Retrieval enhancement generation method and system based on multi-source knowledge fusion and performance verification method

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a retrieval enhancement generation method based on multi-source knowledge fusion, which comprises the following steps: selecting multi-field knowledge sources such as web pages and Mock APIs through a proxy routing system, extracting related information by utilizing sparse retrieval, dense retrieval and text rearrangement technologies, and performing entity recognition and naming in combination with a large language model. And updating the data by adopting a noise fragment fusion technology. Meanwhile, the invention further provides a system and a performance verification method, and accuracy and consistency of generated results are evaluated through regular matching and a large model. According to the method, multi-source data integration, accurate retrieval and generation evaluation are combined, the accuracy, logicality and continuity of the generated content are improved, and the method is suitable for generation tasks in multiple fields. According to the method, dynamic knowledge selection, multi-step retrieval and reasoning and noise fusion technologies are combined, the overall performance of the retrieval enhancement generation model is effectively improved, and accurate evaluation is carried out.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a retrieval enhancement generation method, system and performance verification method based on multi-source knowledge fusion. Background Art

[0002] With the development of Large-Scale Language Models (LLM), Retrieval-Augmented Generation (RAG) technology has gradually become an important technical means in generation tasks. By combining external knowledge bases and real-time retrieval results, RAG systems can generate more accurate and rich natural language texts, and are widely used in question-answering systems, content generation, knowledge base construction and other fields.

[0003] However, existing RAG systems face multiple challenges: first, due to the complexity and diversity of knowledge sources, how to effectively select and integrate this knowledge becomes the key to system performance; second, how to objectively and systematically evaluate the generation performance of RAG systems is a current technical difficulty. Therefore, it is particularly important to propose a systematic evaluation framework that combines multi-source data and reasoning chains. Summary of the invention

[0004] The purpose of the present invention is to provide a retrieval enhancement generation method, system and performance verification method based on multi-source knowledge fusion to solve the problems raised in the background technology.

[0005] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0006] In a first aspect, the present invention proposes a retrieval enhancement generation method based on multi-source knowledge fusion, the method comprising:

[0007] Pre-configure a first interface end and a second interface end for acquiring multi-domain knowledge sources; the first interface end is configured to acquire web page end content, and the second interface end is configured to acquire Mock API interface end content;

[0008] Receiving a multi-domain query request from a user, the first interface end generates first modal data in response to the request, and the second interface end generates second modal data in response to the request;

[0009] Retrieving the divided first modal data based on a preset first retrieval strategy to obtain at least one text block related to the first modal data, wherein the first retrieval strategy includes performing sparse retrieval, dense retrieval, and text rearrangement;

[0010] After performing entity recognition on the second modal data based on the large language model, the entity information returned by the Mock API interface end is retrieved and converted based on the preset second search strategy to generate natural language form data;

[0011] The text block and the natural language form data are updated in combination with the noise segment fusion technology, and the updated text block and the natural language form data are fused based on the chain reasoning method to generate a retrieval result corresponding to the query request.

[0012] Furthermore, before the first interface end and the second interface end for pre-configuring the acquisition of multi-domain knowledge sources, the method further includes:

[0013] Configuring a proxy routing system driven by a large language model, the proxy routing system is used to receive the user's multi-domain query request, respond and generate a jump instruction, and the jump instruction is configured to jump to the first interface end and / or the second interface end;

[0014] After the proxy routing system jumps to the second interface end according to the jump instruction, the large language model preferentially determines the corresponding API interface according to the entity type related to the user's multi-domain query request.

[0015] Furthermore, the method further comprises:

[0016] When the proxy routing system receives the user's multi-domain query request, the large language model preferentially identifies the domain tags in the user's query request, and generates the jump instruction with specific domain knowledge according to the domain tags.

[0017] Furthermore, the first modal data is specifically unstructured data, including web page text, images, videos or audio; the second modal data is specifically structured data, including data in the form of knowledge graphs.

[0018] Furthermore, the first search strategy is specifically:

[0019] The BM25 retrieval method is used to retrieve the first modal data after the block division to obtain the first relevant data;

[0020] Using the bge-m3 dense search method to calculate the first related data based on coding similarity to obtain second related data;

[0021] The second related data is sorted using a rearrangement tool to filter out text blocks related to the user's multi-domain query request.

[0022] Furthermore, the second search strategy is specifically:

[0023] Performing entity recognition and naming on the second modal data based on the large language model, and filtering out entities related to the user's multi-domain query request;

[0024] Determine, through the proxy routing system, an API interface related to the screened entity;

[0025] The information returned by the API interface is converted into natural language data.

[0026] Furthermore, the updating of the text block and the natural language form data by combining the noise segment fusion technology includes:

[0027] After the proxy routing system jumps to the first interface end and / or the second interface end, the fields to which the first modal data and the second modal data belong are matched with the field labels, and the first modal data and the second modal data that have not been matched are screened out as noise segments;

[0028] The noise segment is added to the text block and the natural language form data to complete updating the text block and the natural language form data.

[0029] In a second aspect, the present invention proposes a retrieval enhancement generation system based on multi-source knowledge fusion, the system comprising:

[0030] A first interface module is used to pre-configure and obtain first modal data in a multi-domain knowledge source, wherein the first modal data is specifically unstructured data, including web page text, images, videos or audios, and the first interface module is configured to generate the first modal data from the web page content in response to a user's multi-domain query request by receiving the request;

[0031] The second interface module is used to pre-configure and obtain the second modal data in the multi-domain knowledge source, the second modal data is specifically structured data, including knowledge graph form data, and the second interface module is configured to generate the second modal data from the Mock API interface end content in response to the request by receiving the user's multi-domain query request;

[0032] A proxy routing module, configured with a proxy routing system driven by a large language model, for receiving the user's multi-domain query request, identifying the domain label in the query request, generating a jump instruction according to the domain label, and jumping the request to the first interface module and / or the second interface module in response to the jump instruction, and determining the corresponding API interface according to the entity type related to the user's query request;

[0033] A first retrieval module, configured to retrieve the first modal data after segmentation based on a preset first retrieval strategy to obtain at least one text block related to the first modal data, wherein the first retrieval strategy includes performing sparse retrieval, dense retrieval, and text rearrangement;

[0034] A second retrieval module is used to perform entity recognition on the second modal data based on the large language model, retrieve the entity information returned by the Mock API interface end based on a preset second retrieval strategy, and convert it to generate natural language form data;

[0035] A fusion processing module is used to update the text block and the natural language form data in combination with the noise segment fusion technology, and fuse the updated text block and the natural language form data based on the chain reasoning method to generate a retrieval result corresponding to the query request.

[0036] Furthermore, the first retrieval module specifically uses the BM25 retrieval method to retrieve the first modal data after segmentation to obtain first relevant data, uses the bge-m3 dense retrieval method to calculate the first relevant data based on encoding similarity to obtain second relevant data, and uses a rearrangement tool to sort the second relevant data to filter out text blocks related to the user's multi-field query request;

[0037] The second retrieval module specifically performs entity recognition and naming on the second modal data based on the large language model, filters out entities corresponding to the user's multi-domain query request, determines the API interface related to the filtered entity through the proxy routing module, and converts the information returned by the API interface into natural language form data;

[0038] After the proxy routing module jumps into the first interface module and / or the second interface module, the fusion processing module matches the fields to which the first modal data and the second modal data belong with the field label, filters out the first modal data and the second modal data that have not completed the match as noise segments, and adds the noise segments to the text block and the natural language form data to complete the update of the text block and the natural language form data.

[0039] In a third aspect, the present invention proposes a performance verification method, which is applied to the performance verification of a retrieval enhancement generation system based on multi-source knowledge fusion as described in claim 8, comprising:

[0040] Unify the standard answers corresponding to the preset user multi-field query requests and the letters in the search results into lowercase, and output a first verification indicator through a string matching algorithm;

[0041] Performing coherence and consistency evaluation on the retrieval results through large language model evaluation, and outputting a second verification indicator;

[0042] A performance verification result is determined based on the first verification indicator and the second verification indicator.

[0043] The beneficial effects of the present invention are:

[0044] 1. The present invention effectively solves the problem that the knowledge sources in the existing retrieval enhancement generation (RAG) system are complex and diverse and difficult to effectively select and integrate by proposing a retrieval enhancement generation method based on multi-source knowledge fusion. By configuring the first interface end and the second interface end for acquiring multi-domain knowledge sources, the web page content and the Mock API interface end content are acquired respectively, thereby realizing the comprehensive utilization of unstructured data and structured data. At the same time, with the help of a proxy routing system driven by a large language model, the best knowledge source can be intelligently selected according to the user's multi-domain query request, thereby improving the accuracy and efficiency of knowledge retrieval. This innovative method not only optimizes the knowledge selection process, but also provides a rich and high-quality knowledge base for subsequent retrieval and generation tasks.

[0045] 2. The present invention also proposes a performance verification method for evaluating the performance of the above-mentioned retrieval enhancement generation system based on multi-source knowledge fusion. This method comprehensively evaluates the accuracy and coherence of the retrieval results by combining a string matching algorithm with a large language model evaluation. The string matching algorithm ensures a high degree of consistency in content between the retrieval results and the standard answers, while the large language model evaluation further verifies the logical coherence and rationality of the retrieval results. The introduction of this performance verification method not only improves the objectivity and accuracy of the system evaluation, but also provides strong support for the continuous optimization and improvement of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of a flow chart of a retrieval enhancement generation method based on multi-source knowledge fusion proposed by the present invention;

[0047] Figure 2 A technical roadmap for the retrieval enhancement generation method based on multi-source knowledge fusion proposed by the present invention;

[0048] Figure 3 A structural diagram of a retrieval enhancement generation system based on multi-source knowledge fusion proposed by the present invention;

[0049] Figure 4 A schematic flow chart of the performance verification method proposed in the present invention. DETAILED DESCRIPTION

[0050] The present application is further described in detail below in conjunction with the accompanying drawings. It is necessary to point out here that the following specific implementation methods are only used to further illustrate the present application and cannot be understood as limiting the scope of protection of the present application. Technical personnel in this field can make some non-essential improvements and adjustments to the present application based on the above application content.

[0051] Example 1

[0052] like Figure 1-2 As shown, this embodiment proposes a retrieval enhancement generation method based on multi-source knowledge fusion, the method comprising:

[0053] (1) Pre-configure a first interface end and a second interface end for acquiring multi-domain knowledge sources; the first interface end is configured to acquire web page content, and the second interface end is configured to acquire Mock API interface end content.

[0054] (2) Before pre-configuring the first interface end and the second interface end for acquiring the multi-domain knowledge source, the method further includes:

[0055] A proxy routing system driven by a large language model is configured, the proxy routing system is used to receive a user's multi-domain query request, respond and generate a jump instruction, and the jump instruction is configured to jump to the first interface end and / or the second interface end;

[0056] After the proxy routing system jumps to the second interface end according to the jump instruction, the large language model preferentially determines the corresponding API interface according to the entity type related to the user's multi-domain query request.

[0057] (3) receiving a multi-domain query request from a user, the first interface end generating first modal data in response to the request, and the second interface end generating second modal data in response to the request;

[0058] (4) searching the first modal data after segmentation based on a preset first search strategy to obtain at least one text block related to the first modal data, wherein the first search strategy includes performing sparse search, dense search, and text rearrangement;

[0059] (5) After performing entity recognition on the second modal data based on the large language model, the entity information returned by the Mock API interface end is retrieved and converted based on the preset second search strategy to generate natural language form data;

[0060] (6) The text blocks and natural language form data are updated by combining the noise fragment fusion technology, and the updated text blocks and natural language form data are fused based on the chain reasoning method to generate the retrieval results corresponding to the query request.

[0061] More specifically, this embodiment uses a fine-tuned large language model to determine the external knowledge source that should be used according to the query, and determines the prompt (an initial input or prompt of the large language model) template to be used according to the selection result to determine whether to add the text block or the natural language form data, or add both to the prompt as external knowledge. At the same time, the prompt instructs the large language model to first generate a specific thinking process, and then generates the search results corresponding to the query request according to the thinking process to realize chain reasoning.

[0062] Further preferably, the method further comprises:

[0063] When the proxy routing system receives a user's multi-domain query request, the large language model prioritizes identifying the domain tags in the user's query request and generates jump instructions with specific domain knowledge based on the domain tags.

[0064] Further preferably, the first modal data is specifically unstructured data, including web page text, images, videos or audios; and the second modal data is specifically structured data, including two-dimensional structure table data.

[0065] Further preferably, the first search strategy is specifically:

[0066] The BM25 retrieval method is used to retrieve the first modal data after the block division to obtain the first relevant data;

[0067] The bge-m3 dense retrieval method is used to calculate the first relevant data based on the coding similarity to obtain the second relevant data;

[0068] The second related data is sorted using a rearrangement tool to filter out text blocks related to the user's multi-domain query request.

[0069] Further preferably, the second search strategy is specifically:

[0070] Perform entity recognition and naming on the second modality data based on the large language model, and filter out entities related to the user's multi-domain query requests;

[0071] Determine the API interface related to the filtered entity through the proxy routing system;

[0072] Convert the information returned by the API interface into natural language data.

[0073] It is understandable that in the second search strategy, if Mock API (second interface end) is selected, the named entities contained in the query are first extracted using a large language model as keywords for API search, and then the API interface used for the query is selected, and the entity is used as API input to obtain entity-related information returned by the API.

[0074] Further preferably, updating the text block and the natural language form data in combination with the noise segment fusion technology includes:

[0075] After the proxy routing system jumps to the first interface end and / or the second interface end, the domains to which the first modal data and the second modal data belong are matched with the domain labels, and the first modal data and the second modal data that have not been matched are screened out as noise fragments;

[0076] Noise fragments are added to text blocks and natural language form data. The small amount of noise fragments added can warm up the large language model without hindering it from extracting effective information, thereby improving the reasoning ability and completing the update of text blocks and natural language form data.

[0077] It can be understood that the search enhancement generation method in this embodiment includes the following steps when implemented:

[0078] (1) Knowledge source configuration and selection:

[0079] The first interface and the second interface for acquiring multi-domain knowledge sources are pre-configured. The first interface is configured to acquire web page content, mainly including unstructured data such as web page text, images, videos or audio; the second interface is configured to acquire Mock API interface content, mainly including structured data such as knowledge graph form data.

[0080] A proxy routing system driven by a large language model is configured to receive multi-domain query requests from users, identify domain tags in the query requests, and generate jump instructions according to the domain tags to jump to the first interface end and / or the second interface end.

[0081] (2) Data generation and retrieval:

[0082] A user multi-domain query request is received, the first interface end generates first modal data in response to the request, and the second interface end generates second modal data in response to the request.

[0083] The first modal data (unstructured data) is divided into blocks and then retrieved based on a preset first retrieval strategy. The first retrieval strategy includes performing sparse retrieval (such as using the BM25 retrieval method), dense retrieval (such as using the bge-m3 dense retrieval method) and text re-ranking (such as using the bge-reranker-v2-m3 model for re-ranking) to obtain at least one text block related to the user query request.

[0084] For the second modality data (structured data), entity recognition and naming are performed based on the large language model to filter out entities related to the user's query request. Then, the proxy routing system determines the API interface related to the filtered entity and converts the information returned by the API interface into natural language data.

[0085] (3) Noise fragment fusion and reasoning:

[0086] After the proxy routing system jumps to the first interface end and / or the second interface end, the domains to which the first modal data and the second modal data belong are matched with the domain labels, and the first modal data and the second modal data that have not been matched are screened out as noise fragments.

[0087] Noise fragments are added to text blocks and natural language form data to complete the data update. At the same time, the thinking chain prompt method is adopted to allow the large model to perform more logical reasoning and gradually deduce the correct answer to the question.

[0088] like Figure 3 As shown, this embodiment also proposes a retrieval enhancement generation system based on multi-source knowledge fusion, the system comprising:

[0089] The first interface module is used to pre-configure and obtain first modal data in the multi-domain knowledge source, where the first modal data is specifically unstructured data, including web page text, images, videos or audios. The first interface module is configured to generate the first modal data from the web page content in response to the request by receiving a user's multi-domain query request;

[0090] The second interface module is used to pre-configure and obtain the second modality data in the multi-domain knowledge source, the second modality data is specifically structured data, including two-dimensional structure table data, and the second interface module is configured to generate the second modality data from the content of the Mock API interface end in response to the request by receiving the user's multi-domain query request;

[0091] The proxy routing module is configured with a proxy routing system driven by a large language model, and is used to receive multi-domain query requests from users, identify domain tags in the query requests, generate jump instructions according to the domain tags, and jump the requests to the first interface module and / or the second interface module in response to the jump instructions, and determine the corresponding API interface according to the entity type related to the user query request;

[0092] A first retrieval module is used to retrieve the first modal data after segmentation based on a preset first retrieval strategy to obtain at least one text block related to the first modal data, wherein the first retrieval strategy includes performing sparse retrieval, dense retrieval and text rearrangement;

[0093] The second retrieval module is used to perform entity recognition on the second modal data based on the large language model, retrieve the entity information returned by the Mock API interface end based on the preset second retrieval strategy, and convert it to generate natural language form data;

[0094] The fusion processing module is used to update the text block and natural language form data by combining the noise fragment fusion technology, and fuse the updated text block and natural language form data based on the chain reasoning method to generate the retrieval result of the corresponding query request.

[0095] It should be noted that, in this embodiment, the proxy routing module is not only used for users to input multi-field query requests, but also used for receiving and displaying the search results generated by the fusion processing module.

[0096] Further preferably, the first retrieval module specifically uses the BM25 retrieval method to retrieve the first modal data after segmentation to obtain the first relevant data, uses the bge-m3 dense retrieval method to calculate the first relevant data based on the encoding similarity to obtain the second relevant data, and uses the reranking tool (using the bge-reranker-v2-m3 model to rerank the text blocks) to sort the second relevant data, and screen out the text blocks related to the user's multi-field query request;

[0097] The second retrieval module specifically performs entity recognition and naming on the second modal data based on the large language model, screens out entities related to the user's multi-domain query request, determines the API interface related to the screened entity through the proxy routing module, and converts the information returned by the API interface into natural language form data;

[0098] After the proxy routing module jumps into the first interface module and / or the second interface module, the fusion processing module matches the domains to which the first modal data and the second modal data belong with the domain labels, filters out the first modal data and the second modal data that have not completed the match as noise fragments, and adds the noise fragments to the text block and the natural language form data to complete the update of the text block and the natural language form data.

[0099] In a specific implementation, the system implementation of the present invention includes a first interface module, a second interface module, a proxy routing module, a first search module, a second search module and a fusion processing module. The specific functions of each module are as follows:

[0100] The first interface module is responsible for generating the first modal data, i.e., unstructured data, from the web page content.

[0101] The second interface module is responsible for generating the second modal data, that is, structured data, from the content of the Mock API interface.

[0102] Proxy routing module: It is equipped with a proxy routing system driven by a large language model, which is used to receive user query requests and select the corresponding knowledge source for retrieval based on the domain labels in the query requests.

[0103] The first retrieval module: uses sparse retrieval, dense retrieval and text rearrangement technology to search the first modal data after segmentation to obtain the most relevant information.

[0104] The second retrieval module: performs entity recognition on the second modal data based on the large language model, selects the relevant API interface for entity query through the proxy routing system, and converts the return results of the API into natural language data.

[0105] Fusion processing module: Combining noise fragment fusion technology and chain reasoning method, it integrates and infers information from different knowledge sources to generate the final retrieval results.

[0106] like Figure 4 As shown, this embodiment also proposes a performance verification method, which is applied to the performance verification of the above-mentioned retrieval enhancement generation system based on multi-source knowledge fusion, including:

[0107] The standard answers corresponding to the preset user multi-field query requests and the letters in the search results are unified into lowercase, and the first verification index is output through a string matching algorithm;

[0108] The coherence and consistency of the retrieval results are evaluated through a large language model, and a second verification indicator is output;

[0109] A performance verification result is determined based on the first verification indicator and the second verification indicator.

[0110] As an example, according to all the contents of the above embodiments, there are the following method implementation methods:

[0111] (1) Knowledge selection:

[0112] Through the LLM (Large Language Model) driven proxy routing system, relevant content is selected from multiple knowledge sources (such as web pages, Mock APIs, etc.) according to multi-domain queries (such as sports, movies, etc.) input by users. In the present invention, the first interface end and the second interface end are respectively configured to obtain web page content and Mock API interface end content.

[0113] (2) Knowledge retrieval:

[0114] For web page knowledge, we use broad and focused retrieval techniques. First, we obtain large-scale relevant data through the first interface module, and use the first retrieval module to perform sparse retrieval to quickly filter out irrelevant information. Then, we perform dense retrieval to screen the remaining information more carefully. Finally, we use text rearrangement technology to sort the retrieved text blocks and select the most relevant information.

[0115] For structured knowledge in the API, the second interface module generates second modal data after receiving the user query request. The second retrieval module performs entity recognition on the second modal data based on the large language model to find the entity in the user query. Then, the proxy routing system is used to select the API interface related to the user query to perform entity query and obtain entity information. Finally, the return result of the API is converted into natural language form data.

[0116] (3) Knowledge fusion and reasoning:

[0117] The fusion processing module combines the noise fragment fusion technology to integrate information from different knowledge sources. In this process, the domain to which the user query belongs is detected by designing a route, and question-answer examples from other domains are selected as noise fragments and added to the retrieved text blocks and natural language form data. Then, based on the chain reasoning method, multi-step reasoning methods (such as thinking chain prompts) are used to gradually deduce the answer to the question.

[0118] (4) Performance evaluation:

[0119] The generated search results are verified using the performance verification method. First, the letters in the generated results and the standard answers are unified into lowercase, and the first verification index is output through the string matching algorithm to verify the accuracy of the generated content. Secondly, the coherence and consistency of the generated content are evaluated through the large language model, and the second verification index is output. Finally, the performance verification result is determined based on the first verification index and the second verification index.

[0120] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features can be replaced by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A retrieval enhancement generation method based on multi-source knowledge fusion, characterized in that: The method comprises: Pre-configure a first interface end and a second interface end for acquiring multi-domain knowledge sources; the first interface end is configured to acquire web page content, and the second interface end is configured to acquire Mock API interface end content; Receiving a multi-domain query request from a user, the first interface end generates first modal data in response to the request, and the second interface end generates second modal data in response to the request; Retrieving the divided first modal data based on a preset first retrieval strategy to obtain at least one text block related to the first modal data, wherein the first retrieval strategy includes performing sparse retrieval, dense retrieval, and text rearrangement; After performing entity recognition on the second modal data based on the large language model, the entity information returned by the Mock API interface end is retrieved and converted based on the preset second search strategy to generate natural language form data; The text block and the natural language form data are updated in combination with the noise segment fusion technology, and the updated text block and the natural language form data are fused based on the chain reasoning method to generate a retrieval result corresponding to the query request.

2. According to the method for retrieval enhancement generation based on multi-source knowledge fusion according to claim 1, it is characterized in that: Before the first interface end and the second interface end of pre-configuring the acquisition of multi-domain knowledge sources, the method further includes: Configuring a proxy routing system driven by a large language model, the proxy routing system is used to receive the user's multi-domain query request, respond and generate a jump instruction, and the jump instruction is configured to jump to the first interface end and / or the second interface end; After the proxy routing system jumps to the second interface end according to the jump instruction, the large language model preferentially determines the corresponding API interface according to the entity type related to the user's multi-domain query request.

3. The retrieval enhancement generation method based on multi-source knowledge fusion according to claim 2 is characterized in that: The method further comprises: When the proxy routing system receives the user's multi-domain query request, the large language model preferentially identifies the domain tags in the user's query request, and generates the jump instruction with specific domain knowledge according to the domain tags.

4. The retrieval enhancement generation method based on multi-source knowledge fusion according to claim 1 is characterized in that: The first modal data is specifically unstructured data, including web page text, images, videos or audio; the second modal data is specifically structured data, including data in the form of knowledge graphs.

5. The retrieval enhancement generation method based on multi-source knowledge fusion according to claim 3 is characterized in that: The first search strategy is specifically: The BM25 retrieval method is used to retrieve the first modal data after the block division to obtain the first relevant data; Using the bge-m3 dense search method to calculate the first related data based on coding similarity to obtain second related data; The second related data is sorted using a rearrangement tool to filter out text blocks related to the user's multi-domain query request.

6. The retrieval enhancement generation method based on multi-source knowledge fusion according to claim 5 is characterized in that: The second search strategy is specifically: Performing entity recognition and naming on the second modal data based on the large language model, and filtering out entities related to the user's multi-domain query request; Determine, through the proxy routing system, an API interface related to the screened entity; The information returned by the API interface is converted into natural language data.

7. The retrieval enhancement generation method based on multi-source knowledge fusion according to claim 6 is characterized in that: The updating of the text block and the natural language form data by combining the noise segment fusion technology includes: After the proxy routing system jumps to the first interface end and / or the second interface end, the fields to which the first modal data and the second modal data belong are matched with the field labels, and the first modal data and the second modal data that have not been matched are screened out as noise segments; The noise segment is added to the text block and the natural language form data to complete updating the text block and the natural language form data.

8. A retrieval enhancement generation system based on multi-source knowledge fusion, characterized in that: The system comprises: A first interface module is used to pre-configure and obtain first modal data in a multi-domain knowledge source, wherein the first modal data is specifically unstructured data, including web page text, images, videos or audios, and the first interface module is configured to generate the first modal data from the web page content in response to a user's multi-domain query request by receiving the request; The second interface module is used to pre-configure and obtain the second modal data in the multi-domain knowledge source, the second modal data is specifically structured data, including knowledge graph form data, and the second interface module is configured to generate the second modal data from the Mock API interface end content in response to the request by receiving the user's multi-domain query request; A proxy routing module, configured with a proxy routing system driven by a large language model, for receiving the user's multi-domain query request, identifying the domain label in the query request, generating a jump instruction according to the domain label, and jumping the request to the first interface module and / or the second interface module in response to the jump instruction, and determining the corresponding API interface according to the entity type related to the user's query request; A first retrieval module, configured to retrieve the first modal data after segmentation based on a preset first retrieval strategy to obtain at least one text block related to the first modal data, wherein the first retrieval strategy includes performing sparse retrieval, dense retrieval, and text rearrangement; A second retrieval module is used to perform entity recognition on the second modal data based on the large language model, retrieve the entity information returned by the Mock API interface end based on a preset second retrieval strategy, and convert it to generate natural language form data; A fusion processing module is used to update the text block and the natural language form data in combination with the noise segment fusion technology, and fuse the updated text block and the natural language form data based on the chain reasoning method to generate a retrieval result corresponding to the query request.

9. The retrieval enhancement generation system based on multi-source knowledge fusion according to claim 8 is characterized in that: The first retrieval module specifically uses the BM25 retrieval method to retrieve the first modal data after segmentation to obtain first relevant data, uses the bge-m3 dense retrieval method to calculate the first relevant data based on encoding similarity to obtain second relevant data, and uses a rearrangement tool to sort the second relevant data to filter out text blocks related to the user's multi-field query request; The second retrieval module specifically performs entity recognition and naming on the second modal data based on the large language model, filters out entities corresponding to the user's multi-domain query request, determines the API interface related to the filtered entity through the proxy routing module, and converts the information returned by the API interface into natural language form data; After the proxy routing module jumps into the first interface module and / or the second interface module, the fusion processing module matches the fields to which the first modal data and the second modal data belong with the field label, filters out the first modal data and the second modal data that have not completed the match as noise segments, and adds the noise segments to the text block and the natural language form data to complete the update of the text block and the natural language form data.

10. A performance verification method, characterized in that: Applied to the performance verification of a retrieval enhancement generation system based on multi-source knowledge fusion as described in claim 8, comprising: Unify the standard answers corresponding to the preset user multi-field query requests and the letters in the search results into lowercase, and output a first verification indicator through a string matching algorithm; Performing coherence and consistency evaluation on the retrieval results through large language model evaluation, and outputting a second verification indicator; A performance verification result is determined based on the first verification indicator and the second verification indicator.

Images