Question and answer method, system and equipment based on large model and multiple Agents and medium

Abstract

The invention discloses a question and answer method based on a large model and multiple Agents, belongs to the technical field of large model artificial intelligence, and aims to realize deep understanding and intention recognition of user questions and accurately judge professional fields and processing requirements of the questions through an elaborately designed expert cue word and a large language model based on an intelligent intention recognition mechanism of the large model. Based on the intention recognition result, the most suitable professional Agent is automatically selected for processing; according to the method, a complete tool pool is constructed, two tool calling modes of Function Call and MCP are supported, a professional Agent can automatically select a proper tool according to question requirements, the professionality and accuracy of answering are remarkably improved, and the question processing efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large model artificial intelligence, and in particular to a large model and multi-agent based question and answer method, system, device and storage medium. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, intelligent customer service systems have been widely used in enterprise service fields. The existing intelligent customer service systems mainly adopt the following three modes: the first is a single large language model based intelligent question and answer system, which directly processes user problems through a large language model. Although it has certain natural language understanding ability, it lacks deep knowledge in professional fields and is difficult to provide accurate professional answers. The second is a rule engine based intelligent routing system, which routes problems through predefined rules. Although it has certain classification ability, it lacks flexibility and is difficult to handle complex natural language expressions and emerging business scenarios. The third is a vector retrieval based knowledge question and answer system, which improves the accuracy of knowledge retrieval through semantic vector retrieval, but lacks intelligent intent recognition and professional problem processing ability. SUMMARY

[0003] The purpose of the embodiments of the present application is to provide a large model and multi-agent based question and answer method, which aims to solve the problem of lack of intelligent intent recognition and accurate routing capability in the existing enterprise intelligent customer service system question and answer method, and the inability to automatically select the most suitable processing node according to the professional field of user problems, resulting in low user problem processing efficiency, low answer quality and insufficient professional depth.

[0004] The embodiments of the present application are implemented in the following way: a large model and multi-agent based question and answer method, the method comprising the following steps, precreating at least one domain agent; precreating and storing a domain knowledge database; precreating a tool pool for executing tasks; obtaining user input information; generating template instructions through a prompt word engine; performing intent recognition and routing analysis on the template instructions through a language large model, outputting decision information in a specific format; selecting a corresponding domain agent according to the decision information, and the domain agent calling tools according to the decision information, retrieving the domain knowledge database to generate a processing result and returning.

[0005] Further, in the embodiments of the present application, the decision information in a specific format includes domain category, processing node, tool and tool parameter.

[0006] Further, in the embodiment of the present application, the tool pool comprises a Function Call tool and an MCP service tool; the Function Call tool comprises data query, document generation, calculation and format conversion; and the MCP service tool comprises database connection, API calling, file operation and system monitoring.

[0007] Further, in the embodiment of the present application, the intent recognition and route analysis of the template instruction by the language large model comprises the following steps, performing intent recognition according to the template instruction to determine the core demand of the user; determining the domain category; allocating a processing node; determining the tool name and tool parameters when the tool needs to be called.

[0008] Further, in the embodiment of the present application, the domain Agent calls the tool according to the decision information, retrieves the domain knowledge database to generate a processing result and returns, comprising the following steps: determining whether the tool needs to be called, if the tool is needed, sending a tool calling request; if the tool is not needed, querying the corresponding domain knowledge database and returning an execution result; determining the tool type, if it is a Function Call tool, calling the corresponding tool in the Function Call tool in the tool pool, if it is an MCP service tool, calling the corresponding tool in the MCP service tool in the tool pool; returning the execution result according to the execution of the tool in combination with the corresponding domain knowledge database.

[0009] Another object of the embodiment of the present application is to provide a question and answer system based on a large model and multiple Agents, comprising: a pre-creation module for pre-creating at least one domain Agent, pre-creating and storing a domain knowledge database and pre-creating a tool pool for executing tasks; a user information acquisition module for acquiring input information of a user; a template instruction generation module for generating a template instruction through a prompt word engine; an intent recognition and route analysis module for performing intent recognition and route analysis of the template instruction by a language large model, a decision information output module for outputting decision information in a specific format; an Agent selection module for selecting a corresponding domain Agent according to the decision information; a domain Agent for calling a tool according to decision information, retrieving a database to generate a processing result and returning.

[0010] Further, in the embodiment of the present application, the intention recognition and routing analysis module comprises: a user intention recognition module, which performs intention recognition according to the template instruction to determine the core demand of the user; a domain category determination module, which is configured to determine the domain category; a processing node allocation module, which is configured to allocate a processing node; a tool calling determination module, which is configured to determine the tool name and tool parameters when a tool needs to be called.

[0011] Further, in the embodiment of the present application, the domain Agent comprises: a tool calling judgment processing module, which is configured to judge whether a tool needs to be called, and if a tool is needed, send a tool calling request; if a tool is not needed, query the corresponding domain knowledge database and return an execution result; a tool type judgment processing module, which is configured to judge whether a corresponding tool in the Function Call tool in the tool pool needs to be called if it is a Function Call tool, or whether a corresponding tool in the MCP service tool in the tool pool needs to be called if it is an MCP service tool, an execution result returning module, which is configured to return an execution result according to the execution of the tool in combination with the corresponding domain knowledge database.

[0012] Another object of the embodiment of the present application is to provide a computer device, which comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the large model and multi-Agent based question and answer method of any embodiment of the present application when executing the computer program.

[0013] Another object of the embodiment of the present application is to provide a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the large model and multi-Agent based question and answer method of any embodiment of the present application when executed by a processor.

[0014] Advantages of the present application The present application proposes a large model and multi-Agent based question and answer method, an intelligent intention recognition mechanism based on a large model, through carefully designed expert prompt words and a large language model, deep understanding and intention recognition of user problems are realized, the professional domain and processing demand of the problem are accurately judged, based on the intention recognition result, the most suitable professional Agent is automatically selected for processing; a complete tool pool is constructed, supporting two tool calling modes of Function Call and MCP, the professional Agent can automatically select the appropriate tool according to the problem demand, significantly improving the professionalism and accuracy of the answer, and improving the processing efficiency of the problem. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a large model and multi-agent based question and answer method flowchart of a preferred embodiment of the present application; Figure 2 is a large model and multi-agent based question and answer method flowchart of a preferred embodiment of the present application; Figure 3 is a large model and multi-agent based question and answer method flowchart of a preferred embodiment of the present application; Figure 4 is a large model and multi-agent based question and answer system structure diagram of a preferred embodiment of the present application; Figure 5 is an intent recognition and routing analysis module structure diagram of a preferred embodiment of the present application; Figure 6 is a large model and multi-agent based question and answer method flowchart of a preferred embodiment of the present application; DETAILED DESCRIPTION

[0016] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. For the convenience of description, only parts related to the embodiments of the present application are shown. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0017] The intelligent intent recognition mechanism based on large model of the embodiments of the present application realizes deep understanding and intent recognition of user problems through carefully designed expert prompt words and large language models, accurately judges the professional field and processing requirements of the problem, and automatically selects the most suitable professional Agent for processing based on the intent recognition result. A complete tool pool is constructed to support two tool calling modes of Function Call and MCP, and the professional Agent can automatically select the appropriate tool according to the problem requirements, significantly improving the professionalism and accuracy of the answer and improving the processing efficiency of the problem.

[0018] First, the meanings of several terms involved in the present application are analyzed: The large language model is an artificial intelligence model based on deep learning technology, which can generate natural and fluent text that meets grammar and semantic rules by learning and training a large amount of text data.

[0019] The tool pool is a centralized management of various functional tools, encapsulating query, calculation and other capabilities, providing standardized interfaces for Agents to call on demand to support efficient task execution.

[0020] Prompt word engine is a key tool connecting user needs and large language model (LLM), and the core function is to analyze user's fuzzy needs, supplement key information through preset templates or interactive guidance, and generate structured prompt words that can be accurately understood by LLM. It can manage the life cycle of prompt words, adapt to different model characteristics, reduce user dependence on prompt word engineering, significantly improve the accuracy and practicality of AI output, and is widely used in copywriting, data analysis and other scenarios.

[0021] Function Call (function call) is a bridge for large language models (LLM) to interact with external tools, systems or APIs, allowing the model to actively call preset functions (tools) when processing tasks, and converting natural language requirements into executable code logic, thereby expanding the capability boundary of AI (such as obtaining real-time data, operating software, performing calculations, etc.).

[0022] MCP service tool is an enterprise-level tool set relying on the management control platform (MCP), which solves the security and standardization problems of AI system interaction with enterprise underlying resources. It includes database connection, API call, file operation, system monitoring, etc., with permission verification and log tracing capabilities, and can encapsulate resource differences uniformly, so that upper-layer systems can complete data query, report generation and other operations through standardized interfaces.

[0023] Figure 1 A large model and multi-agent based question and answer method flowchart provided by an embodiment of the present application, the method comprising the following steps: Step S100, at least one domain agent is created in advance; a domain knowledge database is created and stored in advance; a tool pool for executing tasks is created in advance; Step S101, obtaining user input information; In the embodiment of the present application, multi-channel user input information is obtained, which can receive user input information from mobile application programs, web interactive interfaces, intelligent voice interactive devices, wearable smart terminals and other types of devices; In another embodiment of the present application, the user input information is encapsulated through API, and the user's input information is obtained through API gateway; Step S102, generating template instructions through a prompt word engine; Step S103, performing intent recognition and routing analysis on the template instructions through a language large model, Step S104, outputting decision information in a specific format; In the embodiment of the present application, the decision information in a specific format includes domain category, processing node, tool and tool parameter; in the embodiment of the present application, the specific format is JSON output format specification; Step S105, selecting a corresponding domain Agent according to the decision information, the domain Agent calling tools according to the decision information, retrieving a domain knowledge database to generate a processing result and returning.

[0024] In the embodiment of the present application, each domain Agent (intelligent agent) has professional knowledge and processing ability in a specific domain, each domain Agent integrates tool calling and query functions, and can call corresponding tools in a tool pool or query a corresponding domain knowledge database according to decision information.

[0025] In another embodiment, step S105 is replaced by selecting a corresponding processing node for processing according to the decision information; the processing node includes a domain Agent and a non-Agent type processing node (for example, the MCP service interface is a system function interface, and the manual service is a human processing channel). In the embodiment of the present application, the tool pool includes Function Call tools and MCP service tools; the Function Call tools include data query, document generation, calculation and format conversion, etc.; the MCP service tools include database connection, API calling, file operation and system monitoring, etc.

[0026] In the embodiment of the present application, the prompt word engine uses a preset prompt word template, including role definition, task description, available resources, tool pool and output format. In the embodiment of the present application, the role definition is a decision expert of an enterprise intelligent question and answer system. In the embodiment of the present application, the task description includes intent recognition, domain judgment, task allocation and tool selection; the intent recognition is used to determine the core demand and real intention of a user; the domain judgment is used to determine the professional domain to which a question belongs; the task allocation is used to route to the most suitable processing node; and the tool selection is used to determine the tool name and tool parameters when tools need to be called. In the embodiment of the present application, the available resources include domain classification, processing node and tool pool. In the embodiment of the present application, the domain classification includes customer service assistance, business knowledge, data query, tool operation and other domains; the customer service assistance domain further includes domain subcategories of product consultation, operation guide and account problem; the business knowledge domain further includes domain subcategories of futures rules, market policy and transaction mechanism; the data query domain further includes domain subcategories of index statistics, report generation and historical data; the tool operation domain further includes domain subcategories of report export, document generation and calculation execution; and the other domain further includes a subcategory of unclassified problems. In the embodiment of the present application, the processing node includes a customer service assistance Agent, a business knowledge Agent and a ChatBI Agent, which are all intelligent agent processing nodes. Customer service assistants are used to handle customer service issues such as product usage and account management; business knowledge agents are used to answer professional questions such as futures trading rules and compliance policies; ChatBIAgent is used to perform data queries and visual analysis; In another embodiment of the application, the processing node also includes non-agent processing nodes such as MCP service interface, manual service; The MCP service interface is used for reporting export, document generation and other operations; the manual service is used for complex problems or when the system cannot handle them; In an embodiment of the application, the tool pool includes function_call tools and MCP service tools; In an embodiment of the application, the function_call tools include data query, document generation, calculation, and format conversion tools; the data query tool is used to perform real-time data query and statistical analysis; the document generation tool is used to generate professional documents such as reports and contracts; the calculation tool is used to perform complex mathematical calculations and financial models; the format conversion tool is used to support multiple document format conversion; In an embodiment of the application, the MCP service tools include database connection, API call, file operation, and system monitoring; the database connection is used to connect enterprise internal database systems; the API call is used to call third-party service interfaces; the file operation is used to read and write file systems; the system monitoring is used to obtain system running status; In an embodiment of the application, the output format is JSON output format specification. In some embodiments, in addition to JSON format output, the output format can also be adjusted to YAML, XML, table (Markdown / CSV), key-value pair, and other structured formats; In an embodiment of the application, the prompt word engine template is implemented in Python, and in some embodiments, other programming languages such as JavaScript / TypeScript, Java, Go, and PHP can also be used for implementation, and the choice mainly depends on the system stack technology and development needs.

[0027] Figure 2 A method flowchart for intent recognition and routing analysis of the template instruction by a language large model is a preferred embodiment of the application; the method includes the following steps: S1031, according to the template instruction, intent recognition is performed to determine the core needs of the user; S1032, determine the domain category; S1033, assign a processing node; S1034, determine the tool name and tool parameters when the tool needs to be called; Exemplarily, taking the user question "how to export the futures trading report of the past 30 days" as an example, the intention recognition and routing analysis are as follows: Intention recognition: the language large model extracts the core requirement of the user ( "generate and export the futures trading report of the past 30 days" ) through the template instruction; domain judgment: according to the requirement type, it is classified into the "tool operation" domain ( the domain classification is the fourth type in the code) ; node allocation: matching the "MCP service interface" in the "processing node" ( the node is responsible for "report export and document generation" in the code) ; tool routing: synchronously selecting the "MCP service-file operation tool" in the "tool pool", and explicitly specifying the tool parameters ( "time range: the past 30 days, file format: Excel" ) ; output result: outputting the routing decision in the JSON format ( such as specifying agent: "MCP service interface", tools: [{"tool_type": "MCP", "tool_name": "file operation tool"}] ).

[0028] In step S1034 of some embodiments, if the tool does not need to be called, the tool name and the tool parameter are both empty.

[0029] In the embodiment of the application, the corresponding domain Agent is selected according to the decision information for processing, which is specifically: The domain Agent calls the corresponding tool in the tool pool according to the processing node, the tool name and the tool parameter in the decision information, and outputs the result in combination with the corresponding domain knowledge database; In some embodiments, if the tool name and the tool parameter in the decision information are both empty, the domain Agent directly outputs the result after querying the corresponding domain knowledge database; In another embodiment of the application, the corresponding processing node is selected according to the decision information for processing; when the processing node is a non-Agent type processing node ( without the ability to analyze problems or make decisions), it is only responsible for performing a preset operation ( such as exporting a report), Figure 3 It is a domain Agent processing method flow chart of a preferred embodiment of the application; In the embodiment of the application, the domain Agent calls the tool according to the decision information, retrieves the domain knowledge database to generate a processing result, and returns the following steps: S1051, judging whether the tool needs to be called, If the tool is needed, a tool calling request is sent; if the tool is not needed, the corresponding domain knowledge database is queried, and an execution result is returned; S1052, judging the tool type, if it is a Function Call tool, calling the corresponding tool in the Function Call tool in the tool pool, if it is an MCP service tool, calling the corresponding tool in the MCP service tool in the tool pool, S1053, returning an execution result according to the execution of the tool in combination with a corresponding domain knowledge database; The embodiment of the application is based on a large model intelligent intent recognition mechanism, and through carefully designed expert prompt words and a large language model, deep understanding and intent recognition of a user question are realized, and the professional field and processing demand of the question are accurately judged. A collaborative architecture of a master control Agent and multiple professional field Agents is designed, intelligent question routing and professional processing are realized. A complete tool pool is constructed, two tool calling modes of Function Call and MCP are supported, and appropriate tools can be automatically selected according to question demand, and the professionalism and accuracy of the answer are significantly improved.

[0030] The embodiment of the application is based on a large model intelligent intent recognition mechanism, and through carefully designed expert prompt words and a large language model, deep understanding and intent recognition of a user question are realized, and the professional field and processing demand of the question are accurately judged. A collaborative architecture of a master control Agent and multiple professional field Agents is designed, intelligent question routing and professional processing are realized. A complete tool pool is constructed, two tool calling modes of Function Call and MCP are supported, and appropriate tools can be automatically selected according to question demand, and the professionalism and accuracy of the answer are significantly improved.

[0031] Corresponding to the large model and multi-Agent based question and answer method described in the above embodiment, Figure 4 A structure diagram of a large model and multi-Agent based question and answer system provided by the embodiment of the application is shown, only parts related to the embodiment of the application are shown for ease of illustration. The system comprises A pre-creation module is used for pre-creating at least one domain Agent, pre-creating and storing a domain knowledge database, and pre-creating a tool pool for executing tasks; A user information acquisition module is used for acquiring input information of a user; A template instruction generation module is used for generating a template instruction through a prompt word engine; An intent recognition and routing analysis module is used for performing intent recognition and routing analysis on the template instruction through a language large model, A decision information output module is used for outputting decision information in a specific format; An Agent selection module is used for selecting a corresponding domain Agent according to the decision information; a domain Agent, configured to invoke a tool according to decision information, retrieve a database to generate a processing result, and return the processing result; Figure 5 is a system structure diagram of an intent recognition and routing analysis module according to a preferred embodiment of the present application; the intent recognition and routing analysis module comprises a user intent recognition module, configured to perform intent recognition according to the template instruction, and determine a core demand of the user; a domain category determination module, configured to determine a domain category to which the user belongs; a processing node allocation module, configured to allocate a processing node; a tool invocation determination module, configured to determine a tool name and tool parameters when a tool needs to be invoked; Figure 6 is a structure diagram of a domain Agent according to a preferred embodiment of the present application; the domain Agent comprises a tool invocation judgment processing module, configured to judge whether a tool needs to be invoked, and if a tool is needed, send a tool invocation request; if a tool is not needed, query a corresponding domain knowledge database, and return an execution result; a tool type judgment processing module, configured to judge whether a corresponding tool in a Function Call tool in a tool pool needs to be invoked if the tool is a Function Call tool, or whether a corresponding tool in an MCP service tool in the tool pool needs to be invoked if the tool is an MCP service tool, an execution result return module, configured to return an execution result according to tool execution in combination with a corresponding domain knowledge database; An embodiment of the present application creates a plurality of professional domain Agents and professional knowledge databases through a pre-creation module based on a large model and a multi-Agent question and answer system; a complete tool pool is constructed, supporting two tool invocation modes of Function Call and MCP, deep understanding and intent recognition of a user question are realized through a template instruction generation module, an intent recognition and routing analysis module, accurate judgment of a professional domain and processing demand of the question is realized, the most suitable domain Agent is selected through a decision information output module and an Agent selection module, the domain Agent can automatically select a suitable tool according to a question demand, the professional nature and accuracy of an answer are significantly improved, and the processing efficiency of the question is improved.

[0032] The embodiment of the present application also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the question and answer methods based on a large model and a multi-Agent according to the embodiment of the present application when executing the computer program.

[0033] The embodiment of the application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement steps of any large model and multi-agent based question and answer method of the embodiment of the application.

[0034] The embodiment of the application further provides a computer program product, which, when running on a computer device, enables the computer device to perform steps of any large model and multi-agent based question and answer method of the embodiment of the application.

[0035] Those skilled in the art can understand that all or part of the steps of the above-mentioned embodiment methods can be completed by program instruction related hardware, and the program can be stored in a computer readable storage medium, such as ROM, RAM, magnetic disk, optical disc, etc.

[0036] The above merely describes the preferred embodiments of the present application and is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A large model and multi-agent based question and answer method, characterized in that, The method Includes the following steps, At least one domain agent is created in advance; Pre-create and store a domain knowledge database; pre-create a pool of tools to perform tasks; Obtain user input information; Generate template instructions using a prompt word engine; The template instructions are subjected to intent recognition and routing analysis using a large language model. Output decision information in a specific format; Based on the decision information, the appropriate domain agent is selected. The domain agent then invokes tools based on the decision information, retrieves the domain knowledge database, generates processing results, and returns them.

2. The question-answering method based on a large model and multiple agents as described in claim 1, characterized in that, The decision information in the specific format includes domain category, processing node, tool, and tool parameters.

3. The large model and multi-agent based question answering method of claim 2, wherein, The tool pool includes Function Call tools and MCP service tools; Function Call tools include data query, document generation, calculation, and format conversion; MCP service tools include database connection, API call, file operation, and system monitoring.

4. The question-answering method based on a large model and multiple agents as described in claim 3, characterized in that, The process of performing intent recognition and routing analysis on the template instructions using a large language model includes the following steps: Intent recognition is performed based on the template instructions to determine the user's core needs; Determine the relevant field category; Assign processing nodes; Determine the tool name and parameters when you need to invoke the tool.

5. The question-answering method based on a large model and multiple agents as described in claim 4, characterized in that, The domain agent invokes tools based on decision information, retrieves domain knowledge databases to generate processing results, and returns the results including the following steps: Determine whether it is necessary to call the tool. If tools are needed, a tool invocation request is sent; if tools are not needed, the corresponding domain knowledge database is queried and the execution result is returned. Determine the tool type. If it is a Function Call tool, call the corresponding tool in the Function Call tool pool. If it is an MCP service tool, call the corresponding tool in the MCP service tool pool. Based on the execution of the tool and the relevant domain knowledge database, the execution result is returned.

6. A question-answering system based on a large model and multiple agents, characterized in that, The system includes: The pre-creation module is used to pre-create at least one domain agent; pre-create and store a domain knowledge database; and pre-create a pool of tools for performing tasks. The user information acquisition module is used to acquire user input information; The template instruction generation module is used to generate template instructions through the prompt word engine; The intent recognition and routing analysis module is used to perform intent recognition and routing analysis on the template instructions using a large language model. The decision information output module is used to output decision information in a specific format; The Agent selection module is used to select the appropriate domain Agent based on the decision information. Domain Agent is used to invoke tools based on decision information, retrieve databases, generate processing results, and return them.

7. The question-answering system based on a large model and multiple agents as described in claim 6, characterized in that, The intent recognition and routing analysis module includes: The user intent recognition module performs intent recognition based on the template instructions to determine the user's core needs; The domain category determination module is used to determine the domain category to which the domain belongs. The processing node allocation module is used to allocate processing nodes; The tool call determination module is used to determine the tool name and tool parameters when a tool needs to be called.

8. The question-answering system based on a large model and multiple agents as described in claim 7, characterized in that, The domain agent includes: The tool call judgment and processing module is used to determine whether a tool needs to be called. If a tool is needed, a tool call request is sent; if a tool is not needed, the corresponding domain knowledge database is queried and the execution result is returned. The tool type determination module determines whether the tool is a Function Call tool or an MCP service tool. If so, it calls the corresponding tool from the Function Call tool pool; if so, it calls the corresponding tool from the MCP service tool pool. The execution result return module is used to return the execution result based on the tool's execution and the corresponding domain knowledge database.

9. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the question-answering method based on a large model and multiple agents as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the question-answering method based on a large model and multiple agents as described in any one of claims 1 to 5.

Images