Retrieval enhancement generation method of knowledge index based on atlas anchor points

Abstract

The invention belongs to the technical field of retrieval enhancement generation, and discloses a retrieval enhancement generation method of knowledge indexes based on atlas anchor points. According to the method, a graph structure in static knowledge representation is redefined as a dynamically evolved knowledge index, the graph structure is incrementally updated through an iterative retrieval process, key entities and relationships are anchored, and a structured index is generated to guide a large language model to evaluate knowledge sufficiency and construct subsequent sub-queries. The final answer is jointly generated by all retrieval documents and the final evolution diagram. Experiments in a plurality of multi-hop question and answer benchmark tests verify the effectiveness of the method, and reveal that the method can more efficiently associate key information distributed in a retrieval document by adjusting the attention mechanism of a large language model.

Description

Technical Field

[0001] This invention relates to the field of retrieval enhancement generation technology, and in particular to a retrieval enhancement generation method based on knowledge indexing of graph anchors. Background Technology

[0002] Retrieval-enhanced generative models typically help large language models generate accurate and well-founded answers by retrieving relevant paragraphs and using them as contextual input. However, traditional retrieval-enhanced generative methods often face the problem of insufficient knowledge acquisition, especially in multi-hop question-answering scenarios. In such scenarios, the reasoning process needs to integrate evidence distributed across multiple different paragraphs, which traditional methods struggle to handle effectively.

[0003] To mitigate these limitations, existing research has proposed iterative retrieval strategies aimed at progressively finding and accumulating relevant knowledge to support answer generation. Furthermore, some methods perform deep retrieval by interacting with external retrieval tools during the reasoning process, achieving more flexible and adaptive information searching. However, as the number of retrieval steps increases, these retrieval-enhanced generation systems inevitably introduce more noise or irrelevant information. Studies have shown that excessive irrelevant information significantly degrades the quality of the generated output.

[0004] To improve the utilization of knowledge by large language models, existing work has explored methods to reorder retrieved paragraphs or summarize salient information to filter noise. However, these methods typically process retrieved paragraphs in isolation, lacking modeling of relationships between different documents. As a result, cross-document evidence may be overlooked or undercaptured, a deficiency particularly pronounced when it is necessary to aggregate complementary information from multiple sources.

[0005] To better aggregate diverse knowledge from multiple retrieved paragraphs, an increasing number of studies utilize structured representations (including patterns, tables, and graphs) to explicitly extract key information scattered across these documents. Some works construct graphs to identify relevant information across multiple paragraphs or blocks, connecting extracted key entities and facts through graph structures. This allows Big Language to better capture long-range dependencies and multi-hop relationships in long texts or across documents, thereby facilitating the localization of relevant evidence. Other studies explore multiple reasoning paths on constructed knowledge graphs, identifying conflicting paths as input context to mitigate knowledge conflicts between internal and external information.

[0006] However, most existing graph-based methods focus on using graph structures for static dependency capture or conflict identification, or tend to completely replace retrieved paragraphs with structured representations, lacking a mechanism that can dynamically evolve with the iterative retrieval process and actively anchor key concepts to guide knowledge interpretation and subquery generation. Therefore, how to design a dynamic indexing mechanism that can effectively anchor knowledge, reduce noise interference, and support deep reasoning remains a problem that urgently needs to be solved in current technology. Summary of the Invention

[0007] The purpose of this invention is to propose a knowledge indexing retrieval enhancement generation method based on knowledge graph anchors. By introducing an evolutionary knowledge graph as a knowledge anchoring mechanism within an iterative retrieval enhancement generation framework, it addresses the difficulties in knowledge integration and insufficient reasoning depth in existing technologies. This invention transforms the knowledge graph from a mere knowledge representation into an active indexing tool, progressively updating the graph during iterative retrieval to enhance the interpretation of the retrieved knowledge. By updating the graph at each retrieval step and generating the final answer when sufficient knowledge is available, this invention can more effectively organize and interpret the retrieved knowledge, thereby improving the reasoning accuracy and answer generation reliability of large language processing.

[0008] The technical solution of this invention: A retrieval enhancement generation method based on knowledge indexing of graph anchor points, specifically including the following steps:

[0009] Step 1: For a given initial query, initiate the iterative retrieval process;

[0010] During the iterative retrieval process, reasoning is performed through a large language model. Based on the output of the previous reasoning of the large language model, the graph is updated, the reasoning trajectory is generated, and the next query is generated simultaneously. The graph is an evolutionary graph, which serves as a context-aware knowledge index and gradually accumulates key entities and relationships during the iterative retrieval.

[0011] Step 3: Once the iterative retrieval process meets the termination condition, aggregate all retrieved documents and combine them with the final evolutionary graph to generate the final answer for the initial query.

[0012] In the In the next iteration retrieval step, the query generated in the previous iteration is used as the basis for the retrieval process. Retrieve a set of relevant documents from an existing knowledge base. :

[0013]

[0014] in , Indicates the maximum number of iterations for retrieval; This indicates that a dense search engine is used to retrieve a set of documents related to the query from the knowledge base.

[0015] The iterative retrieval specifically refers to:

[0016] Based on the large language model Combined with the initial query Current document being searched and the output of the previous iteration The evolutionary map was updated to obtain :

[0017]

[0018] in, Includes the previous evolutionary map Intermediate inference results and query ;

[0019] Based on the updated evolutionary map Current document being searched And the previous output Generate intermediate inference results for the current step. And the next query :

[0020]

[0021] Among them, the reasoning results It includes reasoning patterns and knowledge sufficiency judgments; if the judgment is insufficient, then based on... generate And then proceed to the next iteration.

[0022] The specific process of updating the evolution map is as follows:

[0023] Employing a stepwise graph evolution mechanism, the evolutionary graph is transformed from... Updated to ; from Transition to At that time, through the steps Documents included in the search The evolutionary map structure is gradually updated, thus forming a gradual evolutionary process:

[0024]

[0025] Among them, the Indexing operations are based on the previous inference trajectory. and query The combination, from relevant documents Extract significant information relevant to the query and integrate it into the evolutionary map.

[0026] The evolutionary map The specific representation of is as follows:

[0027] Define evolutionary map ,in For a collection of entities, For a set of relation triples; the evolutionary graph is transformed into a text sequence through linearization:

[0028] Entities ;

[0029] Relations:

[0030] in, It is a language-based function that contains entities and their attributes. It is an RDF triple language function that includes a head entity, a relation, and a tail entity.

[0031] After completion After the next iteration of retrieval, if the knowledge output by the large language model in step 2.2 is deemed sufficient or the current iteration count reaches the maximum number of iterations, then all documents retrieved in all steps are combined into a union set to form a complete document set. :

[0032] .

[0033] Based on the complete document set Generate instructions Initial query Aggregated document sets and the final evolutionary map According to the large language model Generate the final answer :

[0034]

[0035] Among them, the evolutionary map As anchor points for structured knowledge, they are used to assist in the organization and interpretation of large models. The information in the middle.

[0036] The beneficial effects of this invention are as follows: This invention proposes a retrieval enhancement generation method based on graph anchors for knowledge indexing. By using an evolutionary graph as a structured knowledge anchor, it preserves significant factual and relational structures. Compared with existing methods, this invention does not merely extract knowledge for question answering, but focuses on knowledge anchoring. It introduces evolutionary graphs to anchor key knowledge in retrieved documents within an iterative retrieval enhancement generation framework, better assisting the model in generating subqueries to obtain higher-quality retrieved documents and improve the accuracy of the final question answer. Specifically, this invention facilitates the retrieval of more relevant and diverse documents in multiple retrieval steps and enables large language models to better interpret and utilize the retrieved knowledge. As the retrieval process progresses, the graph anchors more and more relevant entities and relationships into the index, thereby modulating the attention mechanism of the large language model to aggregate evidence from scattered retrieved documents, thus significantly improving the accuracy of reasoning and the reliability of generated answers. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of a retrieval enhancement generation method based on knowledge indexing using graph anchors. Detailed Implementation

[0038] like Figure 1 As shown, a retrieval enhancement generation method based on knowledge indexing of graph anchor points is presented.

[0039] The specific steps are as follows:

[0040] Step 1: Initialization Setup. Given an initial query that requires multi-hop retrieval. The system introduces an evolutionary knowledge graph. As a structured indexing mechanism, this invention utilizes a graph to simultaneously support inference, next query generation, and final answer prediction, unlike existing methods that only perform iterative retrieval to accumulate evidence.

[0041] Step 1.1: Document Retrieval. Based on the previous query... Retrieve a set of related documents :

[0042]

[0043] in This represents the maximum number of iterations.

[0044] Step 2: Perform iterative retrieval and graph evolution. In the... In this iteration, the following sub-steps are executed:

[0045] Step 2.1: Stepwise spectral evolution. Utilizing a large language model. Based on the initial query New search documents and including , , Previous output Updated map:

[0046]

[0047] Specifically, from arrive The transformation is an incremental update process:

[0048]

[0049] in This indicates a document indexing operation, which is based on the previous inference path. Previous query As a condition. These two signals together guide the flow from... Significant, query-related information is extracted for graph updates, thereby gradually anchoring key entities and relationships into the evolutionary graph structure.

[0050] The map The representation uses a linearization method. The spectrum is defined as... ,in It is a collection of entities. It is a set of relations extracted from documents and represented as RDF triples. Graph Linearization is achieved by linguisticizing its entities and triples:

[0051] Entities ;

[0052] Relations:

[0053] in Entity language functions that represent the encoded entity names and their attributes. Represents a language-specific RDF triple. The result is wrapped in a specific tag. <graph> and< / graph> middle.

[0054] Step 2.2: Joint Reasoning and Query Generation. Using the updated graph... Document retrieval and the previous output As a condition, utilize large language models Generate intermediate inference results and next query :

[0055]

[0056] The reasoning results Included by <think> and< / think> The reasoning behind the package, and what it was... <judgement> and< / judgement> The knowledge sufficiency of the package is assessed. If the assessment is "insufficient," then a knowledge sufficiency test is performed. Continue iterating; if it is "sufficient", then terminate the iteration.

[0057] Step 3: Final answer generation.

[0058] Step 3.1: Aggregate all The documents retrieved in each iteration form a complete set. :

[0059]

[0060] Step 3.2: Utilize a large language model based on aggregated documents and the final evolutionary map Generate the final answer :

[0061]

[0062] As the number of iterations increases, the document set Inevitably, noise accumulates, and the evolutionary map... It is refined step by step in each iteration, serving as a structured knowledge anchor point to help the model organize and interpret the retrieved knowledge more effectively, thereby achieving high-precision reasoning and responses.

[0063] This invention is applied to question-and-answer scenarios, particularly complex user question-and-answer scenarios. In this field, during the process of answering user questions, this invention utilizes an evolutionary knowledge graph to progressively construct a document index, helping large language models anchor key information and improving their knowledge understanding and answer generation capabilities.

[0064] Table 1 shows the optimal performance of this invention in multiple multi-hop and complex problem scenarios. As can be seen from Table 1, the accuracy of this invention surpasses traditional retrieval enhancement generation methods, including direct retrieval enhancement generation, adaptive retrieval enhancement generation, and multi-round knowledge refinement methods. Furthermore, this invention can further improve the interpretability of knowledge utilization by using knowledge graphs for document knowledge indexing and anchoring. This fully demonstrates the effectiveness and efficiency advantages of this invention in retrieval enhancement generation.

[0065] Table 1 Comparison of Model Performance

[0066]

Claims

1. A retrieval enhancement generation method based on knowledge indexing using graph anchor points, characterized in that, The specific steps are as follows: Step 1: For a given initial query, initiate the iterative retrieval process; During the iterative retrieval process, reasoning is performed through a large language model. Based on the output of the previous reasoning of the large language model, the graph is updated, the reasoning trajectory is generated, and the next query is generated simultaneously. The graph is an evolutionary graph, which serves as a context-aware knowledge index and gradually accumulates key entities and relationships during the iterative retrieval. Step 3: Once the iterative retrieval process meets the termination condition, aggregate all retrieved documents and combine them with the final evolutionary graph to generate the final answer for the initial query.

2. The retrieval enhancement generation method based on knowledge indexing of graph anchor points according to claim 1, characterized in that, In the In the next iteration retrieval step, the query generated in the previous iteration is used as the basis for the retrieval process. Retrieve a set of relevant documents from an existing knowledge base. : in , Indicates the maximum number of iterations for retrieval; This indicates that a dense search engine is used to retrieve a set of documents related to the query from the knowledge base.

3. The retrieval enhancement generation method based on knowledge indexing of graph anchor points according to claim 1, characterized in that, The iterative retrieval specifically refers to: Based on the large language model Combined with the initial query Current document being searched and the output of the previous iteration The evolutionary map was updated to obtain : in, Includes the previous evolutionary map Intermediate inference results and query ; Based on the updated evolutionary map Current document being searched And the previous output Generate intermediate inference results for the current step. And the next query : Among them, the reasoning results It includes reasoning patterns and knowledge sufficiency judgments; if the judgment is insufficient, then based on... generate And then proceed to the next iteration.

4. The retrieval enhancement generation method based on knowledge indexing of graph anchor points according to claim 3, characterized in that, The specific process of updating the evolution map is as follows: Employing a stepwise graph evolution mechanism, the evolutionary graph is transformed from... Updated to ; from Transition to At that time, through the steps Documents included in the search The evolutionary map structure is gradually updated, thus forming a gradual evolutionary process: Among them, the Indexing operations are based on the previous inference trajectory. and query The combination, from relevant documents Extract significant information relevant to the query and integrate it into the evolutionary map.

5. The retrieval enhancement generation method based on knowledge indexing of graph anchor points according to claim 1, characterized in that, The evolutionary map The specific representation of is as follows: Define evolutionary map ,in For a collection of entities, For a set of relation triples; the evolutionary graph is transformed into a text sequence through linearization: Entities ; Relations: ; in, It is a language-based function that contains entities and their attributes. It is an RDF triple language function that includes a head entity, a relation, and a tail entity.

6. The retrieval enhancement generation method based on knowledge indexing of graph anchor points according to claim 1, characterized in that, After completion After the next iteration of retrieval, if the knowledge output by the large language model in step 2.2 is deemed sufficient or the current iteration count reaches the maximum number of iterations, then all documents retrieved in all steps are combined into a union set to form a complete document set. : 。 7. The retrieval enhancement generation method based on knowledge indexing of graph anchor points according to claim 6, characterized in that, Based on the complete document set Generate instructions Initial query Aggregated document sets and the final evolutionary map According to the large language model Generate the final answer : ； Among them, the evolutionary map As anchor points for structured knowledge, they are used to assist in the organization and interpretation of large models. The information in the middle.

Images