A problem anchor graph based retrieval augmentation generation method

Abstract

The application discloses a retrieval enhancement generation method based on a problem anchor graph, and belongs to the technical field of natural language processing and information retrieval. The method comprises the following steps: in the index stage, a document corpus is divided into text blocks, a large language model is used to generate answerable questions for each text block and record the generation relationship, an association relationship is established based on the semantic similarity between the questions, and a heterogeneous problem-block graph containing text block nodes, question nodes, generation edges and association edges is constructed; in the retrieval stage, a user query is matched with the question nodes to obtain a seed question set, the text blocks are directly retrieved along the generation edges and expanded along the association edges, and a traceable retrieval tree from the user query to the retrieved text blocks is constructed; and in the generation stage, hierarchical summarization and answer generation are performed based on the text blocks and question nodes in the retrieval tree. By introducing the question nodes as semantic intermediaries, the application changes the traditional direct matching of "query to text block" into an indirect matching path of "query to question and then to text block", effectively alleviates the semantic alignment deviation, and provides an explicit multi-hop reasoning chain through the retrieval tree, thereby significantly enhancing the explainability and evidence recall capability of the retrieval process.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing and information retrieval technology, and particularly relates to a question-anchored graph based retrieval-augmented generation (QAG-RAG) method and system. More specifically, this invention relates to constructing a question-block heterogeneous graph using a large language model to bridge user queries and document content, mitigating semantic alignment bias through a retrieval path of "query to question to text block," and constructing a traceable retrieval tree based on graph traversal to enhance the interpretability of the retrieval process. Ultimately, it can be applied to scenarios such as multi-hop question answering, knowledge-intensive reasoning, and domain-specific information retrieval. Background Technology

[0002] In recent years, large language models have demonstrated remarkable capabilities in natural language processing and text generation tasks. However, the "illusion" phenomenon in their generated content severely limits their reliability in real-world applications. To address this issue, retrieval-enhanced generation techniques have emerged. Their core idea is to combine external knowledge base retrieval mechanisms with the generative model, enabling the model's output to be based on external knowledge sources. This improves factual accuracy and supports knowledge-intensive tasks such as question answering, information synthesis, and decision support.

[0003] However, existing retrieval enhancement generation systems face two fundamental challenges in practical deployment. First, semantic alignment bias: user queries are typically highly abstract and generalized, while document fragments are specific and context-dependent fine-grained content, resulting in a significant difference in information granularity. Traditional dense retrieval methods rely on vector embedding for similarity matching, which can capture surface semantic relationships but struggles to accurately map the user's deep query intent to specific document content. Although existing research has proposed hypothetical document generation, query rewriting, and expansion schemes to mitigate this bias, these methods essentially still rely on direct similarity matching between queries and text blocks, lacking explicit semantic alignment mechanisms and failing to effectively bridge the gap between abstract queries and concrete content.

[0004] Second, the retrieval process lacks interpretability: the retrieval phase of traditional retrieval enhancement systems is often a "black box" operation. The uninterpretability of vector embeddings means the system cannot provide a clear reasoning path to explain why specific evidence was chosen over other content. This is particularly critical in fields requiring auditing and traceable decision-making (such as healthcare, law, and finance). Some existing technologies attempt to introduce entity knowledge graphs to enhance structured reasoning, but such methods typically require extensive manual annotation, making automated construction and cross-domain transfer difficult. Other studies propose dynamically constructing reasoning paths during the reasoning phase, but these methods fail to establish a reusable and searchable question space during the indexing phase, resulting in a lack of stable structured support for reasoning path construction and difficulty in ensuring global consistency and traceability of the retrieval process.

[0005] Furthermore, in multi-hop question-answering scenarios, user queries often involve complex reasoning across multiple document fragments. Existing retrieval systems, limited by the direct matching paradigm, struggle to effectively uncover indirect evidence that is semantically distant from the query but topically relevant. While recursive retrieval and query decomposition techniques can expand the retrieval scope to some extent, the generated subqueries are typically limited to the semantic neighborhood of the original query, lacking flexible associative expansion capabilities and failing to provide an explicit reasoning chain from the query to supporting evidence.

[0006] In summary, existing retrieval enhancement generation technologies still have significant shortcomings in terms of semantic alignment accuracy, retrieval process interpretability, and multi-hop reasoning capabilities. There is an urgent need for a technical solution that can pre-build a structured question space in the indexing stage, achieve traceable multi-hop association expansion in the retrieval stage, and synthesize answers based on explicit reasoning paths in the generation stage. Summary of the Invention

[0007] In view of the technical problems existing in retrieval enhancement generation technology, such as semantic alignment bias, lack of interpretability of the retrieval process, and insufficient multi-hop reasoning ability, this invention provides a retrieval enhancement generation method based on question anchoring graph.

[0008] The technical solution adopted in this invention is as follows: a retrieval enhancement generation method based on a question anchoring graph, comprising an indexing stage, a retrieval stage, and a generation stage. In the indexing stage, the document corpus is divided into several text blocks. For each text block, at least one answerable question is generated using a large language model, and the generation relationship between each question and its source text block is recorded. Based on the semantic similarity of the questions, association relationships are established between the questions, thereby constructing a heterogeneous question-block graph containing text block nodes, question nodes, generation edges, and association edges.

[0009] In the retrieval phase, a user query is received, and the user query is matched with the question nodes in the question-block graph to obtain a seed question set. Based on the seed question set, directly related text blocks are retrieved along the generating edges, and multi-hop expansion is performed along the related edges to traverse to the related questions and their corresponding source text blocks, thereby constructing a retrieval tree from the user query to each retrieved text block.

[0010] In the generation stage, based on the text blocks and corresponding question nodes in the retrieval tree, hierarchical summaries and answers are generated to obtain the final answer.

[0011] The generation of answerable questions involves: for each text block, using a large language model to generate several answerable questions based on structured prompts, and generating an interpretability description for each question. This interpretability description records the textual scope or knowledge motivation upon which the question is based. This design provides a traceable auditable basis for the question generation process, alleviating the "black box" problem of traditional retrieval enhancement generation systems during the indexing stage.

[0012] The method of establishing associations based on semantic similarity includes: calculating the semantic similarity between the embedding vectors corresponding to any two question nodes; if the semantic similarity exceeds a preset threshold, then establishing an undirected association edge between the two question nodes. In this way, semantically similar questions form associations across text blocks, and thus, without directly establishing text block-to-text block edges, cross-block semantic associations are achieved through an implicit bridging path of "text block—question—associated question—text block," significantly reducing the computational complexity of graph construction.

[0013] The process of matching user queries with question nodes includes: decomposing the user query into several sub-queries using a large language model; calculating the embedding similarity between the user query and each sub-query and each question node in the question-block graph; and taking the union of the top K question nodes with the highest similarity as the seed question set. Through query decomposition, the high-information-density original query is aligned to specific, focused question nodes, improving the accuracy of the initial matching.

[0014] The multi-hop expansion along the associated edges includes: in the h-th layer expansion, starting from the question node in the (h-1)-th layer, associated question nodes are discovered along the associated edges, and then the source text blocks corresponding to the associated question nodes are obtained along the generating edges, serving as the retrieval results for the h-th layer; this process is iteratively executed until a preset maximum retrieval depth is reached, resulting in a multi-hop retrieval text block set. This multi-hop traversal mechanism enables the system to overcome the semantic neighborhood limitations of the original query and discover evidence that is semantically distant but indirectly related to the topic.

[0015] After constructing the retrieval tree, a feature filtering step is included: extracting entity and relation features from the user query, subquery, seed question set, and text blocks directly retrieved in the first layer, respectively, and constructing a feature matrix; calculating the cosine similarity between the embedding vector of the text block to be filtered and each feature in the feature matrix, and obtaining a relevance score based on weighted summation; pruning text blocks with scores below the filtering threshold from the retrieval tree. This filtering mechanism effectively suppresses noise introduced during multi-hop expansion, ensuring high evidence quality entering the generation stage.

[0016] The hierarchical summary and answer generation includes: aggregating text blocks and corresponding question nodes layer by layer upwards according to the retrieval depth in the retrieval tree; in each layer, the text blocks, corresponding question nodes, and aggregated summaries from deeper layers are input into a large language model to generate an aggregated summary for that layer; until the top layer generates the final summary, and based on the final summary, a final answer to the user query is generated. This bottom-up hierarchical aggregation strategy ensures that the answer synthesis process strictly corresponds to the explicit reasoning path of the retrieval tree, making each part of the final answer traceable to a specific question-evidence. This invention also provides a retrieval enhancement generation system based on a question anchoring graph, including an index module, a retrieval module, and a generation module. The index module is configured to construct the aforementioned heterogeneous question-block graph; the retrieval module is configured to generate a retrieval enhancement generation system based on the user query.

[0017] Seed question matching, multi-hop traversal, and feature filtering are performed in the question-block graph to construct a retrieval tree; the generation module is configured to perform hierarchical summary and answer generation based on the retrieval tree.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] First, by introducing question nodes as explicit semantic intermediaries between user queries and document blocks, the traditional direct matching paradigm of "querying to a text block" is transformed into an indirect matching path of "querying to a question and then to a text block," effectively mitigating the semantic alignment deviation between abstract queries and concrete content. Second, the edges generated in the question-block heterogeneous graph provide explicit source relationships from questions to source blocks, while the associated edges provide semantic bridging capabilities across blocks. This allows the retrieval process to present a complete root-to-leaf reasoning chain through the retrieval tree, significantly enhancing the interpretability and auditability of the retrieval process. Third, the multi-hop traversal mechanism based on associated edges can automatically discover indirectly relevant evidence in the pre-built question space during the indexing phase, breaking through the semantic neighborhood limitations of the original query and improving the evidence recall rate in multi-hop question answering and composite reasoning tasks. Fourth, the question generation process during the indexing phase is accompanied by interpretability explanations, making the motivation and basis for graph construction auditable and verifiable, enhancing the system's transparency from the data source. Fifth, the hierarchical answer synthesis mechanism strictly corresponds to the deep structure of the retrieval tree, ensuring that the final answer generation process is built on a traceable question-evidence aggregation path, achieving full-link interpretability from retrieval to generation. Attached Figure Description

[0020] Operation flowchart: Figure 1 Question-Chunk Graph Construction Figure 2 QAG-RAG search process Figure 3 QAG-RAG Reasoning Process Figure 4 "What are Yamanaka factors?" is the search intent in the Question-Chunk Graph. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0022] Overall Architecture Overview

[0023] like Figure 1 As shown, the retrieval enhancement generation system based on question anchoring graphs described in this embodiment mainly includes two core data structures: a heterogeneous question-chunk graph and a retrieval tree. The question-chunk graph is pre-built during the indexing phase to bridge user queries and document content; the retrieval tree is dynamically generated during the retrieval phase to record the explicit reasoning path from the user query to supporting evidence.

[0024] Indexing Phase: Problem-Block Graph Construction

[0025] like Figure 1 As shown, the construction process of the question-block graph includes two stages: question generation and graph assembly.

[0026] In the problem generation phase, given a document corpus First, it is divided into several non-overlapping sets of text blocks. Each text block Represents a semantically coherent unit. For each text block... Using a large language model based on structured prompts Generate a set of answerable questions: Each of the following questions All are anchored to their source text blocks Contextual information. Unlike traditional keyword extraction or summary generation, this embodiment requires the large language model to simultaneously generate interpretability descriptions for each question. These interpretability descriptions record the textual scope or knowledge motivation upon which the question is based, thereby forming auditable metadata during the indexing stage.

[0027] In the implementation of this embodiment, the structured hints The large language model is required to decompose the text into a series of clear, answerable, and logically ordered sub-questions, each limited to 50 words in length, and to explain the reason for asking the question.

[0028] In the graph assembly stage, such as Figure 1 As shown in Step 2, all text block nodes and question nodes are heterogeneously integrated. Node set Contains text block nodes With the problem node Edge set It contains two types of edges:

[0029] Generate edges ( For each text block The generated problem Establish an undirected edge This is used to ensure traceability from the issue to its source text block.

[0030] Associated edges ( For any two problems Calculate the semantic similarity of their embedding vectors; if the similarity exceeds a preset threshold tau_q (for example, in this embodiment, tau_q is taken as tau_q), then... If ), then an undirected associative edge is established. In this embodiment, all embeddings are computed using the bge-m3 model, generating 1024-dimensional vectors.

[0031] It is worth noting that this embodiment does not directly establish edges between text blocks. When an associated edge connects text blocks from different sources (i.e., ...), ... and Automatically generates implicit bridging paths across text blocks: This design significantly reduces the computational complexity of graph construction and avoids exhaustive pairwise comparisons of text blocks.

[0032] Search Phase: Graph-Guided Search

[0033] like Figure 2 As shown, after a user query is given, the retrieval phase executes four steps in sequence: subquery generation, question matching, multi-hop traversal, and feature filtering.

[0034] Subquery generation: User queries typically have high information density, making precise matching with individual question nodes difficult. Therefore, this embodiment utilizes a large language model to decompose the user query into a set of focused subqueries. Each subquery captures a different aspect of the original query.

[0035] Question matching: Calculate the user query and each subquery separately. The embedding similarity with all problem nodes in the problem-block graph is used to retrieve the top K most similar problem nodes (K=5 in this embodiment) for each query, and the union of these nodes is taken as the seed problem set. .

[0036] Multi-hop traversal: Constructing a retrieval tree based on the seed problem set. Layer 1 is direct retrieval: for each seed problem... Generate edges along it Directly retrieve the corresponding source text block. (H-th layer) For association expansion: Starting from the problem node at level h-1, follow the association edge Discover the associated problem nodes, and then follow the generated edges. Retrieve the source text block as the search result for the h-th level. Iterate through this process until the preset maximum search depth H is reached.

[0037] Feature filtering: Multi-hop traversal may introduce noisy text blocks with low relevance to the original user query. Therefore, this embodiment constructs a feature-based filtering mechanism. Entities and relations are extracted as discriminative features from the following four sources: (1) the original user query; (2) the expanded subquery. (3) The set of seed problems matched during the problem matching phase (4) The first layer retrieves text blocks directly through the generated edges. After embedding the above features, a feature matrix is ​​constructed. For each text block c outside the first level of the retrieval tree, compute its embedding vector. The cosine similarity to each column of the feature matrix is ​​used, and the relevance score is obtained by weighted summation: The weight vector w distributes the weights evenly across the feature elements, i.e. Scores below the filtering threshold will be filtered out. (In this embodiment, we take) Text blocks are pruned from the retrieval tree.

[0038] In the search tree In this context, each root-to-leaf path represents a complete multi-hop inference chain from the user query through intermediate question nodes to the supporting text block.

[0039] Generation Phase: Layered Answer Synthesis

[0040] like Figure 3 As shown, after obtaining the filtered search tree, the text blocks are organized into a hierarchical structure according to the search depth, where This represents the set of text blocks retrieved in the h-th hop. This indicates the problem node corresponding to this layer.

[0041] Hierarchical summarization employs a bottom-up aggregation strategy. Starting from the maximum retrieval depth H, it utilizes a large language model based on prompts... Summarize the deepest problem nodes and text blocks: For the intermediate layer This will combine the current layer's question nodes, text blocks, and aggregated summaries from deeper layers. All inputs are aggregated into a large language model: Finally, top-level summary It encapsulates a complete multi-hop reasoning chain as the context for answer generation: This hierarchical mechanism ensures that each step of answer synthesis is based on traceable question-evidence pairs in the retrieval tree.

[0042] Implementation Case: Multi-hop Question Answering in the Biomedical Field

[0043] like Figure 4 As shown, the complete retrieval process of this implementation case is illustrated using the query "What is Yamanaka factors?" from the mini-BioASQ dataset as an example.

[0044] In this local retrieval tree, the nodes marked with red circles represent seed questions obtained by matching user queries with embedded similarity, such as "What are Yamanaka factors?" or "What is the role of Yamanaka factors in cell reprogramming?". These seed questions serve as entry points for the retrieval and are directly connected to their source text blocks (blue nodes) along the generating edges.

[0045] Nodes marked with green circles represent target text blocks discovered through traversal of associative edges. For example, a seed question might be related to another question, "What are the key transcription factors for induced pluripotent stem cells?". Although this question is not highly similar to the literal wording of the user's query, a semantic bridge is established through the associative edges between question nodes. By tracing the generating edges of this related question, indirectly related but highly relevant text blocks can be retrieved. These text blocks may contain specific protein names of Yamanakafactors (such as Oct3 / 4, Sox2, Klf4, c-Myc) and their molecular mechanisms in somatic cell reprogramming.

[0046] This case demonstrates that this embodiment successfully retrieved indirect evidence that was not directly similar to the user's query but was related to the topic by using the question node as a semantic mediator. Furthermore, the entire retrieval process provided an explicit and auditable reasoning chain from the query to the supporting evidence through the retrieval tree.

[0047] Experimental verification parameters

[0048] In the experimental verification of this embodiment, the question-block graph is stored in the Neo4j graph database, and the node embedding vectors are indexed through the Milvus vector database to support efficient similarity retrieval. The retrieval depth H is configured according to the characteristics of the dataset and is set to 8 hops in the ablation experiment of mini-BioASQ. The large language model generators used include, but are not limited to, Qwen3-1.7B, GLM-Z1-9B, and DeepSeek-V3.2, all of which perform question generation, subquery expansion, hierarchical summarization, and answer generation in a zero-shot manner without fine-tuning.

Claims

1. A retrieval enhancement generation method based on question anchor graphs, characterized in that... The method includes: S1. Indexing Phase Construction Issues - Block Graph: Dividing the document corpus into a set of non-overlapping text blocks. For each text block Generate a set of answerable questions using a large language model And corresponding interpretability descriptions, and assemble heterogeneous question-block graphs based on the text blocks and the answerable questions. , where the set of nodes By text block node and problem nodes Composition, edge set Includes generated edges and associated edges The generated edges connect each question node to its source text block node, and the associated edges connect nodes whose semantic similarity exceeds a preset threshold. Two problem nodes; S2. Search Phase Execution Graph Guides Search: Receiving User Queries Decompose it into a set of subqueries using a large language model. Based on the embedding vectors of the user query and the subquery set, a seed question set is retrieved from the question-block graph. Starting from the seed problem set, a multi-hop traversal is performed along the associated edges and generating edges to construct a retrieval tree layer by layer and obtain the set of retrieval text blocks. ; S3. Generation Phase: Perform hierarchical answer synthesis: Organize the text blocks in the retrieval tree hierarchically according to the retrieval depth. Using a bottom-up approach, starting from the deepest layer, use a large language model to summarize the question nodes and text blocks at each layer. Use the aggregated summaries from deeper layers as additional context for the shallower summaries to obtain the top-level summary. The final answer to the user query is generated based on the top-level summary.

2. The method according to claim 1, characterized in that... The process of generating answerable questions in step S1 includes: using structured prompts The large language model is invoked to decompose the text block into a set of clear, answerable, and logically ordered sub-questions, with each sub-question limited to a preset word count. For each generated sub-problem, an interpretability description is also output. The interpretability description records the textual scope or knowledge motivation on which the sub-problem is based, serving as auditable metadata for the indexing phase.

3. The method according to claim 1, characterized in that... The process of assembling the heterogeneous problem-block diagram in step S1 includes: For each text block The generated problem Establish undirected generated edges Ensure traceability from the issue node to the source text block node; for the global issue set Any two questions in and ,calculate When the similarity is greater than or equal to the threshold At that time, establish undirected associative edges. ; When a certain associated edge The two connected questions originate from different text blocks. and At that time, without explicitly creating text block-to-text block edges, via path Form an implicit bridge across text blocks.

4. The method according to claim 1, characterized in that... In step S2, the seed problem set is obtained. The process includes: Query the user respectively and each of its subqueries Calculate the embedding vector; based on the embedding vector, in the set of problem nodes... The top K question nodes with the highest cosine similarity are retrieved from each query; the union of the top K question nodes corresponding to the user query and all subqueries is taken as the seed question set. .

5. The method according to claim 1, characterized in that... The multi-hop traversal process described in step S2 includes: First-level direct retrieval: For each seed problem in the seed problem set. , along its generated edges Retrieve the source text block to obtain the first-level text block set. ; The h-th layer association extension (h≥2): Let First, expand along the associated edges from the problem nodes at level h-1 to obtain the set of problem nodes at level h. Then, by searching along the generation edges, the set of text blocks at the h-th level is obtained. ; Iterate through the above association expansion process until the preset maximum retrieval depth H is reached, then merge all layers to obtain the set of retrieved text blocks. .

6. The method according to claim 5, characterized in that... The method further includes a step of feature filtering on the retrieval tree: Entities and relationships are extracted as discriminative features from four sources: the user query, the subquery set, the seed question set, and the text blocks retrieved through edge generation in the first layer, to obtain a feature set. ; The embedding is calculated for each feature element in the feature set to obtain the feature matrix. ; For each text block outside the first level of the retrieval tree Calculate its embedding vector The cosine similarity with each column of the feature matrix is ​​used, and the relevance score is obtained by weighted summation according to the weight vector w. ; Those with relevance scores below the filtering threshold The text blocks are pruned from the search tree to obtain a refined search tree.

7. The method according to claim 1, characterized in that... The process of synthesizing the layered answer in step S3 includes: For the deepest layer h=H, based on summary hints The large language model for the problem nodes of this layer and text block collection Summary: ; For the middle layer where 1 ≤ h < H, input the problem nodes of this layer , the set of text blocks , and the aggregated summary from deeper layers into the large language model for aggregation: ; Based on top-level summary Answer generation hints The large language model generates the final answer to the user's query. .

8. The method according to claim 1, characterized in that... The search tree In the vertex set It includes at least the user query node, the seed question set, the question nodes obtained from each layer of expansion, and the retrieved text blocks, and the edge set. The generated edges and associated edges traversed during the multi-hop traversal are recorded. Each path from the root to the leaf in the retrieval tree represents a traceable multi-hop reasoning chain from the user query through intermediate question nodes to the supporting text block, which is used to provide interpretability of the retrieval process.

9. The method according to claim 1, characterized in that... The embedding vectors of the question nodes and text block nodes are generated by a unified embedding model. The nodes in the question-block graph are stored in a graph database, and the embedding vectors of each node are simultaneously indexed in a vector database to support efficient similarity retrieval based on approximate nearest neighbor search.

10. A retrieval enhancement generation system based on question anchor graphs, characterized in that... The system includes: The question-block graph construction module is used to segment the document corpus into blocks during the indexing stage, and generate answerable questions and interpretable descriptions for each text block using a large language model. It then assembles a heterogeneous question-block graph based on generated edges and associated edges. The graph-guided retrieval module receives user queries during the retrieval phase, generates subqueries using a large language model, matches seed question sets based on embedded similarity, and constructs a retrieval tree through multi-hop traversal along associated and generated edges. The feature filtering module extracts entities and relationships as discriminative features from the user query, subqueries, seed questions, and first-level text blocks, scores text blocks in the retrieval tree based on weighted cosine similarity, and prunes text blocks with scores below a filtering threshold. The hierarchical answer synthesis module is used to organize text blocks according to retrieval depth, aggregate evidence using a bottom-up hierarchical summarization approach, and generate the final answer based on the top-level summary; wherein, the question-block graph construction module, graph-guided retrieval module, feature filtering module, and hierarchical answer synthesis module work together to implement the method described in any one of claims 1 to 9.

Images