Large language model agent mass tool calling method and system based on knowledge graph retrieval enhancement

Abstract

The invention relates to the technical field of agent application of a large language model, in particular to a large language model agent mass tool calling method and system based on knowledge graph retrieval enhancement, and the method comprises a tool graph construction stage, a tool semantic knowledge base construction stage and a tool retrieval calling stage. The method has the beneficial effect that the tool combination calling capability of the model is improved through the tool dependency relationship structure of the knowledge graph. Through vector retrieval and pre-retrieval of the knowledge graph, the number scale of tools for direct intention recognition of the large language model is reduced, and the error rate of multi-tool selection and tool parameter extraction slot positions of the large language model is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agent application of large language models, in particular to a large language model agent mass tool calling method and system based on knowledge graph retrieval enhancement. BACKGROUND

[0002] With the continuous extension of the application scenarios of large language models, the demand for the active calling of large language models to tools in the tool library and the multi-round interaction with the environment is highlighted. The existing large language models generally use the method of storing and retrieving related tools from the vector database on a large scale, and then using the large language model based on the small-scale recall results for further intent recognition and tool calling, to improve the tool selection and calling ability of the model and reduce the tool calling and execution error rate. However, this method has the inherent limitation of missing representation of the association and dependence relationship between tools, which makes the large language model often have tool error calling, intent recognition failure or tool calling omission in the mass tool calling task as the size of the tool library increases.

[0003] The traditional method uses a keyword matching method for candidate tool retrieval. The new mass tool management method based on RAG knowledge base generally uses a semantic matching method for storage and retrieval. However, in actual application, although the existing method can play a role in improving the model tool calling ability in the conventional scale scenario. However, in the mass tool calling scenario, its semantic vector retrieval scheme still has considerable limitations and cannot directly meet the diversified needs of the model for tool dependence relationship and similar tool recognition and calling in the tool library.

[0004] Therefore, the present application provides a mass tool management and retrieval method based on knowledge graph assistance to overcome the above limitations. SUMMARY

[0005] The purpose of the present application is to provide a large language model agent mass tool calling method and system based on knowledge graph retrieval enhancement to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides the following technical solution: a large language model agent mass tool calling method based on knowledge graph retrieval enhancement, comprising a tool graph construction stage, a tool semantic knowledge base construction stage and a tool retrieval calling stage; The tool graph construction stage comprises: Triplet extraction: according to the graph structure described by the knowledge graph module, custom prompt words are formulated for graph extraction, and graph triplets and relationship extraction are performed on the text based on the custom prompt words by using the large language model, concept nodes and node relationship edges are constructed, and the extraction results are converted into graph database instructions and then entered into the prompt graph database; Semantic disambiguation and context node merging: using a semantic encoder model to encode the text semantics of entity nodes to obtain node semantic similarity, using a knowledge graph semantic disambiguation algorithm to perform semantic disambiguation on the knowledge graph constructed by open extraction, combining semantic similarity and graph community discovery algorithm to perform clustering and community discovery of graph nodes, and forming a hierarchical structure with relatively dense relationship edges within the community and relatively sparse relationship edges between communities; The tool semantic knowledge base construction stage includes: Tool name and description text synthesis: according to the preset field, design a semantic coding prompt word template, and substitute the tool specific field text into the prompt word template to synthesize a complete context prompt word; Synthetic text semantic vector encoding and database: using a semantic vector encoding model to encode the synthesized context prompt word to obtain a semantic vector, and injecting the semantic vector as a field of the corresponding knowledge graph node into the knowledge graph database, realizing the fusion of the graph and the vector library; The tool retrieval and calling stage includes question instruction rewriting and expansion, vector initial retrieval, recall result reordering, graph seed tool node expansion, graph expansion recall result reordering, and tool calling steps.

[0007] Preferably, the tool retrieval and calling stage specifically includes: Question instruction rewriting and expansion: design a question rewriting and expansion prompt word, substitute the user's original question prompt word into the prompt word template, obtain a diversified question retrieval text after expansion and rewriting, and improve the retrieval recall effect; Vector initial retrieval: using a semantic vector encoding model to encode the expanded instruction to obtain a semantic vector, and performing vector similarity retrieval on the vector library to obtain an initial recall tool set; Recall result reordering: using a reordering model to reorder the initial recall tools and take out the Topn result; Graph seed tool node expansion: taking the seed tool node recalled in the previous step Topn result as the starting point, respectively running the graph depth-first search algorithm to expand the associated tools and auxiliary tools of the tool seed node; Graph expansion recall result reordering: design a prompt word, reorder the tools recalled by the graph expansion traversal combined with the prompt word instruction, and take out the TopK tools after reordering as the final result of tool retrieval; Tool calling: substitute the final result of tool retrieval into a specific prompt word combined with the user's original question, and the large model performs the last intent recognition and tool calling according to the K tools and performs data extraction of parameter slots.

[0008] Preferably, in the triple extraction step of the tool graph construction phase, a customized prompt word is formulated according to the explicit graph structure in the knowledge graph module, which is used to guide the large language model to accurately identify and extract graph triples and relationships in the text. The extracted content includes concept nodes and relationship edges between nodes. After extraction, the results are converted into instructions that meet the requirements of the graph database and are entered into the prompt graph database.

[0009] Preferably, in the vector initial retrieval step of the tool retrieval calling phase, the semantic vector encoding model converts the expanded instructions into vector representations with specific semantic features when performing semantic encoding. These vectors accurately reflect the semantic information of the instructions. By calculating the similarity between the vectors in the vector library, the vector closest to the instruction semantics is found, and the initial recall tool set is obtained.

[0010] Preferably, in the tool calling step of the tool retrieval calling phase, the large model performs intent recognition and tool calling based on the reordered TopK tools. Since K is much smaller than the number of tools N in the massive tool library, the number of tools processed by the large model is reduced, the computational complexity is reduced, the accuracy of tool calling and parameter extraction is improved, and the final called tool is highly matched with the user's original questioning intent, and the relevant parameter slot data is accurately extracted.

[0011] A large language model agent massive tool calling system based on knowledge graph retrieval enhancement, comprising a massive tool dependency relationship knowledge graph module and a massive tool semantic vector retrieval knowledge base module; In the massive tool dependency relationship knowledge graph module, the graph nodes represent available tools for agents, and the node attributes include tool name, tool function description, tool parameter name, tool parameter type, tool parameter meaning, and tool description semantic vector fields. The fusion retrieval of knowledge graph and vector retrieval is realized, and each node is directly recalled as a seed node for graph access through vector retrieval. Graph nodes are divided into independent tool nodes and auxiliary tool nodes. Independent tool nodes can directly or indirectly meet the user's instruction requirements, and auxiliary tool nodes serve independent tool nodes. Graph edges are divided into four types: "tool directly depends on", "tool indirectly depends on", "parameter directly depends on", and "parameter indirectly depends on", which are used to represent the dependency and hierarchical relationships of massive tools. The massive tool semantic vector retrieval knowledge base module is composed of a large number of retrievable semantic vectors encoded by semantic models. The encoded texts include the name field, function description field, and formal parameter description field of the tool functions in the tool library. These description information is injected into the knowledge base through the tool registration mechanism, and the recall of the graph retrieval seed node is realized in the retrieval stage.

[0012] Preferably, in the mass tool dependency relationship knowledge graph module, the tool description semantic vector field of the graph node attribute is encoded by a semantic model to obtain a vector representation of tool-related text, and the vector can accurately reflect the semantic information of the tool, so that the vector retrieval mode can be combined in the knowledge graph retrieval process to improve the accuracy and comprehensiveness of retrieval, and comprehensive representation of tool information is realized from the semantic level and the graph structure level.

[0013] Preferably, in the mass tool dependency relationship knowledge graph module, the division mode of the independent tool node and the auxiliary tool node is that the independent tool node refers to a tool node whose execution result generally directly meets the user instruction demand, and the auxiliary tool node refers to a tool node that generally serves the independent tool node; through this division, the true value tool meeting the user's intention can be clearly found in the tool retrieval process, and the efficiency and accuracy of tool calling are improved.

[0014] Preferably, in the mass tool dependency relationship knowledge graph module, the four types of edges are defined as follows: "tool directly depends on" means that a tool must execute the dependent tool to execute normally; "tool indirectly depends on" means that a tool can optionally execute the dependent tool, and the optional condition depends on the intention instruction input by the user; "parameter directly depends on" means that the input parameter of a tool must be obtained by executing its dependent tool; "parameter indirectly depends on" means that the input parameter of a tool can be optionally obtained by executing its dependent tool.

[0015] Preferably, in the mass tool semantic vector retrieval knowledge base module, the name field is the concise and easily distinguishable command name of the function tool, the function description field is the perfect and detailed function text introduction of the tool function, and the formal parameter description field is the complete description text of the type and meaning of all formal parameters of the tool function; by injecting these description information into the semantic vector retrieval knowledge base, the model can fully understand the function of all calling tools and the difference between the corresponding function and parameter slot filling semantics in the retrieval stage, fine-grained tool recognition and intention recognition are realized, and the accuracy and pertinence of the mass tool calling of the large language model agent are improved.

[0016] Compared with the prior art, the present application has the following advantages: The application provides a mass tool calling method and system of a large language model intelligent agent based on knowledge graph retrieval enhancement, which improves the tool combination calling capability of the model through the tool dependency relationship structure of the knowledge graph. Through vector retrieval and pre-retrieval of the knowledge graph, the number of tools for direct intent recognition by the large language model is reduced, and the error rate of tool selection and tool parameter extraction slot of the large language model is reduced. Through vector retrieval and pre-retrieval of the knowledge graph, the number of tools for direct intent recognition by the large language model is reduced, and the reasoning algorithm and memory overhead of the large language model are reduced, and the reasoning response speed is improved. Through vector retrieval and pre-retrieval of the knowledge graph, the length of the input large language model context prompt word is reduced, and the problem of large language model prompt word length window overflow is alleviated. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The method flowchart of the application. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme of the application clear, complete and the advantages more clear and obvious, the embodiments of the application will be further described in detail below with reference to the drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the application, not all embodiments, and are only used to explain the embodiments of the application, and do not limit the embodiments of the application. All other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0019] Embodiment one, the application provides a technical scheme: a mass tool calling method of a large language model intelligent agent based on knowledge graph retrieval enhancement, comprising the following steps: (1) Tool graph construction stage This stage mainly includes the following two steps.

[0020] a. Triple extraction: according to the graph structure described in the knowledge graph module, formulate graph extraction customized prompt words. Use the large language model to extract graph triples and relationships based on the customized prompt word instructions. Specifically, extract and construct concept nodes and node relationship edges. And the extraction result is converted into a graph database instruction and then entered into the prompt graph database.

[0021] b. Semantic disambiguation and context node merging: use a semantic encoder model to encode the text semantics of entity nodes to obtain node semantic similarity. Use a knowledge graph semantic disambiguation algorithm to perform semantic disambiguation on the knowledge graph obtained by open extraction and construction. At the same time, for example, the semantic similarity combines the community discovery algorithm of the graph to perform clustering and community discovery of the graph nodes, realize the hierarchical structure that the relationship edges are relatively dense within the community and the relationship edges are relatively sparse between the communities, and create a favorable hierarchical structure for graph retrieval.

[0022] (2) Tool semantic knowledge base construction phase This phase mainly includes the following two steps.

[0023] a. Tool name and description text synthesis: according to the described field, design a semantic coding prompt word template, and synthesize a complete context prompt word by substituting the tool specific field text into the prompt word template.

[0024] b. Synthesis text semantic vector encoding and database: use the semantic vector encoding model to encode the synthesized context prompt word to obtain the semantic vector. Inject the semantic vector as a field of the corresponding knowledge graph node into the knowledge graph database, and realize the fusion of the graph and the vector library.

[0025] (3) Tool retrieval calling phase This phase includes the following steps.

[0026] a. Rewriting and expanding the question instruction: design a rewriting and expanding prompt word for the question Query text, substitute the user's original question prompt word into the prompt word template, and obtain a diversified question retrieval text after expanding and rewriting. Improve the retrieval recall effect.

[0027] b. Vector initial retrieval: use the semantic vector encoding model to encode the expanded instruction to obtain the semantic vector. Perform vector similarity retrieval on the vector library using the semantic vector to obtain an initial recall tool set.

[0028] c. Recall result reordering: use the reordering model to reorder the initial recall tools and take out the Topn results.

[0029] d. Graph seed tool node expansion: take the seed tool node recalled in the previous step Topn result as the starting point, and run the graph depth-first search algorithm to expand the associated tools and auxiliary tools of the tool seed node.

[0030] e. Graph expansion recall result reordering: design a prompt word, combine the tools recalled by the graph expansion traversal with the prompt word instruction for reordering, and take out the TopK tools after reordering as the final result of tool retrieval.

[0031] f. Tool calling: substitute the final result of tool retrieval into a specific prompt word combined with the user's original question, and the large model performs the final intent recognition and tool calling according to the K tools and performs parameter slot data extraction. At this time, K is much smaller than the tool quantity scale N of the massive tool library, which can effectively improve the accuracy of tool calling and parameter extraction.

[0032] Embodiment two, on the basis of embodiment one, proposes a large language model agent mass tool calling system based on knowledge graph retrieval enhancement, including (1) Mass tool dependency knowledge graph module In the mass tool knowledge graph of the present scheme, the nodes of the graph represent the tools available to the agent, and the edges of the graph represent the associated dependency relationships between the tools. The attributes of the graph nodes include tool name, tool function description, tool parameter name, tool parameter type, and tool parameter meaning fields, which represent the tool information from multiple angles. In addition, the tool description semantic vector field is added to the graph node attributes, which realizes the fusion retrieval of knowledge graph and vector retrieval. Each graph node can be recalled as a seed node for graph access through vector retrieval.

[0033] The nodes in the graph are divided into two types: independent tool nodes and auxiliary tool nodes. The execution results of independent tool nodes can generally directly meet the user's instruction requirements, such as weather information query tools or flight information query tools, but they may also need to indirectly call auxiliary tool nodes to obtain the data they depend on. Auxiliary tools generally serve independent tool nodes, such as current date tools and unit conversion tools. Through the classification of tool node types, the dependency relationships of mass tools are represented, and the hierarchical relationships between tools are expressed, which helps to discover true values in the tool retrieval process.

[0034] The types of edges in the graph are divided into "tool directly depends on" and "tool indirectly depends on", as well as "parameter directly depends on" and "parameter indirectly depends on". "Tool directly depends on" means that a tool must execute the dependent tool to execute normally. "Tool indirectly depends on" means that a tool can optionally execute the dependent tool, and the optional condition depends on the user's input intent instruction. "Parameter directly depends on" means that the input parameter of a tool must be obtained by executing its dependent tool. "Parameter indirectly depends on" means that the input parameter of a tool can be optionally obtained by executing its dependent tool. For example, when the user's instruction input is "What's the weather tomorrow?", the date acquisition tool needs to be called. When the user's input is "What's the weather on January 1, 2025?", the date acquisition tool does not need to be called. At this time, there is a "parameter indirectly depends on" relationship between the weather query tool and the date query tool.

[0035] (2) Mass tool semantic vector retrieval knowledge base module The tool semantic vector retrieval library module in the scheme is composed of a large number of retrievable semantic vector sets coded by semantic models. The coded text mainly includes the clear and clear description fields of the tool functions in the tool library, including the name field, the function description field and the formal parameter description field. The name field is the concise and easily distinguishable command name of the function tool, the function description field is the perfect and detailed function text introduction of the tool function, and the formal parameter description field is the complete description text of the type and meaning of all formal parameters of the tool function. These description information is injected into the semantic vector retrieval knowledge base through the tool registration mechanism, and in the retrieval stage, the recall of the graph retrieval seed node can be realized, so that the model can fully understand all the calling tool functions and the difference between the corresponding function and parameter slot filling semantics, and realize fine-grained tool recognition and intent recognition.

[0036] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for mass tool invocation of large language model agents based on knowledge graph retrieval enhancement, characterized in that: The tool retrieval calling stage includes question instruction rewriting expansion, vector initial retrieval, recall result reordering, graph seed tool node expansion, graph expansion recall result reordering and tool calling steps. The tool graph construction stage includes: Triplet extraction: form a graph extraction customized prompt word according to the graph structure described by the knowledge graph module, use a large language model to extract graph triplets and relationships from text based on the customized prompt word instruction, construct concept nodes and node relationship edges, and then convert the extraction results into graph database instructions and record them in the prompt graph database; Semantic disambiguation and context node merging: use a semantic encoder model to encode the text semantics of entity nodes to obtain node semantic similarity, use a knowledge graph semantic disambiguation algorithm to disambiguate the knowledge graph constructed by open extraction, combine the semantic similarity and the community discovery algorithm of the graph to cluster and discover communities in the graph, and form a hierarchical structure with relatively dense relationship edges within the community and relatively sparse relationship edges between communities; The tool semantic knowledge base construction stage includes: Tool name and description text synthesis: design a semantic encoding prompt word template according to the preset field, and synthesize a complete context prompt word by substituting the tool field text into the prompt word template; Synthetic text semantic vector encoding and database injection: use a semantic vector encoding model to encode the synthesized context prompt word to obtain a semantic vector, and inject the semantic vector as a field of the corresponding knowledge graph node into the knowledge graph database to realize the fusion of the graph and the vector database; The tool retrieval calling stage includes question instruction rewriting expansion, vector initial retrieval, recall result reordering, graph seed tool node expansion, graph expansion recall result reordering and tool calling steps.

2. The method of claim 1, wherein the method is characterized by: The tool retrieval calling stage specifically includes: Question instruction rewriting expansion: design a question rewriting expansion prompt word, substitute the user's original question prompt word into the prompt word template, and obtain a diversified question retrieval text after rewriting expansion to improve the retrieval recall effect; Vector initial retrieval: use a semantic vector encoding model to encode the expanded instruction to obtain a semantic vector, and perform vector similarity retrieval on the vector database using the semantic vector to obtain an initial recall tool set; Recall result reordering: use a reordering model to reorder the initial recall tools and take out the Topn results; Graph seed tool node expansion: take the seed tool node recalled in the previous step Topn result as the starting point, and run a graph depth-first search algorithm to expand the associated tools and auxiliary tools of the tool seed node; Graph expansion recall result reordering: design a prompt word, reorder the tools recalled by the graph expansion traversal using the prompt word instruction, and take out the TopK tools after reordering as the final result of tool retrieval; Tool calling: substitute the final result of tool retrieval into a specific prompt word combined with the user's original question, and the large model performs the final intent recognition and tool calling according to the K tools and extracts the data of the parameter slot.

3. The method of claim 2, wherein the method is characterized by: In the triple extraction step of the tool graph construction phase, a graph extraction customized prompt word is formulated according to the explicit graph structure in the knowledge graph module, which is used to guide the large language model to accurately identify and extract graph triples and relationships in the text. The extracted content includes concept nodes and relationship edges between nodes. After extraction, the results are converted into instructions that meet the requirements of the graph database and are entered into the prompt graph database.

4. The method of claim 3, wherein the method is characterized by: In the vector initial retrieval step of the tool retrieval and calling phase, the semantic vector encoding model encodes the expanded instructions into vector representations with specific semantic features. These vectors accurately reflect the semantic information of the instructions. By calculating the similarity between the vectors in the vector library, the vector closest to the instruction semantics is found, and the initial recall tool set is obtained.

5. The method of claim 4, wherein the method is characterized by: In the tool calling step of the tool retrieval and calling phase, the large model performs intent recognition and tool calling based on the reordered TopK tools. Since K is much smaller than the number of tools N in the massive tool library, the number of tools processed by the large model is reduced, the computational complexity is reduced, the accuracy of tool calling and parameter extraction is improved, and the final called tool is highly matched with the user's original question intent, and the relevant parameter slot data is accurately extracted.

6. A large language model agent mass tool invocation system based on knowledge graph retrieval enhancement, applying the method of claim 5, characterized in that: It includes a massive tool dependency relationship knowledge graph module and a massive tool semantic vector retrieval knowledge base module. In the massive tool dependency relationship knowledge graph module, the graph nodes represent available tools for intelligent agents. The node attributes include tool name, tool function description, tool parameter name, tool parameter type, tool parameter meaning, and tool description semantic vector fields. The fusion retrieval of knowledge graph and vector retrieval is realized, and each node is directly recalled as a seed node for graph access through vector retrieval. Graph nodes are divided into independent tool nodes and auxiliary tool nodes. Independent tool nodes can directly or indirectly meet the user's instruction requirements, and auxiliary tool nodes serve independent tool nodes. Graph edges are divided into "tool direct dependency", "tool indirect dependency", "parameter direct dependency", and "parameter indirect dependency", which are used to represent the dependency relationship and hierarchical relationship of massive tools. The massive tool semantic vector retrieval knowledge base module is composed of a large number of retrievable semantic vectors encoded by semantic models. The encoded texts include the name field, function description field, and formal parameter description field of the tool library. These description information is injected into the knowledge base through the tool registration mechanism, and the recall of the graph retrieval seed node is realized in the retrieval stage.

7. The large language model agent mass tool invocation system based on knowledge graph retrieval enhancement according to claim 6, characterized in that: In the massive tool dependency relationship knowledge graph module, the tool description semantic vector field of the graph node attribute is obtained by encoding the tool-related text through a semantic model. This vector accurately reflects the semantic information of the tool, allowing for a combination of vector retrieval during knowledge graph retrieval to improve the accuracy and comprehensiveness of the retrieval and achieve comprehensive representation of tool information at the semantic and graph structure levels.

8. The large language model agent mass tool invocation system based on knowledge graph retrieval enhancement according to claim 7, characterized in that: In the mass tool dependency relationship knowledge graph module, the division mode of the independent tool node and the auxiliary tool node is: the independent tool node refers to a tool node whose execution result generally directly meets the user instruction demand; the auxiliary tool node refers to a tool node that generally serves the independent tool node; through the division, the true value tool meeting the user intention is clearly found in the tool retrieval process, and the efficiency and accuracy of tool calling are improved.

9. The large language model agent mass tool invocation system based on knowledge graph retrieval enhancement according to claim 8, characterized in that: In the mass tool dependency relationship knowledge graph module, the four types of edges are defined as follows: "Tool directly depends on" means that a tool must execute the dependent tool to execute normally; "Tool indirectly depends on" means that a tool can optionally execute the dependent tool, and the optional condition depends on the intention instruction input by the user; "Parameter directly depends on" means that the input parameter of a tool must be obtained by executing its dependent tool; "Parameter indirectly depends on" means that the input parameter of a tool can optionally be obtained by executing its dependent tool.

10. The large language model agent mass tool invocation system based on knowledge graph retrieval enhancement according to claim 9, characterized in that: In the mass tool semantic vector retrieval knowledge base module, the name field is the concise and easily distinguishable command name of the function tool, the function description field is the perfect and detailed function text introduction of the tool function, and the formal parameter description field is the complete description text of the type and meaning of all formal parameters of the tool function; by injecting these description information into the semantic vector retrieval knowledge base, the model fully understands the function of all called tools and the difference of the corresponding function and parameter slot filling semantics in the retrieval stage, fine-grained tool recognition and intention recognition are realized, and the accuracy and pertinence of the mass tool calling of the large language model agent are improved.

Images