsibility Definition and Chapter 5 Emergency Procedures").

[0138] Logical explanation supplement: For comprehension-based multi-step questions, add a brief explanation of the reasoning path (such as "Based on the definition of clause A → combined with the conditions of clause B → conclusion C is reached") to enhance the interpretability of the answer;

[0139] Timeliness reminder: Indicate the effective status of the cited clauses (currently in effect / revised / repealed) to avoid answering based on outdated regulations.

[0140] Example 2:

[0141] An enhanced enterprise internal information retrieval generation method based on local graph structure, such as Figure 2 As shown, the method of this embodiment is described below.

[0142] Step 1: After the user inputs a question, the question type is determined. Based on the rule-based prompt template and lightweight classification model, the sentence structure, keyword features, and entity mentions of the question are analyzed. If the question contains a clear document name, chapter number, or clause number (such as "Article 12," "Chapter 2"), it is determined to be a clause retrieval question, the question type is set to SIMPLE, the complexity score is low, and the process proceeds to Step 3. If the question involves logical relationships between multiple entities, conditional constraints, or requires cross-paragraph information integration (such as "what obligations need to be fulfilled"), it is determined to be a comprehension-based multi-hop question, the question type is set to COMPLEX, and Step 2 is executed.

[0143] Step 2: For complex questions identified as COMPLEX, perform semantic decomposition to break them down into a series of sub-questions, then proceed to Step 3. Each sub-question includes the original question reference identifier, sub-question text, expected retrieval entity type, and priority weight. The specific method for question decomposition is as follows:

[0144] First, identify the core entity set E={e1,e2,...,e...} in the problem. n This includes the responsible party (e.g., "department head"), the type of incident (e.g., "safety accident"), and the actions taken (e.g., "reporting obligation").

[0145] Secondly, analyze the semantic relations between entities R={r1,r2,...,r m}, identify logical structures such as conditional relationships, causal relationships, parallel relationships, and referential relationships;

[0146] Then, based on the entity-relationship structure, the original problem is decomposed into several subproblems Q={q1,q2,...,q k Each sub-question corresponds to a specific reasoning sub-goal, such as "What is the definition of a safety accident?" or "What are the responsibilities of a department head?"

[0147] Finally, the sub-problems are stored in the sub-problem queue module according to their logical dependencies, and the priority weights of each sub-problem are set based on the semantic importance of the problem in the original sentence and the degree centrality of the nodes in the knowledge graph.

[0148] Step 3: Knowledge Graph Hybrid Retrieval. Based on the question type identifiers determined in Step 1 (including question type (clause retrieval type / comprehension-based multi-hop type), complexity score, and decomposition status flags) and the sub-questions decomposed in Step 2, a differentiated retrieval strategy is executed.

[0149] For clause retrieval questions (SIMPLE), the key identifiers in the question, namely the document name, chapter number, and clause number, are used as query conditions. Precise matching is performed on the graph state index to locate the target clause node. The complete clause text, effective date, and revision record are extracted from the node attributes. At the same time, relevant context fragments (such as the clause's pre-conditions and subsequent exceptions) are recalled through vector semantic matching to construct a single-hop evidence set.

[0150] For the understanding-based multi-hop problem (COMPLEX), a dual-path parallel search mechanism is initiated, where path A is the global decomposition path and path B is the incremental iterative path, such as... Figure 3 As shown.

[0151] The process of global path decomposition is as follows:

[0152] Iterate through each subproblem q in the subproblem queue i The process involves performing text semantic retrieval and knowledge graph subgraph extraction and expansion, respectively. Text semantic retrieval encodes sub-questions into semantic vectors and retrieves Top-K relevant text fragments from the document vector library. Knowledge graph subgraph extraction and expansion uses key entities from the sub-questions as seed nodes and performs a limited-depth breadth-first expansion (depth d=3 in this embodiment) on the knowledge graph. The limited depth is dynamically adjusted according to the problem complexity, generating a local subgraph G. i =(V i E i ), where V i Let E be the set of adjacent nodes. i Let it be the set of relation edges.

[0153] Then, all subgraphs are integrated to identify shared nodes and connecting edges across subgraphs, constructing a multi-hop candidate evidence set. If there are logical relationships between subgraphs (such as jointly referencing a clause or having conditional dependencies), the relationships are marked in the evidence chain buffer module to form a preliminary evidence chain structure.

[0154] The incremental iteration path process is as follows:

[0155] In the first iteration, the original problem Q was used. o For the query, semantic vector retrieval is used to recall Top-K relevant document fragments D. (1) From D (1) Identify the key entity set E in the middle (1) In knowledge graphs, E (1) Extract a first-order neighborhood subgraph G from the seed node (1) Document evidence D (1) With subgraph G (1) Merge and construct the current evidence set. (1) =(D (1) G(1) ), marking the cumulative round as 1.

[0156] In the t-th (t≥2) iteration, the sub-problem Sub-Q generated in the (t-1)-th iteration is used. (t) To retrieve the newly added document fragment D, perform semantic retrieval. (t) From D (t) Identifying new key entity E (t) , with E (t) Extracting neighborhood subgraph G from the knowledge graph for the seed (t) The newly acquired evidence in this round is combined with historical evidence to form Evidence. (t) =Evidence (t-1) ∪(D (t) G (t) The consolidation of evidence specifically includes:

[0157] Document evidence deduplication: Remove duplicates based on text content hash value, and reorder by relevance and timeliness;

[0158] Subgraph merging: unify node identifiers, integrate newly added nodes and edge relationships, identify entity associations across rounds, and expand the graph structure in the local subgraph cache;

[0159] Evidence Chain Update: In the evidence chain cache module, mark the source sub-problem, round, and relationship with historical evidence of the newly added evidence in this round.

[0160] Step 4: Verify the search results, i.e., evaluate the quality of candidate evidence in the evidence chain caching module. If the verification fails, increase the graph path depth (d+1), increase the semantic recall quantity (K×1.5), or adjust the sub-question priority, and return to Step 3 to search again. If the verification passes, proceed to the next step. The specific method for quality evaluation is as follows:

[0161] First, calculate the semantic relevance score between the evidence and the question; encode the evidence text and the question text into vectors respectively, calculate the cosine similarity, and set a similarity threshold θ1 (e.g., 0.75). Evidence with a similarity lower than this threshold is marked as low relevance.

[0162] Secondly, verify the integrity of the graph structure; check whether the chain of evidence forms a closed reasoning path (whether there is a connected path from the question entity to the answer entity), and whether there are isolated evidence fragments (textual but no structural connection or structure but no textual support).

[0163] Then, the verification result is determined; the verification passes if the semantic relevance score is higher than the threshold and the structure is intact; if there is low relevance evidence or structural breakage, the verification fails. The retrieval can be retried up to N times (e.g., N=2).

[0164] Step 5: Use the large language model to combine the retrieved text evidence and the graph path to generate a draft of the natural language answer, and based on the merged evidence, judge whether the question is fully answered, check whether the answer covers all sub-goals of the question, whether sufficient clause bases are cited, and whether the conditional constraints and exceptions are clarified.

[0165] If it is determined to be sufficient, terminate the iteration, set the retrieval termination condition flag to "fully resolved", obtain the final evidence set and the draft answer, and proceed to Step 6; if it is determined to be insufficient (such as no corresponding evidence for key sub-goals, unclear conditional relationships) and the number of iterations has not reached the upper limit, analyze the evidence gap, continue to generate the next sub-question, specify the entity or relationship type that needs to be supplemented for retrieval, return to Step 3, perform directional supplementary retrieval, and enter the next round of retrieval iteration; if the number of iterations reaches the upper limit, set the retrieval termination condition flag to "reached the upper limit", terminate the iteration, obtain the final evidence set and the draft answer, and proceed to Step 6; if there is no new valid evidence in two consecutive iterations during the iteration process (convergence determination to avoid infinite loops), obtain the final evidence set and the draft answer, then terminate the iteration and enter Step 6.

[0166] In the first round of iteration, based on Evidence (1) Judge whether the question has been fully answered. If it is determined to be sufficient, terminate the iteration, set the retrieval termination condition flag to "fully resolved", and proceed to Step 6 to generate the answer; if it is determined to be insufficient, analyze the evidence gap and generate the next sub-question Sub-Q (2) (clearly pointing to the missing information dimension, such as "What is the specific definition of X" or "What are the triggering conditions of Y"), store Sub-Q (2) in the sub-question queue module, and set the generation round mark to 2.

[0167] In the t (t≥2) -th round of iteration, based on the merged evidence Evidence (t) , judge again whether the question has been fully answered. If it is determined to be sufficient, set the retrieval termination condition flag to "fully resolved", terminate the iteration, and proceed to Step 6 to generate the answer. If it is determined to be insufficient and t < T_max (the preset maximum number of iterations, for example, set to 3 - 5 times, which can be specifically configured according to the complexity of the question), analyze the evidence gap, continue to generate the next sub-question Sub-Q (t+1) , store it in the sub-question queue module, set the generation round mark to t + 1, and enter the next round of iteration. If it is determined to be insufficient and t = T_max, set the retrieval termination condition flag to "reached the upper limit", terminate the iteration, and proceed to Step 6 to generate the answer.

[0168] Step 6: Format and output the answer; based on the specification requirements of the scenario, perform structured processing on the answer and then output it, including:

[0169] Citation of clauses: Clearly indicate the name of the cited document, chapter number, and clause number in your answer, in the format of "Article 12 of the Safety Management System (2024 Revised Edition)";

[0170] Source Notes: List all content involved in the answer and explain the logical relationship between each part (e.g., "Comprehensive analysis based on Chapter 2 Responsibility Definition and Chapter 5 Emergency Procedures").

[0171] Logical explanation supplement: For comprehension-based multi-step questions, add a brief explanation of the reasoning path (such as "Based on the definition of clause A → combined with the conditions of clause B → conclusion C is reached") to enhance the interpretability of the answer;

[0172] Timeliness reminder: Indicate the effective status of the cited clauses (currently in effect / revised / repealed) to avoid answering based on outdated regulations.

[0173] The core of this embodiment is the design of a retrieval enhancement algorithm based on local graph structure expansion. This algorithm is a key technological improvement for enhancing evidence recall coverage and path integrity in multi-hop question answering scenarios. The core idea of ​​the algorithm is to achieve a balance between "breadth coverage" and "depth mining" through the synergy of two paths: global decomposition and incremental iteration.

[0174] The algorithm flow for global path decomposition is as follows:

[0175] Input: Subproblem queue Q = {q1, q2, ..., q} k}, knowledge graph G=(V,E), maximum expansion depth d_max.

[0176] For each subproblem q i ∈Q:

[0177] Entity recognition: from q i Extract the key entity set E i ={e i1 , e i2 , ...}, matching the corresponding nodes in the knowledge graph;

[0178] Seed initialization: E i The node in the diagram is labeled as seed node S. i Initialize the local subgraph G i =(S i , );

[0179] Neighborhood expansion: for the current subgraph G i For a boundary node, query its first-order neighbor node N(v) and associated edge E(v) in G, and add N(v) to V. i E(v) is added to E i ;

[0180] Depth control: Repeat the neighborhood expansion steps until the maximum depth d_max is reached or the number of newly expanded nodes is zero;

[0181] Document Recall: Based on Subgraph G i The node attributes (such as clause number, keywords) in the document library are used to retrieve relevant text fragments. i ;

[0182] Evidence integration: G i D i ) as a subproblem q i The candidate evidence is stored in the evidence chain buffer module.

[0183] After all sub-problems have been processed, perform cross-subgraph association analysis: identify different Gs. i If shared nodes exist, then an association edge between sub-problems is established in the evidence chain buffer module and marked as "logical association".

[0184] The algorithm flow for the incremental iteration path is as follows:

[0185] Input: Original question Q o Knowledge graph G, maximum number of iterations T_max, convergence threshold δ;

[0186] Initialization: t=1, seed node set S (1) =ExtractEntities(Q o The set of visited nodes, V_visited= ;

[0187] Iterative execution:

[0188] Subgraph expansion: with S (t) Using a seed, perform a neighborhood expansion with a restricted depth (d=1) in G to generate an incremental subgraph ΔG. (t) ;

[0189] Node update: ΔG (t) New nodes are added to V_visited, and nodes with expansion potential (degree > 1 and not yet fully expanded) are selected as seeds S for the next round. (t+1) ;

[0190] Document retrieval: Retrieve relevant document fragments D based on node attributes in V_visited. (t) ;

[0191] Sufficiency check: Call the defined module to determine whether the current evidence is sufficient to answer the question; if so, terminate early.

[0192] Convergence criterion: If |S (t+1)If | < δ (the number of new seeds is less than the threshold), then the iteration terminates;

[0193] Iteration control: t=t+1, if t>T_max then terminate, otherwise return to the subgraph expansion step.

[0194] Output: The subgraph accumulated during the iteration process With document evidence .

[0195] Dual-path fusion mechanism:

[0196] The global decomposition path and the incremental iteration path are executed independently, and the results are finally fused in the evidence chain buffer module.

[0197] Deduplication of evidence: Based on the hash value of the text content and the unique identifier of the graph node, duplicate evidence is removed;

[0198] Confidence score: Calculate a comprehensive score for each piece of evidence, taking into account semantic relevance, graph path length (shorter is more reliable), and number of discovery paths (higher confidence is when discovered by multiple paths);

[0199] Priority ranking: Sort by confidence score in descending order, and use high-confidence evidence to generate answers first;

[0200] Complementary and supplementary: If evidence found in one path is missing in another path, but the confidence level is higher than the threshold, it is retained to ensure comprehensive coverage.

[0201] To verify the effectiveness of the retrieval enhancement generation method based on local graph structure proposed in this embodiment in the enterprise system question-and-answer scenario, a knowledge base based on real rules and regulations documents was constructed, and a test set generated using GPT-4 was used for system evaluation.

[0202] Experimental conditions and methods:

[0203] Knowledge base construction: Collect 60 real corporate rules and regulations documents (including safety management system, financial reimbursement system, attendance management system, personnel appointment and dismissal system, etc.), with a total of 1433 clause nodes. Construct a knowledge graph through entity recognition and relationship extraction, containing 2010 entity nodes and 3955 relationship edges.

[0204] Test set generation: Based on the GPT-4 model, two types of test questions were generated for the knowledge base content: 80 clause retrieval questions (SIMPLE) and 120 comprehension-based multi-hop questions (COMPLEX), totaling 200 test questions. Among them, the multi-hop questions cover 2-hop reasoning (60 questions), 3-hop reasoning (40 questions), and complex reasoning with 4 hops or more (20 questions), involving typical scenarios such as cross-chapter references, combination of conditions and constraints, and tracing the responsible party.

[0205] Comparison method: Traditional vector retrieval RAG (Baseline-V) and standard GraphRAG (Baseline-G) are selected as comparison baselines. The method in this embodiment is referred to as Ours (dual-path hybrid mechanism).

[0206] Evaluation metrics: Answer accuracy (human judgment of whether the answer is correct), evidence recall (whether key supporting evidence is retrieved), average response time (Latency, the time from question input to answer output), and LLM call count (API Cost, the average number of large model calls per question).

[0207] The experimental results are shown in Table 1.

[0208] Compared with existing technologies, the dual-path hybrid retrieval mechanism proposed in this invention also demonstrates significant advantages in small- to medium-sized knowledge base scenarios. For COMPLEX type problems, the accuracy of this invention reaches 76.7%, an improvement of 36.1 percentage points compared to Baseline-V and 22.8 percentage points compared to Baseline-G. This indicates that even with a limited document size (60 documents), this invention, through the global decomposition of path A and the incremental iterative collaboration of path B, can still effectively capture implicit relationships across documents, solving the coverage problem of insufficient coverage of a single retrieval mechanism.

[0209] In terms of evidence recall, this invention achieved 83.2%, higher than Baseline-V (54.3%) and Baseline-G (67.8%). Especially in complex reasoning scenarios with 4 hops or more, Baseline-V suffers from broken evidence chains due to a lack of graph structure support, and Baseline-G introduces redundant nodes due to fixed-depth expansion. In contrast, this invention generates precise location of key evidence through dynamic sub-problems, achieving a recall rate of 78.6% in this scenario, significantly outperforming the comparison methods.

[0210] In terms of response efficiency, the average response time of this invention is 1.6 seconds, which is better than Baseline-G (2.1 seconds). Through sufficiency-driven termination control, this invention achieves termination in the first or second iteration in 81.3% of problems, avoiding invalid computation of the fixed-round strategy. At the same time, the number of LLM calls is reduced to 1.9 times, a 32% reduction compared to Baseline-G (2.8 times), effectively controlling computational costs in small to medium-scale scenarios.

[0211] Table 1 Comparison results of the three methods

[0212] Baseline-V 58.7% 85.8% 40.6% 54.3% 0.8s 1.2 Baseline-G 65.3% 82.5% 53.9% 67.8% 2.1s 2.8 Ours 81.3% 88.3% 76.7% 83.2% 1.6s 1.9

[0213] Finally, it should be noted that the above embodiments 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

Claims

1. An enhanced enterprise internal information retrieval and generation system based on a local graph structure, characterized in that: It includes a problem decomposition module, a sub-problem queue module, a knowledge graph hybrid retrieval module, a graph database, a vector database, an evidence chain caching module, a verification module, a definition module, and a formatting module; The problem decomposition module is used to analyze the sentence structure, keyword features, and entity mentions of the original input problem based on rule-based prompt templates and a lightweight classification model. Based on the analysis results, the problem is decomposed. If the problem contains a clear file name, chapter number, or clause number, it is determined to be a clause retrieval problem, the problem type is marked as SIMPLE, the complexity score is low, and no problem decomposition is performed. If the problem involves logical relationships, constraints, or requires cross-paragraph information integration, it is determined to be a comprehension-based multi-hop problem, the problem type is marked as COMPLEX, and a semantic decomposition operation is performed to decompose the problem into a series of sub-problems. The sub-problem queue module is used to store the sequence of sub-problems decomposed by the problem decomposition module. Each sub-problem includes the original problem reference identifier, sub-problem text, expected retrieval entity type, and priority weight. The knowledge graph hybrid retrieval module is used to receive question type identifiers and execute differentiated retrieval strategies for clause-based retrieval questions and comprehension-based multi-hop questions; For clause retrieval questions, the key identifiers in the question, namely the document name, chapter number, and clause number, are used as query conditions. The graph state index is used for precise matching to locate the target clause node, extract the complete clause text, effective date, and revision record from the node attributes, and recall relevant context fragments through vector semantic matching to construct a single-hop evidence set. For the understanding-based multi-hop problem, a dual-path parallel retrieval mechanism is initiated, with path A being the global decomposition path and path B being the incremental iteration path. The global path decomposition process is as follows: traverse each sub-problem in the sub-problem queue, and perform text semantic retrieval and knowledge graph subgraph extraction and expansion respectively; then integrate all subgraphs, identify shared nodes and connecting edges across subgraphs, and construct a multi-hop candidate evidence set; if there is a logical relationship between subgraphs, mark the relationship in the evidence chain buffer module to form a preliminary evidence chain structure. The incremental iteration path process is as follows: Starting with the original input question, bootstrap iterative reasoning and graph expansion are performed. In each iteration, the sub-question generated in the previous round is used as the query to perform semantic retrieval and recall new document fragments. New key entities are identified from the new document fragments, and neighborhood subgraphs are extracted from the knowledge graph using the new key entities as seeds. The evidence newly acquired in this round is accumulated and merged with historical evidence and stored in the evidence chain cache module. The definition module is invoked to determine whether the question has been adequately answered based on the merged evidence. If the determination is sufficient, the iteration is terminated, the search termination condition flag is set to "sufficiently resolved", and the formatting module is entered to generate the answer; If the judgment is insufficient and the number of iterations has not reached the upper limit, the definition module is called to analyze the evidence gap, the next sub-problem is generated and stored in the sub-problem queue module, and the next round of iteration begins; if the number of iterations reaches the upper limit, the retrieval termination condition flag is set to "reached the upper limit" and the iteration is terminated. If no new valid evidence is added in two consecutive iterations during the iteration process, the iteration is terminated; In the first iteration, the original question is used as the query, and semantic retrieval recalls the Top-K related document fragments. When merging evidence, the Top-K related document fragments are merged with the neighborhood subgraph. The graph database is used to store the node and edge relationships of the enterprise system knowledge graph, including entity nodes and relationship edges; Entity nodes include institutional clauses, responsible entities, process nodes, and conditional constraints; relationship edges include referencing relationships, preconditions, postconditions, and logical AND / OR relationships. The vector database is used to store text blocks, which are sub-blocks formed by dividing the preprocessed text of the original rules and regulations documents. The evidence chain caching module is used to store a set of candidate evidence in the multi-hop reasoning process, including textual evidence (original clause content), structural evidence (graph path), evidence source identifier, and evidence timeliness marker. The verification module is used to evaluate the quality of candidate evidence in the evidence chain caching module. The definition module is used to generate a natural language response draft by utilizing a large language model, combining retrieved textual evidence and graph paths, and to determine whether the question has been adequately answered. The formatting module is used to structure the answers based on the scenario's specifications.

2. The enterprise internal information retrieval enhancement generation system based on local graph structure according to claim 1, characterized in that: The problem decomposition module decomposes the comprehension-based multi-hop problem as follows: First, identify the core entity set E = {e1, e2,..., e n} in the problem, including the responsible subject, event type, and action. Secondly, the semantic relations R = {r1, r2,..., r m} between entities are analyzed, and logical structures are identified, including conditional relations, causal relations, parallel relations and reference relations. Then, the original question is decomposed into several sub-questions Q = {q1, q2,..., qn} based on the entity-relation structure, each of which corresponds to a specific inference sub-goal. k} Finally, the sub-problems are stored in the sub-problem queue module according to their logical dependencies, and the priority weights of each sub-problem are set based on the semantic importance of the problem in the original sentence and the degree centrality of the nodes in the knowledge graph.

3. The enterprise internal information retrieval enhancement generation system based on local graph structure according to claim 2, characterized in that: During the global decomposition process, the text semantic retrieval operation is as follows: encode the sub-question into a semantic vector, and recall Top-K relevant text fragments from the document vector library; The operation of the knowledge graph subgraph extraction and expansion is: taking the key entity in the sub-problem as a seed node, performing a breadth-first expansion with a limited depth in the knowledge graph, the limited depth being dynamically adjusted according to the problem complexity, to generate a local subgraph G i =(V i ,E i ), wherein V i is a set of adjacent nodes, and E i is a set of relationship edges; The iterative process of the incremental iteration path is as follows: Round 1 iteration: Starting with the original problem Q o For the query, semantic vector retrieval is used to recall Top-K relevant document fragments D. (1) From D (1) Identify the key entity set E in the middle (1) In knowledge graphs, E (1) Extract a first-order neighborhood subgraph G from the seed node (1) Document evidence D (1) With subgraph G (1) Merge and construct the current evidence set. (1) =(D (1) G (1) The evidence is stored in the evidence chain cache module and marked as having accumulated round number 1; the definition module is called based on Evidence. (1) Determine whether the question has been adequately answered; If the determination is sufficient, the iteration is terminated, the search termination condition flag is set to "sufficiently resolved", and the formatting module is entered to generate the answer; If the judgment is insufficient, the definition module analyzes the evidence gap and generates the next sub-question, Sub-Q. (1) Sub-Q (1) Store it in the subproblem queue module and set the generation round marker to 2; Iteration round t (t≥2): The sub-problem Sub-Q generated in round t-1. (t) To retrieve the newly added document fragment D, perform semantic retrieval. (t) From D (t) Identifying new key entity E (t) , with E (t) Extracting neighborhood subgraph G from the knowledge graph for the seed (t) ; The newly acquired evidence in this round is combined with historical evidence, i.e., Evidence (t) =Evidence (t-1) ∪(D (t) G (t) ), and store it in the evidence chain cache module; Based on the merged evidence Evidence (t) The defined module is called again to determine whether the question has been fully answered; If the determination is sufficient, set the search termination condition flag to "sufficiently solved" and terminate the iteration; if the determination is insufficient and t < T_max, define the module to analyze the evidence gap and continue to generate the next sub-question Sub-Q (t+1) , store it in the sub-question queue module, set the generation round mark to t + 1, and enter the next round of iteration; If the determination is insufficient and t=T_max, set the search termination condition flag to "reached the upper limit" and terminate the iteration; Specifically, evidence consolidation includes: Document evidence deduplication: Remove duplicates based on text content hash value, and reorder by relevance and timeliness; Subgraph merging: unify node identifiers, integrate newly added nodes and edge relationships, identify entity associations across rounds, and expand the graph structure in the local subgraph cache; Evidence chain update: In the evidence chain cache module, mark the source sub-problem, round to which the newly added evidence belongs, and its relationship with historical evidence; The definition module determines whether a question has been adequately answered as follows: Check whether the answer covers all sub-objectives of the question, whether it cites sufficient terms and conditions, and whether it clarifies the conditions, constraints, and exceptions. If the evidence is deemed insufficient, the feedback is sent to the knowledge graph hybrid retrieval module, specifying the entities or relation types that need to be supplemented for retrieval, and a targeted supplementary retrieval is performed. If the evidence is deemed sufficient, the final set of evidence and the draft response will be passed to the formatting module.

4. The enterprise internal information retrieval enhancement generation system based on local graph structure according to claim 3, characterized in that: The specific method for the verification module to perform quality assessment is as follows: First, calculate the semantic relevance score between the evidence and the question; encode the evidence text and the question text into vectors respectively, calculate the cosine similarity, set a similarity threshold θ1, and mark evidence below this threshold as low relevance; Secondly, verify the integrity of the graph structure; check whether the chain of evidence forms a closed reasoning path, that is, whether there is a connected path from the question entity to the answer entity, and whether there are isolated evidence fragments, that is, text without structural association or structure without text support. Then, the verification results are determined; if the semantic relevance scores are all higher than the threshold and the structure is complete, the verification is passed and the system enters the positioning module; if there is low relevance evidence or structural breakage, the re-retrieval mechanism is triggered to expand the depth of the knowledge graph path, increase the number of semantic recalls, or adjust the priority of sub-questions for re-retrieval, and return to the knowledge graph hybrid retrieval module to perform supplementary retrieval, with a maximum of N retries.

5. The enterprise internal information retrieval enhancement generation system based on local graph structure according to claim 1, characterized in that: The structured processing of the formatting module includes: Citation of clauses: Clearly indicate the name of the cited policy document, chapter number, and clause number in your answer; Source Notes: List all content involved in the answer and explain the logical relationship between each part; Additional logical explanation: For comprehension-based multi-hop problems, a brief explanation of the reasoning path is provided; Timeliness reminder: Indicate the effective status of the cited clauses, i.e., currently in effect / revised / repealed.

6. An enhanced generation method for enterprise internal information retrieval based on local graph structure, characterized in that: The enhanced enterprise internal information retrieval generation system based on local graph structure as described in claim 1 is implemented by including the following steps: Step 1: After the user inputs a question, the question type is determined. Based on the rule-based prompt template and lightweight classification model, the sentence structure, keyword features, and entity mentions of the question are analyzed. If the question contains a clear document name, chapter number, or clause number, it is determined to be a clause retrieval question, the question type is set to SIMPLE, the complexity score is low, and the process proceeds to Step 3. If the question involves logical relationships, constraints, or requires cross-paragraph information integration, it is determined to be a comprehension-based multi-hop question, the question type is set to COMPLEX, and the process proceeds to Step 2. Step 2: For complex questions identified as COMPLEX, perform semantic decomposition to break down the question into a series of sub-questions, and then proceed to Step 3; each sub-question contains the original question reference identifier, sub-question text, expected retrieval entity type, and priority weight; Step 3: Knowledge Graph Hybrid Retrieval; Based on the question type identifier identified in Step 1 and the sub-questions decomposed in Step 2, execute differentiated retrieval strategies; For clause retrieval questions, the key identifiers in the question, namely the document name, chapter number, and clause number, are used as query conditions. The graph state index is used for precise matching to locate the target clause node, extract the complete clause text, effective date, and revision record from the node attributes, and recall relevant context fragments through vector semantic matching to construct a single-hop evidence set. For the understanding-based multi-hop problem, a dual-path parallel retrieval mechanism is initiated, with path A being the global decomposition path and path B being the incremental iteration path. The global path decomposition process is as follows: traverse each sub-problem in the sub-problem queue, and perform text semantic retrieval and knowledge graph subgraph extraction and expansion respectively; then integrate all subgraphs, identify shared nodes and connecting edges across subgraphs, and construct a multi-hop candidate evidence set; if there is a logical relationship between subgraphs, mark the relationship in the evidence chain buffer module to form a preliminary evidence chain structure. The incremental iteration path process is as follows: Starting with the original input question, bootstrap iterative reasoning and graph expansion are performed. In each iteration, the sub-question generated in the previous round is used as the query to perform semantic retrieval to recall newly added document fragments. New key entities are identified from the newly added document fragments, and neighborhood subgraphs are extracted from the knowledge graph using the new key entities as seeds. The newly acquired evidence in this round is accumulated and merged with historical evidence and stored in the evidence chain cache module, and then step 4 is executed. In the first iteration, the original question is used as the query, and semantic retrieval recalls the Top-K related document fragments. When merging evidence, the Top-K related document fragments are merged with the neighborhood subgraph. Step 4: Verify the search results, i.e., evaluate the quality of candidate evidence in the evidence chain cache module; if the verification fails, increase the depth of the graph path, increase the number of semantic recalls, or adjust the priority of sub-questions, and return to Step 3 to search again; if the verification passes, proceed to the next step. Step 5: Using a large language model, combine the retrieved textual evidence and graph paths to generate a natural language response draft. Based on the merged evidence, determine whether the question has been adequately answered. If adequately answered, terminate the iteration, set the retrieval termination condition flag to "adequately resolved," and proceed to Step 6. If inadequately answered and the iteration count has not reached the upper limit, analyze the evidence gap, generate the next sub-question, return to Step 3, and proceed to the next round of retrieval iteration. If the iteration count reaches the upper limit, set the retrieval termination condition flag to "upper limit reached," terminate the iteration, and proceed to Step 6. If no new valid evidence is added in two consecutive iterations, terminate the iteration and proceed to Step 6. Step 6: Format the answer output; Based on the scenario's requirements, the answer is output after being structured.

7. The enterprise internal information retrieval enhancement generation method based on local graph structure according to claim 6, characterized in that: The specific method for problem decomposition in step 2 is as follows: First, identify the core entity set E={e1,e2,...,e...} in the problem. n This includes the responsible party, the type of event, and the actions taken; Secondly, analyze the semantic relations between entities R={r1,r2,...,r m } Identify logical structures, including conditional relationships, causal relationships, parallel relationships, and referential relationships; Then, based on the entity-relationship structure, the original problem is decomposed into several subproblems Q={q1,q2,...,q k Each sub-problem corresponds to a specific reasoning sub-goal; Finally, the sub-problems are stored in the sub-problem queue module according to their logical dependencies, and the priority weights of each sub-problem are set based on the semantic importance of the problem in the original sentence and the degree centrality of the nodes in the knowledge graph.

8. The enterprise internal information retrieval enhancement generation method based on local graph structure according to claim 7, characterized in that: In step 3, during the global path decomposition process, the text semantic retrieval operation is as follows: the sub-question is encoded into a semantic vector, and Top-K related text fragments are retrieved from the document vector library; The subgraph extraction and expansion operation of the knowledge graph is as follows: using the key entities in the sub-problems as seed nodes, a breadth-first expansion with a limited depth is performed in the knowledge graph. The limited depth is dynamically adjusted according to the problem complexity, generating a local subgraph G. i =(V i E i ), where V i Let E be the set of adjacent nodes. i For the set of relation edges; The iterative process of the incremental iteration path is as follows: In the first iteration, the original problem Q was used. o For the query, semantic vector retrieval is used to recall Top-K relevant document fragments D. (1) From D (1) Identify the key entity set E in the middle (1) In knowledge graphs, E (1) Extract a first-order neighborhood subgraph G from the seed node (1) Document evidence D (1) With subgraph G (1) Merge and construct the current evidence set. (1) =(D (1) G (1) ), marking the cumulative round as 1; In the t-th (t≥2) iteration, the sub-problem Sub-Q generated in the (t-1)-th iteration is used. (t) To retrieve the newly added document fragment D, perform semantic retrieval. (t) From D (t) Identifying new key entity E (t) , with E (t) Extracting neighborhood subgraph G from the knowledge graph for the seed (t) ; The newly acquired evidence in this round is combined with historical evidence, i.e., Evidence (t) =Evidence (t-1) ∪(D (t) G (t) ); Specifically, evidence consolidation includes: Document evidence deduplication: Remove duplicates based on text content hash value, and reorder by relevance and timeliness; Subgraph merging: unify node identifiers, integrate newly added nodes and edge relationships, identify entity associations across rounds, and expand the graph structure in the local subgraph cache; Evidence chain update: In the evidence chain cache module, mark the source sub-problem, round to which the newly added evidence belongs, and its relationship with historical evidence; The method for determining whether the question has been adequately answered in step 5 is as follows: Check whether the answer covers all sub-objectives of the question, whether it cites sufficient terms and conditions, and whether it clarifies the conditions, constraints, and exceptions. If the evidence is deemed insufficient, return to step 3, specify the entities or relation types that need to be supplemented for retrieval, and perform targeted supplementary retrieval. If the evidence is deemed sufficient, the final evidence set and draft response are obtained, and step 6 is executed.

9. The enterprise internal information retrieval enhancement generation method based on local graph structure according to claim 8, characterized in that: The specific method for quality assessment in step 4 is as follows: First, calculate the semantic relevance score between the evidence and the question; encode the evidence text and the question text into vectors respectively, calculate the cosine similarity, set a similarity threshold θ1, and mark evidence below this threshold as low relevance; Secondly, verify the integrity of the graph structure; check whether the chain of evidence forms a closed reasoning path, that is, whether there is a connected path from the question entity to the answer entity, and whether there are isolated evidence fragments, that is, text without structural association or structure without text support. Then, the verification result is determined; the verification passes if the semantic relevance score is higher than the threshold and the structure is complete; if there is low relevance evidence or structural breakage, the verification fails; the retrieval can be retried up to N times.

10. The enterprise internal information retrieval enhancement generation method based on local graph structure according to claim 6, characterized in that: The structuring process in step 6 includes: Citation of clauses: Clearly indicate the name of the cited policy document, chapter number, and clause number in your answer; Source Notes: List all content involved in the answer and explain the logical relationship between each part; Additional logical explanation: For comprehension-based multi-hop problems, a brief explanation of the reasoning path is provided; Timeliness reminder: Indicate the effective status of the cited clauses, i.e., currently in effect / revised / repealed.

Images