Multi-time reasoning system, method and equipment based on large model and medium

Abstract

The invention discloses a multi-time reasoning system and method based on a large model, equipment and a medium. In the application, the selector focuses on the selection of the intelligent agent, does not perform task disassembly, and encapsulates the implementation logic of the task in the specific intelligent agent. Each agent can focus on the specific task field which the agent is good at, optimization and adjustment are carried out according to the characteristics of the field, and the task processing efficiency and quality are improved. According to the technical scheme, the selector redescribes the current user problem by combining the selected intelligent agent and the interaction information recorded by the memory intelligent agent instead of extracting API parameters, so that information loss or misunderstanding caused by simple API parameter extraction is avoided, and the context of historical interaction and the functional characteristics of the intelligent agent are considered. Rich interaction information stored in the memory provides precious decision basis for the selector. According to the system, through a combination mode of multi-round selection and task internal reasoning, the processing difficulty of complex problems in a special field is relieved.

Description

Technical Field

[0001] This application relates to the technical fields of natural language processing and deep learning, and particularly to a multi-inference system, method, device, and medium based on a large model. Background Art

[0002] In the current era of the rapid development of large model applications, complex problem reasoning has become a key factor restricting the application effect and scope of large models. Although frameworks such as Workflow (workflow framework), ReAct (reactive action framework), and Function Call (function call framework) have shown certain value in engineering applications, when dealing with complex problems, they each have significant limitations, hindering the effective implementation of large model intelligence.

[0003] The Workflow framework, with its fixed processing flow, is suitable for clear and single-task application scenarios. However, in the face of complex and changing reasoning requirements, its lack of self-planning ability makes Workflow ineffective when dealing with unstructured or dynamically changing tasks. The fixity of this framework limits its flexibility in adapting to complex problems.

[0004] The ReAct framework attempts to achieve autonomous reasoning through cycles of planning, execution, observation, and stopping conditions. Although this method has the ability of autonomous planning in theory, in practical applications, the over-reliance of large models on preset prompt words and the lack of sufficient professional background knowledge often lead to the unreliability of reasoning results. In addition, the fully autonomous planning method of the ReAct framework also increases the uncertainty and risk of the system.

[0005] Therefore, how to achieve effective and accurate reasoning for complex problems in a professional field remains an urgent technical problem to be solved. Summary of the Invention

[0006] This application provides a multi-inference system, method, device, and medium based on a large model, which is used to solve the problems of poor flexibility, high uncertainty, and insufficient processing ability existing in the process of dealing with complex problems.

[0007] In a first aspect, this application provides a multi-inference system based on a large model, and the system includes: a memory, a selector with the ability to select multi-round intelligent agents, and multiple intelligent agents;

[0008] The memory is used to store the interaction information between the selector and the multiple agents, so as to provide historical data support for the selector in any round of agent selection; wherein, the interaction information includes: each historical user question, the question answers respectively corresponding to the historical user questions, and the agent information respectively selected for the historical user questions;

[0009] The selector is used to receive a user question and solve the user question based on at least one round of agent selection process; wherein, in the first round of selection, based on the interaction information stored in the memory, one or more target agents for processing the user question are determined from the multiple agents; for each determined target agent, based on the user question and the pre - saved functional scope corresponding to the target agent, a restatement question is generated and input to the target agent; and, the response information fed back by any one of the target agents is received, and based on the interaction information stored in the memory, it is determined whether the response information meets the pre - configured output requirements; if it is determined that the response information does not meet the output requirements, the problem is analyzed again based on the interaction information stored in the memory to generate a new intermediate processing question, and a corresponding target agent is re - allocated for the intermediate processing question. Based on the functional scope of the re - allocated target agent and the intermediate processing question, a new restatement question is generated and output to the re - allocated target agent for processing, and so on, continuously performing multiple rounds of agent selection and problem processing until it is determined that the received response information meets the output requirements and the response information is output as the question answer;

[0010] The multiple agents are used to receive the restatement question input by the selector; determine each subtask corresponding to the restatement question and the execution relationship between the subtasks; execute each subtask according to the execution relationship between the subtasks; determine the response information corresponding to the restatement question based on the execution results respectively corresponding to the subtasks; and send the response information to the selector to support the selector for possible next - round agent selection.

[0011] In a second aspect, the present application provides a multiple - inference method based on a large model, and the method includes:

[0012] The selector receives a user question; based on the interaction information stored in the memory, one or more target agents for processing the user question are determined from the multiple agents; for each determined target agent, based on the user question and the pre - saved functional scope corresponding to the target agent, a restatement question is generated and input to the target agent;

[0013] For any of the target agents, receive the retelling question input by the selector; determine each subtask corresponding to the retelling question and the execution relationship between the subtasks; execute each subtask according to the execution relationship between the subtasks; based on the execution results respectively corresponding to each subtask, determine the response information corresponding to the retelling question; send the response information to the selector to support the selector to perform the next possible agent selection;

[0014] The selector receives the response information fed back by any of the target agents, and determines whether the response information meets the pre-configured output requirements based on each interaction information stored in the memory; if it is determined that the response information does not meet the output requirements, then analyze the problem again based on each interaction information stored in the memory to generate a new intermediate processing problem, and re-assign a corresponding target agent to the intermediate processing problem, and generate and output a new retelling question to the re-assigned target agent for processing according to the function scope of the re-assigned target agent and the intermediate processing problem, and so on in a loop, continuously performing multiple rounds of agent selection and problem processing until it is determined that the received response information meets the output requirements and output the response information as the problem answer.

[0015] In a third aspect, the present application provides a computer device, which includes a processor, and the processor is used to implement the steps of the above-mentioned multiple inference method based on a large model when executing a computer program stored in a memory.

[0016] In a fourth aspect, the present application provides a computer-readable storage medium, which stores a computer program, and the computer program is used to implement the steps of the above-mentioned multiple inference method based on a large model when executed by a processor.

[0017] The beneficial effects of the present application are as follows:

[0018] 1. Since the selector focuses on the selection of agents without performing task decomposition, this design has significant advantages. It encapsulates the implementation logic of tasks in specific agents, making the system architecture clearer and more modular. Each agent can focus on a specific task area it is good at, optimize and adjust according to the characteristics of this area, and improve the efficiency and quality of task processing. On this basis, the maintainability and scalability of the system are greatly enhanced. When new task types need to be added, only need to develop the corresponding agents and register them in the system.

[0019] 2. The selector uses the method of problem restatement instead of extracting API parameters, avoiding information loss or misunderstanding that may be caused by simple API parameter extraction. By combining the selected agent and the interaction information recorded by the memory agent, the selector can re-describe the current user problem. This restatement process takes into account the context of historical interactions and the functional characteristics of the agent, enabling the problem to be presented in a way that is easier for the agent to understand and process.

[0020] 3. With the support of the memory, the selector has the ability to select agents multiple times. The rich interaction information stored in the memory provides valuable decision-making basis for the selector. When dealing with complex problems, it is often impossible to obtain satisfactory results through a single agent selection. Multiple agents may need to work together to gradually solve the problem in stages. This mechanism of selecting agents multiple times enables the system to handle complex and changing problems, improving the success rate and accuracy of problem-solving. Through a combination of "multiple-round selection + internal task reasoning", the difficulty of handling complex problems in a specific domain is alleviated.

[0021] 4. Since the selector does not perform task decomposition, the agent needs to determine by itself whether the current request needs to be decomposed and how to decompose it. The agent can conduct a detailed analysis of the restated problem based on its own task logic and the domain knowledge it has, so as to decompose it into subtasks. These subtasks can be arranged according to the corresponding execution relationships to ensure that the task can be completed efficiently and accurately. The agent's ability to decompose tasks in this way enables the system to handle various complex problems more flexibly and give full play to the professional advantages of each agent. Brief Description of the Drawings

[0022] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0023] Figure 1 It is a schematic structural diagram of a multiple-inference system based on a large model provided by an embodiment of the present application;

[0024] Figure 2 It is a schematic process diagram of multiple inferences based on a large model provided by an embodiment of the present application;

[0025] Figure 3 It is a schematic structural diagram of a computer device provided by an optional embodiment of the present application. Detailed Embodiments

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following will further describe this application in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of them. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of this application.

[0027] To achieve effective and accurate reasoning for complex problems in a professional field, this application provides a multi-inference system, method, device, and medium based on a large model.

[0028] Embodiment 1:

[0029] This application provides a multi-inference system based on a large model. Figure 1 FIG. 11 is a schematic structural diagram of a multi-inference system based on a large model provided by an embodiment of this application. The system includes: a memory 11, a selector 12 with the ability to select multi-round agents, and multiple agents 13.

[0030] The memory 11 is used to store the interaction information between the selector 12 and the multiple agents 13 to provide historical data support for the selector 12 in any round of agent selection. Among them, the interaction information includes: each historical user question, the question answers respectively corresponding to the historical user questions, and the agent information selected for each historical user question.

[0031] The selector 12 is used to receive a user question and solve the user question based on at least one round of agent selection process. Among them, in the first round of selection, based on the interaction information stored in the memory 11, one or more target agents for processing the user question are determined from the multiple agents 13. For each determined target agent, a retelling question is generated and input to the target agent based on the user question and the pre-saved functional scope corresponding to the target agent. And, receive the response information fed back by any one of the target agents, and determine whether the response information meets the pre-configured output requirements based on the interaction information stored in the memory 11. If it is determined that the response information does not meet the output requirements, the problem is analyzed again based on the interaction information stored in the memory 11 to generate a new intermediate processing question, and a corresponding target agent is re-allocated for the intermediate processing question. Based on the functional scope of the re-allocated target agent and the intermediate processing question, a new retelling question is generated and output to the re-allocated target agent for processing. This process is repeated continuously, and multiple rounds of agent selection and problem processing are carried out until it is determined that the received response information meets the output requirements and the response information is output as the problem answer.

[0032] The multiple agents 13 are configured to receive the restated question input by the selector 12; determine each subtask corresponding to the restated question and the execution relationship between the subtasks; execute each subtask according to the execution relationship between the subtasks; determine the response information corresponding to the restated question based on the execution results respectively corresponding to each subtask; and send the response information to the selector 12 to support the selector 12 in performing the next possible agent selection.

[0033] This application provides a multi-inference system based on a large model. The system mainly consists of a memory 11, a selector 12 with the ability of multi-round agent selection, multiple agents 13, and an interaction module. The core goal of the system is to solve complex problems raised by users through multi-round agent selection and reasoning, and provide effective and accurate answers for users. The following is a detailed introduction to each module included in the multi-inference system based on the large model:

[0034] (I) Memory 11

[0035] The memory 11 serves as the data storage center of the system and plays a crucial role. It stores the interaction information between the selector 12 and the multiple agents 13, and this interaction information can provide historical data support for each round of agent selection by the selector 12. Among them, the stored interaction information includes: each historical user question, the question answers respectively corresponding to each historical user question, and the agent information selected when processing each historical user question. Optionally, the interaction information may also include the user intent respectively corresponding to each historical user question.

[0036] Among them, an agent is a large model trained based on specific professional field tasks and can independently handle related problems within the field.

[0037] Compared with the existing framework, the existence of the memory 11 enables the system to learn from historical experience. For example, the Workflow framework lacks effective utilization of historical interactions, while the memory 11 of this system allows the selector 12 to refer to the processing methods of past similar questions, thereby more accurately selecting agents and improving the efficiency and accuracy of problem-solving.

[0038] In a possible implementation manner, the memory 11 will also update the locally stored interaction information in a timely manner according to the data written by the selector 12. Through real-time update, the memory 11 can timely reflect the latest situation of the interaction between the system and the user, ensuring that the system always makes decisions based on the most accurate and comprehensive data.

[0039] (II) Selector 12

[0040] Selector 12 is the decision-making core of the system and has the ability to select agents in multiple rounds. After receiving a user's question, it will carefully determine one or more target agents in the first round based on the interaction information in the memory 11. Among them, the process of selecting agents is flexible and diverse. It can be based on the analysis of the user's intention or factors such as the matching degree (e.g., text similarity) between the current user question and the historical user questions. Taking the following two ways of selecting agents as examples, the process of selecting target agents will be described:

[0041] Method 1: If the target agent is selected based on the matching degree between the current user question and the historical user questions, selector 12 can first extract the features of the current user question. These features can include keywords, the theme of the question, the fields involved, etc. Taking keywords as an example, if the current user question is "Query the financial statements of a certain company in the past month", selector 12 will extract keywords such as "a certain company", "the past month", and "financial statements". Then, selector 12 will compare the extracted features with the features of each historical user question stored in the memory 11. For each historical user question, calculate its feature matching degree with the current user question. There can be multiple ways to calculate the matching degree. For example, a simple count of the number of matching keywords can be used. If the historical user question contains more keywords that are the same as the current user question, it is considered to have a higher matching degree. Or a more complex semantic matching algorithm can be used, considering the semantic relevance of keywords and the overall semantic similarity of the questions. Suppose there is a historical user question in the memory 11 "Query the financial statements of another company last month". Through keyword matching, it is found that the keyword "financial statements" is the same, and "the past month" and "last month" are semantically similar. Thus, it can be judged that this historical user question has a certain matching degree with the current user question. According to the high or low matching degree, selector 12 will screen out the historical user questions with higher matching degrees. Then, view the agent information selected by these historical user questions respectively. If multiple historical user questions have selected the same agent, then this agent is very likely to be determined as one of the target agents for processing the current user question. For example, in the above example, if multiple historical user questions related to "querying financial statements" have all selected the database query agent, then selector 12 may determine the database query agent as the target agent in the first round because, from historical experience, this agent is effective in processing similar questions.

[0042] Method 2: If the interaction information recorded in the memory 11 also includes the user intentions corresponding to each historical user question, then the target agent can be selected by analyzing the user intention of the current user question. Exemplarily, the selector 12 is specifically used for:

[0043] If the interaction information further includes the user intents corresponding to the respective historical user questions, then perform intent analysis on the user question to determine the target intent; determine, from the interaction information stored in the memory 11, the target interaction information associated with the target intent; and determine, according to the agent information recorded in the target interaction information, the target agent for processing the user question.

[0044] If a target agent is selected based on user intent analysis, the selector 12 performs intent analysis on the current user question. Natural language processing techniques, such as text classification and sentiment analysis, can be used to determine the target intent of the current user question. For example, by analyzing the question "Query the financial statements of a certain company in the past month", its target intent is determined to be obtaining the financial data of a specific company for a specific period. Then, from the interaction information stored in the memory 11, find the target interaction information associated with this target intent, and determine the target agent for processing the current user question according to the agent information recorded in the target interaction information. This method combining user intent analysis and the matching degree of historical questions enables the selector 12 to more accurately determine the target agent, improving the efficiency and accuracy of the system in solving problems. Next, find the target interaction information associated with the target intent from the memory 11, and determine the target agent for processing the user question according to the agent information recorded therein. If the user intent is not included, the selector 12 determines the target agent based on other factors such as the matching degree between the historical user question and the current user question.

[0045] After the selector 12 determines each target agent for processing the current user question based on the above embodiments, for each determined target agent, a restated question can be generated by combining the user question and the pre-saved functional scope corresponding to this target agent and input into this target agent. For example, if the target agent is a database query agent and the user question is "Find the sales data of Company A", the selector 12 may restate the question as "Find the sales data of which the company is A". Another example, for a weather query agent, the user question is "Query the weather tomorrow". Since weather queries usually require specifying a location, if the city where the user is located is recorded as "Beijing" in the memory 11, the selector 12 will restate the question as "Query the weather condition in Beijing tomorrow".

[0046] Generating a restated question can improve the efficiency and accuracy of the agent in processing questions. By transforming the user question into a form that is easy for the agent to understand and process, the understanding cost of the agent is reduced, enabling it to locate the problem faster and take corresponding processing measures. At the same time, the clear restated question also helps to avoid incorrect processing caused by unclear information, thus enhancing the performance and reliability of the entire system.

[0047] Based on the above embodiments, the target intelligent agent can process the restated question output by the selector 12, and thus feedback response information to the selector 12. When the selector 12 receives the response information feedback by any target intelligent agent, it can determine whether the response information meets the pre-configured output requirements based on the interaction information stored in the memory 11. Among them, the output requirements can be that the response information needs to match the question answers of similar historical user questions in the memory 11, or the response information needs to completely cover the key information involved in the user question. If the selector 12 determines that the response information does not meet the output requirements, it will analyze the question again based on the interaction information stored in the memory 11 to generate a new intermediate processing question. For example, if it is analyzed that the response information lacks key content, such as the user asks about the price, performance, and after-sales policy of a certain electronic product, and the intelligent agent feedback only includes the price, then the selector 12 can generate an intermediate processing question to supplement the two pieces of information of "performance" and "after-sales policy". If it is analyzed that there is an error in the response information, such as the provided price is too different from the actual market price, the selector 12 can generate an intermediate processing question about checking the provided price. If it is analyzed that the response information is ambiguous, for example, only mentions that a certain product has good performance but does not specifically describe the performance indicators, the selector 12 can generate an intermediate processing question to supplement the specific information of the performance indicators. The selector 12 can reassign the corresponding target intelligent agent to the intermediate processing question according to the characteristics of the intermediate processing question. The selector 12 combines the function scope of the reassigned target intelligent agent and the intermediate processing question, generates and outputs a new restated question to the reassigned target intelligent agent for processing. For example, if the pre-configured output requirements are not met because the answer is incomplete, the selector 12 may generate a new intermediate processing question, such as "Supplement the inventory quantity information of a certain product", and then select a suitable intelligent agent (such as an inventory management intelligent agent) to process this new question. This process repeats continuously, and multiple rounds of intelligent agent selection and question processing are carried out until it is determined that the received response information meets the output requirements and the response information is output as the question answer of the current user question.

[0048] In a possible implementation, when determining whether the response information meets the pre-configured output requirements, the associated interaction information corresponding to the user's question can be obtained from the memory 11. Among them, the historical user questions in the associated interaction information match the current user question. Here, the match can refer to the text match between the current user question and the historical user question, or the user intention match between the current user question and the historical user question, which is not specifically limited here. After finding the associated interaction information, the selector 12 can compare the response information fed back by the target agent with the question answer in the associated interaction information. If the question answer in the associated interaction information matches the current response information, it is determined that the response information meets the output requirements; otherwise, it is determined that it does not. For example, if the comparison result shows that the two are consistent in key entities and formats, such as price entities, product specification entities, etc., it is considered that the response information meets the output requirements; otherwise, it does not.

[0049] In this application, the selector 12 focuses on the selection of agents without performing task decomposition. This design has significant advantages. It encapsulates the implementation logic of the task in specific agents, making the system architecture clearer and more modular. Each agent can focus on a specific task area it is good at, optimize and adjust according to the characteristics of this area, and improve the efficiency and quality of task processing. For example, in a complex system involving multiple fields such as finance, healthcare, and education, the financial agent can specifically handle the analysis and prediction of financial data, the healthcare agent focuses on disease diagnosis and treatment plan recommendation, and the education agent is responsible for matching learning resources and formulating learning plans. In this way, the maintainability and scalability of the system are greatly enhanced. When a new task type needs to be added, only the corresponding agent needs to be developed and registered in the system.

[0050] And the selector 12 adopts the method of question paraphrasing instead of API parameter extraction, which is a more flexible and practical approach. By combining the selected agent and the interaction information recorded by the memory agent, the selector 12 can re-describe the current user question. This paraphrasing process takes into account the context of historical interactions and the functional characteristics of the agent, enabling the question to be presented in a way that is easier for the agent to understand and process. For example, when the user asks the question "Query the financial status of a certain company", after selecting the financial data query agent, the selector 12 will refer to the usage record of this agent in the memory 11 and the processing methods of previous similar questions, and paraphrase the question as "In the specified financial database, query the balance sheet, income statement, and cash flow statement data of a certain company for the most recent fiscal year". This paraphrasing not only clarifies the data source and specific content of the query but also combines the actual business scenario, avoiding information loss or misunderstanding that may be caused by simple API parameter extraction.

[0051] In addition, with the support of the memory 11, the selector 12 has the ability to select agents multiple times. The rich interaction information stored in the memory 11 provides valuable decision-making basis for the selector 12. When dealing with complex problems, it is often impossible to obtain satisfactory results through a single agent selection. Multiple agents 13 may need to work together to gradually solve the problem in stages. For example, when dealing with a problem involving market research and product recommendation, first select the market data collection agent to obtain relevant market data. Based on the feedback results, the selector 12 may judge that further data analysis is needed, so it reselects the data analysis agent for in-depth mining. If the data analysis results show that product recommendation needs to be combined with user preferences, the selector 12 will select the product recommendation agent. This process will continue until the selector 12 determines based on the information in the memory 11 that there is no need to continue calling agents, and then returns the currently obtained result to the user. This mechanism of selecting agents multiple times enables the system to handle complex and changing problems, improving the success rate and accuracy of problem-solving.

[0052] (III) Multiple agents 13

[0053] Multiple agents 13 are the "executors" of the system, each responsible for specific professional tasks in their respective fields, and all are large models trained based on their respective professional tasks. After receiving the restated problem input by the selector 12, they will decompose the problem to determine each subtask and the execution relationship between the subtasks. This way of problem decomposition can break down complex problems into multiple simple subtasks, facilitating more efficient processing. Specifically, for any agent, after receiving the restated problem input by the selector 12, it can determine each subtask corresponding to the restated problem and the execution relationship between these subtasks. Among them, the execution relationship can include one or more of serial, parallel, multiple-choice, etc. For example, for a complex data analysis problem, it may be decomposed into subtasks such as data collection, data cleaning, and data analysis, and there are execution relationships such as serial, parallel, and multiple-choice among these subtasks. According to the execution relationship between each subtask, each subtask is executed in turn. Then, the execution results of each subtask are integrated to determine the response information corresponding to the restated problem. This response information is then fed back to the selector 12 to support the selector 12 in making the next possible agent selection.

[0054] In a possible implementation, when the agent sequentially executes each subtask according to the execution relationship between the subtasks, it can combine a pre-constructed knowledge base that contains various knowledge and rules required to complete the task, ensuring that the agent can make accurate decisions and operations. For example, the knowledge base of a database query agent may include database table structures, query statement rules, etc. Another example is that a weather query agent can query and process weather information according to the rules for obtaining and analyzing weather data.

[0055] Since the selector 12 does not perform task decomposition, the agent needs to independently determine whether the current request needs to be decomposed and how to decompose it. The agent can conduct a detailed analysis of the restated problem based on its own task logic and the domain knowledge it has mastered. Taking the database query agent as an example, if the restated problem received is "Query the sales data of all stores of a large chain supermarket in the past month and rank them according to sales volume", the agent can decompose this problem into several subtasks: First, filter out the sales records of all stores of the large chain supermarket in the database; then, filter out the sales data in the past month; next, calculate the sales volume of each store; finally, rank the stores according to sales volume. These subtasks are executed according to the corresponding execution relationship to ensure that the task can be completed efficiently and accurately. This task decomposition ability of the agent enables the system to handle various complex problems more flexibly and give full play to the professional advantages of each agent.

[0056] Optionally, the system may further include an interaction module. As the interface between the system and the user, the interaction module serves as a bridge for communication between the system and the user, responsible for receiving the questions input by the user and feeding back the final answer of the system to the user.

[0057] In a possible implementation, the interaction module is also responsible for receiving the feedback information of the user on the question answer and passing it to the memory 11, so that the memory 11 can adjust the stored interaction information according to the user's feedback. This feedback mechanism enables the system to continuously optimize its own performance based on the user's feedback and improve user satisfaction.

[0058] Compared with the existing framework, the interaction module enhances the user experience and optimizability of the system. The existing framework often lacks an effective interaction and feedback mechanism with the user, while this system can timely understand the needs and opinions of the user through the interaction module and make targeted adjustments and improvements to the system.

[0059] The beneficial effects of this application are as follows:

[0060] 1. Since the selector 12 focuses on the selection of agents without task decomposition, this design has significant advantages. It encapsulates the implementation logic of tasks in specific agents, making the system architecture clearer and more modular. Each agent can focus on a specific task area it is good at, optimize and adjust according to the characteristics of this area, and improve the efficiency and quality of task processing. On this basis, the maintainability and scalability of the system are greatly enhanced. When new task types need to be added, only the corresponding agents need to be developed and registered in the system.

[0061] 2. The selector 12 uses the method of question restatement instead of API parameter extraction, avoiding information loss or misunderstanding that may be caused by simple API parameter extraction. By combining the selected agent and the interaction information recorded by the memory agent, the selector 12 can re-describe the current user question. This restatement process takes into account the context of historical interactions and the functional characteristics of the agent, making the question presented in a way that is easier for the agent to understand and process.

[0062] 3. With the support of the memory 11, the selector 12 has the ability to select agents multiple times. The rich interaction information stored in the memory 11 provides valuable decision-making basis for the selector 12. When dealing with complex problems, it is often impossible to obtain a satisfactory result through a single agent selection. Multiple agents 13 may need to work together and solve the problem step by step in stages. This mechanism of selecting agents multiple times enables the system to handle complex and changeable problems, improving the success rate and accuracy of problem-solving. Through a combination of "multiple-round selection + internal task reasoning", it alleviates the difficulty of dealing with complex problems in a specific domain.

[0063] 4. Since the selector 12 does not perform task decomposition, the agent needs to judge by itself whether the current request needs to be decomposed and how to decompose it. The agent can conduct a detailed analysis of the restated question based on its own task logic and the domain knowledge it has mastered, so as to decompose sub-tasks. These sub-tasks can be executed according to the corresponding execution relationships according to the actual situation to ensure that the task can be completed efficiently and accurately. This task decomposition ability of the agent enables the system to handle various complex problems more flexibly and give full play to the professional advantages of each agent.

[0064] Embodiment 2:

[0065] Based on the same inventive concept, the present application also provides a multiple inference method based on a large model. Figure 2 FIG. is a schematic diagram of the process of multiple inference based on a large model provided by an embodiment of the present application. The process includes:

[0066] S201: The selector receives the user's question; based on each interaction information stored in the memory, one or more target agents for processing the user's question are determined from the multiple agents; for each determined target agent, a restated question is generated and input to the target agent based on the user's question and the pre - saved functional scope corresponding to the target agent.

[0067] S202: For any one of the target agents, the restated question input by the selector is received; each subtask corresponding to the restated question and the execution relationship between the subtasks are determined; the subtasks are executed according to the execution relationship between the subtasks; based on the execution results corresponding to the subtasks respectively, the response information corresponding to the restated question is determined; the response information is sent to the selector to support the selector for the next possible agent selection.

[0068] S203: The selector receives the response information fed back by any one of the target agents, and based on each interaction information stored in the memory, determines whether the response information meets the pre - configured output requirements; if it is determined that the response information does not meet the output requirements, the problem is analyzed again based on each interaction information stored in the memory to generate a new intermediate processing problem, and corresponding target agents are re - allocated for the intermediate processing problem. Based on the functional scope of the re - allocated target agent and the intermediate processing problem, a new restated question is generated and output to the re - allocated target agent for processing. This process is repeated continuously, and multiple rounds of agent selection and problem processing are carried out until it is determined that the received response information meets the output requirements and the response information is output as the answer to the question.

[0069] The multi - inference method based on the large model provided by this application is applied to a computer device, which can be an intelligent device, such as a mobile terminal, a computer, etc., or a server, such as a business server, an application server, etc.

[0070] Since the principle of the above - mentioned multi - inference method based on the large model for solving problems is similar to that of the multi - inference system based on the large model, for specific details, please refer to the embodiments of the above - mentioned system, and the repeated parts will not be elaborated here.

[0071] Embodiment 3:

[0072] Please refer to Figure 3 , Figure 3 which is a schematic structural diagram of a computer device provided by an optional embodiment of this application, as Figure 3As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including a high-speed interface and a low-speed interface. Each component communicates with each other using different buses and can be installed on a common motherboard or in other ways as needed. The processor can process instructions executed within the computer device, including instructions stored in the memory or on the memory to display graphical information of the GUI on an external input / output device (such as a display device coupled to the interface). In some alternative embodiments, if needed, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides part of the necessary operations (such as an array of servers, a set of blade servers, or a multi-processor system). Figure 3 In the figure, a processor 10 is taken as an example.

[0073] The processor 10 can be a central processing unit, a network processor, or a combination thereof. Among them, the processor 10 can further include a hardware chip. The above-mentioned hardware chip can be an application-specific integrated circuit, a programmable logic device, or a combination thereof. The above-mentioned programmable logic device can be a complex programmable logic device, a field-programmable gate array, a generic array logic, or any combination thereof.

[0074] Among them, the memory 20 stores instructions executable by at least one processor 10, so that at least one processor 10 executes the method shown in the above embodiments.

[0075] The memory 20 can include a program storage area and a data storage area. Among them, the program storage area can store an operating system and application programs required for at least one function; the data storage area can store data created according to the use of the computer device presented by a kind of landing page of a small program, etc. In addition, the memory 20 can include a high-speed random access memory and can also include a non-transitory memory, such as at least one disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some alternative embodiments, the memory 20 can optionally include a memory remotely set relative to the processor 10, and these remote memories can be connected to the computer device through a network. Examples of the above-mentioned network include but are not limited to the Internet, an enterprise intranet, a local area network, a mobile communication network, and a combination thereof.

[0076] The memory 20 can include a volatile memory, such as a random access memory; the memory can also include a non-volatile memory, such as a flash memory, a hard disk, or a solid-state drive; the memory 20 can also include a combination of the above-mentioned types of memories.

[0077] The computer device further includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30, and the output device 40 may be connected via a bus or other means. Figure 3 Take the connection via the bus as an example.

[0078] The input device 30 can receive input digital or character information, and generate key signal inputs related to the user settings and function controls of the computer device, such as a touch screen, a keypad, a mouse, a trackpad, a touchpad, a pointing stick, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 may include a display device, an auxiliary lighting device (e.g., an LED), and a haptic feedback device (e.g., a vibration motor), etc. The above display device includes but is not limited to a liquid crystal display, a light-emitting diode, a display, and a plasma display. In some alternative embodiments, the display device may be a touch screen.

[0079] Embodiment 4:

[0080] Based on the above embodiments, the embodiment of the present application further provides a computer-readable storage medium, in which a computer program executable by a processor is stored. When the program runs on the processor, the processor is caused to perform the following steps when executing:

[0081] The selector receives the user's question; based on the respective interaction information stored in the memory, one or more target agents for processing the user's question are determined from the multiple agents; for each determined target agent, a restatement question is generated and input to the target agent based on the user's question and the function scope corresponding to the target agent pre-saved.

[0082] For any one of the target agents, receive the restatement question input by the selector; determine the respective subtasks corresponding to the restatement question and the execution relationship between the subtasks; execute the subtasks according to the execution relationship between the subtasks; based on the execution results respectively corresponding to the subtasks, determine the response information corresponding to the restatement question; send the response information to the selector to support the selector in making the next possible agent selection.

[0083] The selector receives the response information fed back by any of the target agents, and determines whether the response information meets the pre-configured output requirements based on the interaction information stored in the memory. If it is determined that the response information does not meet the output requirements, the problem is analyzed again based on the interaction information stored in the memory to generate a new intermediate processing problem, and a corresponding target agent is re-allocated for the intermediate processing problem. Based on the functional scope of the re-allocated target agent and the intermediate processing problem, a new paraphrased problem is generated and output to the re-allocated target agent for processing. This process is repeated in a loop, continuously performing multiple rounds of agent selection and problem processing until it is determined that the received response information meets the output requirements and the response information is output as the problem answer.

[0084] Since the principle of solving problems by the above computer-readable storage medium is similar to the multiple inference method based on a large model, the implementation of the above computer-readable storage medium can refer to the embodiments of the method, and the repeated parts will not be described again.

[0085] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these changes and modifications.

Claims

1. A multi-inference system based on a large model, characterized in that: The system includes: a memory, a selector with multi-round agent selection capability, and a plurality of agents; The memory is used to store the interaction information between the selector and the multiple agents, so as to provide historical data support for the selector in any round of agent selection; wherein the interaction information includes: each historical user question, the answers to the questions corresponding to the historical user questions, and the agent information selected by the historical user questions; The selector is used to receive user questions and solve the user questions based on at least one round of agent selection process; wherein, in the first round of selection, based on each interactive information stored in the memory, one or more target agents for processing the user questions are determined from the multiple agents; for each determined target agent, based on the user question and the pre-stored functional scope corresponding to the target agent, a repetition question is generated and input to the target agent; and, response information fed back by any of the target agents is received, and based on each interactive information stored in the memory, whether the response information meets the pre-configured output requirements; if it is determined that the response information does not meet the output requirements, the problem is analyzed again based on each interactive information stored in the memory to generate a new intermediate processing problem, and the corresponding target agent is re-assigned to the intermediate processing problem, and based on the functional scope of the re-assigned target agent and the intermediate processing problem, a new repetition question is generated and output to the re-assigned target agent for processing, and this cycle is repeated, and multiple rounds of agent selection and problem processing are continuously performed until it is determined that the received response information meets the output requirements and the response information is output as the answer to the question; The multiple agents are used to receive the retelling question input by the selector; determine the subtasks corresponding to the retelling question and the execution relationship between the subtasks; execute the subtasks according to the execution relationship between the subtasks; determine the response information corresponding to the retelling question based on the execution results corresponding to the subtasks; and send the response information to the selector to support the selector in making the next round of possible agent selection.

2. The system according to claim 1, characterized in that The selector is specifically used for: If the interaction information also includes the user intentions corresponding to the historical user questions, then performing intention analysis on the user questions to determine the target intention; Determining target interaction information associated with the target intent from the interaction information stored in the memory; According to the agent information recorded in the target interaction information, a target agent for processing the user question is determined.

3. The system according to claim 1, characterized in that The selector is specifically used for: Acquire the associated interaction information corresponding to the user question stored in the memory; wherein the historical user question in the associated interaction information matches the user question; If the answer to the question in the associated interaction information matches the response information, determining that the response information meets the output requirement; Otherwise, it is determined that the response information does not meet the output requirement.

4. The system according to claim 1, characterized in that The multiple intelligent agents are specifically used for: According to the execution relationship between the subtasks, the subtasks are executed based on a pre-built knowledge base.

5. The system according to claim 1, wherein: The memory is also used to update the stored interaction information in real time.

6. The system according to claim 1, characterized in that The system further comprises: an interaction module; The interaction module is used to interact with the user, receive the user question input by the user, and output the answer to the user question to the user.

7. The system according to claim 6, characterized in that The interaction module is further used to receive feedback information from the user on the answer to the question; and output the feedback information to the memory, so that the memory adjusts the interaction information including the user's question according to the feedback information.

8. A multiple reasoning method based on a large model, characterized in that: The method comprises: The selector receives the user question; based on each interaction information stored in the memory, determines one or more target intelligent agents for processing the user question from the multiple intelligent agents; for each determined target intelligent agent, based on the user question and the pre-stored functional scope corresponding to the target intelligent agent, generates and inputs a restatement question to the target intelligent agent; For any of the target intelligent agents, receiving the retelling question input by the selector; determining the subtasks corresponding to the retelling question and the execution relationship between the subtasks; executing the subtasks according to the execution relationship between the subtasks; determining the response information corresponding to the retelling question based on the execution results corresponding to the subtasks; and sending the response information to the selector to support the selector in making the next round of possible intelligent agent selection; The selector receives response information fed back by any of the target agents, and determines whether the response information meets the pre-configured output requirements based on the interaction information stored in the memory; if it is determined that the response information does not meet the output requirements, the problem is analyzed again based on the interaction information stored in the memory to generate a new intermediate processing problem, and the corresponding target agent is reallocated to the intermediate processing problem, and a new repetition problem is generated and output to the reallocated target agent for processing based on the functional scope of the reallocated target agent and the intermediate processing problem, and this cycle is repeated for multiple rounds of agent selection and problem processing until it is determined that the received response information meets the output requirements and the response information is output as the answer to the question.

9. A computer device, characterized in that: The computer device comprises a processor, and the processor is used to implement the steps of the large model-based multiple reasoning method as claimed in claim 8 when executing the computer program stored in the memory.

10. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements the steps of the large model-based multiple reasoning method as described in claim 8 above.

Images