Multi-agent system-based information processing method and multi-agent system

Abstract

The present disclosure provides a multi-agent system-based information processing method and a multi-agent system. The multi-agent system comprises multiple agents. The method is applied to a target agent, and the target agent is any agent among the multiple agents. The method comprises: determining information to be processed corresponding to a target agent; on the basis of an information processing model, when said information does not match a function corresponding to the target agent, determining an agent to which a function matching said information corresponds; and sending said information to the matching agent for processing. In the present disclosure, information to be processed can be quickly and accurately transferred between agents by utilizing the analytical capability of an information processing model, and is processed by a suitable agent, thereby effectively improving the processing effect of tasks.

Description

Information processing method based on multi-agent system and multi-agent system

[0001] This disclosure claims priority to the Chinese patent application filed with the China Patent Office on January 12, 2024, with application number 202410052559.2 and application name “Information processing method based on multi-agent system and multi-agent system”, the entire contents of which are incorporated by reference into this disclosure. Technical Field

[0002] The present disclosure relates to artificial intelligence technology, and in particular to an information processing method based on a multi-agent system and a multi-agent system. Background Art

[0003] With the continuous development of artificial intelligence, agent technology has received more and more attention and has gradually become an important research topic in the field of artificial intelligence.

[0004] At present, how to effectively use intelligent agents to process tasks and ensure the processing effect of tasks is a problem that needs to be solved. Summary of the Invention

[0005] The present disclosure provides an information processing method based on a multi-agent system and a multi-agent system to improve the task processing effect of the agents.

[0006] In a first aspect, an embodiment of the present disclosure provides an information processing method based on a multi-agent system, wherein the multi-agent system includes multiple agents; the method is applied to a target agent, wherein the target agent is any one of the multiple agents; the method comprises:

[0007] Determining the information to be processed corresponding to the target agent;

[0008] Based on the information processing model, when the information to be processed does not match the function corresponding to the target agent, determining an agent whose function matches the information to be processed;

[0009] The information to be processed is sent to a matching agent for processing.

[0010] Optionally, the information processing model is used to: generate corresponding reply information when the information to be processed matches the function corresponding to the target intelligent agent; and output the identifier of the intelligent agent that matches the information to be processed when the information to be processed does not match the function corresponding to the target intelligent agent.

[0011] Optionally, based on the information processing model, when the information to be processed does not match the function corresponding to the target agent, determining an agent whose function matches the information to be processed includes:

[0012] The information to be processed and the description information of the functions corresponding to each of the multiple agents are input into the information processing model so that when the information to be processed does not match the function corresponding to the target agent, the information processing model determines the agent whose function matches the information to be processed.

[0013] Optionally, inputting the information to be processed and description information of the function corresponding to each of the multiple agents into the information processing model includes:

[0014] Inputting the information to be processed, the description information of the function corresponding to the target agent, and first prompt information into the information processing model, wherein the first prompt information is used to prompt the information processing model to generate corresponding reply information when it is determined based on the information to be processed and the description information that the information to be processed matches the function corresponding to the target agent, and output a preset message indicating that a reply is impossible when the information does not match;

[0015] If the output of the information processing model is the preset information, the information to be processed, the descriptive information of the functions corresponding to the remaining intelligent agents and the second prompt information are input into the information processing model, wherein the remaining intelligent agents are the intelligent agents other than the target intelligent agent among the multiple intelligent agents, and the second prompt information is used to prompt the information processing model to determine the intelligent agent whose function matches the information to be processed based on the information to be processed and the descriptive information corresponding to the remaining intelligent agents.

[0016] Optionally, inputting the information to be processed and description information of the function corresponding to each of the multiple agents into the information processing model includes:

[0017] Inputting the information to be processed, description information of the function corresponding to each of the multiple agents, and third prompt information into the information processing model;

[0018] Among them, the third prompt information is used to prompt the information processing model to determine whether the information to be processed matches the function corresponding to the target intelligent agent based on the description information of the information to be processed and the functions corresponding to each intelligent agent, and generate corresponding reply information when they match, and determine the intelligent agent whose function matches the information to be processed when they do not match.

[0019] Optionally, the information input into the information processing model further includes: fourth prompt information;

[0020] The fourth prompt information is used to prompt the information processing model to reply to the information to be processed based on the knowledge learned during the training process when it is determined that there is no agent among the multiple agents whose functions match the information to be processed.

[0021] Optionally, the information to be processed includes user input information;

[0022] The target agent is an agent in the multi-agent system that is marked as an entry agent; the information to be processed input into the multi-agent system is received and processed by the entry agent.

[0023] Optionally, the multi-agent system is configured to conduct at least one round of dialogue with a user; in any round, the information to be processed input to the multi-agent system includes input information of the user in that round;

[0024] Sending the information to be processed to a matching agent for processing, including:

[0025] Sending the information to be processed to a matching agent for processing, and transferring the tag of the entry agent to the matching agent, so that the matching agent receives and processes the information to be processed in the next round input to the multi-agent system;

[0026] The entry agent corresponding to the first round is the agent pre-selected in the multi-agent system.

[0027] Optionally, the information to be processed includes user input information; sending the information to be processed to a matching agent for processing includes:

[0028] Acquiring perception configuration information, where the perception configuration information is used to indicate whether to inform the user to transfer the information to be processed to a matching agent for processing;

[0029] If so, the transfer action is notified to the user, and the information to be processed is transferred to the matching agent for processing.

[0030] Optionally, the method further includes:

[0031] After the multi-agent system is constructed, sending description information of the function corresponding to the target agent to the remaining agents in the multi-agent system, and receiving description information of the corresponding functions sent by the remaining agents;

[0032] The received description information is stored in the memory module, so that when the information to be processed is obtained, the description information corresponding to each agent is read from the memory module and input into the information processing model for processing.

[0033] In a second aspect, an embodiment of the present disclosure provides a method for constructing a multi-agent system, comprising:

[0034] Outputting a list of agents contained in an agent resource library, wherein the agent resource library includes at least two trained agents;

[0035] Obtaining a plurality of agents selected by a client for a target application from the list;

[0036] Build a multi-agent system corresponding to the target application based on the agent selected by the customer;

[0037] Wherein, any agent in the multi-agent system is used to execute any method described in the first aspect.

[0038] Optionally, the method further includes:

[0039] After constructing the multi-agent system corresponding to the target application, a notification message is sent to each agent in the multi-agent system, where the notification message is used to instruct the multiple agents to forward description information of their corresponding functions to each other.

[0040] Optionally, the method further comprises at least one of the following:

[0041] Acquire the agent trained or uploaded by the customer, and add the agent trained or uploaded by the customer to the agent resource library;

[0042] Obtaining description information added by a client for one or more agents in the agent resource library;

[0043] Obtaining an entry agent and / or entry configuration information specified by a client for the multi-agent system, wherein the entry configuration information indicates whether to transfer a tag of the entry agent to a matching agent when information to be processed does not match the function of the current entry agent;

[0044] Acquire the perception configuration information configured by the client for the multi-agent system; wherein the perception configuration information is used to indicate whether to inform the user to transfer the information to be processed to a matching agent for processing.

[0045] Optionally, the method further includes:

[0046] Obtaining function and / or capability requirement information input by a customer for one or more agents in the agent resource library;

[0047] Determine, based on the function and / or capability requirement information of each agent, a capability enhancement method corresponding to each agent, and process each agent according to the corresponding capability enhancement method;

[0048] Among them, the capability enhancement method includes at least one of the following: fine-tuning the information processing model in the intelligent agent, reinforcement learning for a single intelligent agent, reinforcement learning with other intelligent agents, and configuring prompt information for the intelligent agent.

[0049] In a third aspect, the present disclosure further provides an information processing method based on a multi-agent system, comprising:

[0050] Obtaining information to be processed corresponding to the multi-agent system;

[0051] determining an entry agent in the multi-agent system;

[0052] The information to be processed is sent to the entry agent so that when the information to be processed matches the function of the entry agent, the entry agent generates corresponding reply information. When the information to be processed does not match the function of the entry agent, the agent in the multi-agent system whose function matches the information to be processed is determined based on the information processing model, and the information to be processed is sent to the matching agent for processing.

[0053] In a fourth aspect, an embodiment of the present disclosure further provides an information processing method based on a multi-agent system, wherein the multi-agent system includes a plurality of agents; the method is applied to a target agent among the plurality of agents; the method comprises:

[0054] Get user input information;

[0055] If the input information matches the function corresponding to the target agent, the corresponding reply information is generated through the information processing model;

[0056] If the input information does not match the function corresponding to the target agent, the input information is sent to an agent whose function matches the input information for processing; wherein the matching agent is determined by the information processing model.

[0057] Optionally, the input information and the reply information are used to be displayed in the group chat interface;

[0058] Sending the input information to an agent whose function matches the input information for processing, including:

[0059] Acquiring perception configuration information, where the perception configuration information is used to indicate whether to inform the user to transfer the information to be processed to a matching agent for processing;

[0060] If so, the user is notified of the forwarding behavior through the group chat interface, and the input information is forwarded to the matching agent for processing, so that the matching agent generates corresponding reply information and displays it to the user through the group chat interface.

[0061] In a fifth aspect, an embodiment of the present disclosure further provides an information processing method, the method comprising:

[0062] Get information to be processed;

[0063] Inputting the information to be processed and description information of the functions corresponding to the multiple information processing models into a target information processing model, so that the target information processing model outputs corresponding reply information when the information to be processed matches the function of the target information processing model, and determines the information processing model that matches the information to be processed among the multiple information processing models when there is no match; wherein the target information processing model is any information processing model among the multiple information processing models;

[0064] If the output of the target information processing model is used to represent the matched information processing model, the information to be processed is sent to the matched information processing model for processing.

[0065] In a sixth aspect, an embodiment of the present disclosure further provides a multi-agent system, comprising a plurality of agents, wherein the agents are used to execute the method described in any one of the first and fourth aspects.

[0066] In a seventh aspect, the embodiments of the present disclosure further provide a multi-agent based application platform, including a control module and an agent resource library;

[0067] The control module is used to execute the method described in any one of the second aspect or the third aspect.

[0068] In an eighth aspect, an embodiment of the present disclosure provides an electronic device, including:

[0069] at least one processor; and

[0070] a memory communicatively coupled to the at least one processor;

[0071] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the electronic device to execute the method described in any one of the above aspects.

[0072] In a ninth aspect, an embodiment of the present disclosure provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When a processor executes the computer-executable instructions, the method described in any one of the above aspects is implemented.

[0073] In a tenth aspect, an embodiment of the present disclosure provides a computer program product, comprising a computer program, which implements the method described in any of the above aspects when executed by a processor.

[0074] The information processing method based on a multi-agent system and the multi-agent system provided by the embodiments of the present disclosure can construct a multi-agent system based on multiple agents, realize task processing by using the multi-agent system, and better complete the task through the cooperation between the functions of each agent. Moreover, after any target agent in the multi-agent system receives the information to be processed, if the information to be processed is not within the functional scope of the target agent, the agent that matches the information to be processed can be determined through the information processing model corresponding to the target agent, and the information to be processed can be transferred to the matching agent for processing, thereby utilizing the understanding and analysis capabilities of the information processing model to quickly and accurately realize the flow of information to be processed between agents, and having the appropriate agent process it, effectively improving the processing effect of the task. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0076] FIG1 is a schematic diagram of an application scenario provided by an embodiment of the present disclosure;

[0077] FIG2 is a flow chart of an information processing method based on a multi-agent system provided by an embodiment of the present disclosure;

[0078] FIG3 is a schematic diagram of an interaction principle of an intelligent agent provided by an embodiment of the present disclosure;

[0079] FIG4 is a schematic diagram of an architecture of an intelligent agent provided by an embodiment of the present disclosure;

[0080] FIG5 is a flow chart of a method for constructing a multi-agent system according to an embodiment of the present disclosure;

[0081] FIG6 is a schematic diagram showing the construction principle of a multi-agent system for a target application provided by an embodiment of the present disclosure;

[0082] FIG7 is a schematic diagram of an interactive interface provided by an embodiment of the present disclosure;

[0083] FIG8 is a flow chart of another information processing method based on a multi-agent system provided by an embodiment of the present disclosure;

[0084] FIG9 is a schematic diagram of a group chat interface provided by an embodiment of the present disclosure;

[0085] FIG10 is a schematic structural diagram of an electronic device provided in an embodiment of the present disclosure.

[0086] The above drawings illustrate specific embodiments of the present disclosure, which will be described in more detail below. These drawings and textual descriptions are not intended to limit the scope of the present disclosure in any way, but rather to illustrate the concepts of the present disclosure to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0087] Herein, exemplary embodiments will be described in detail, examples of which are shown in the accompanying drawings.The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure.

[0088] It should be noted that the user information (including but not limited to user device information, user attribute information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards, and corresponding operation entrances must be provided for users to choose to authorize or refuse.

[0089] The disclosed embodiments can be implemented through a large model, for example, a large language model. Among them, a large language model refers to a deep learning language model with large-scale language model parameters, which usually contains hundreds of millions, tens of billions, hundreds of billions, trillions or even more than ten trillion language model parameters. The large language model can also be called a cornerstone language model / foundation model (Foundation Model). The large language model is pre-trained through large-scale unlabeled corpus to produce a pre-trained language model with more than 100 million parameters. This language model can adapt to a wide range of downstream tasks, and the language model has good generalization ability, such as a large-scale language model (LLM), a multi-modal pre-training language model, etc.

[0090] In practical applications, large language models only require a small number of samples to fine-tune the pre-trained language model and can be applied to different tasks. Large language models can be widely used in natural language processing (NLP), computer vision and other fields. Specifically, they can be applied to computer vision tasks such as visual question answering (VQA), image caption (IC), and image generation, as well as natural language processing tasks such as text-based sentiment classification, text summary generation, and machine translation. The main application scenarios of large language models include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design.

[0091] First, the terms involved in this disclosure are explained:

[0092] Agent: A solution implemented by a model that maintains descriptive information about functions, inputs and outputs, etc.

[0093] Multi-agent group chat: Multiple agents exchange information through two-to-two interactions, thereby understanding what other agents are doing.

[0094] The application scenarios of the present disclosure are first described below.

[0095] The intelligent agent can process tasks through interaction with the external environment and is applicable to various scenarios in various fields. Optionally, the intelligent agent can be implemented based on a trained information processing model, which processes the acquired information. The information processing model can be a large language model, or other models. For example, a model with a small number of parameters can be constructed and trained to serve as the information processing model.

[0096] Taking the intelligent customer service scenario as an example, the intelligent agent can obtain the user's input information, generate reply information corresponding to the input information based on the information processing model, and output it to the user. In the process of interacting with the user, it can also implement API (Application Programming Interface) calls, document queries and other operations based on the information processing model, so as to better handle user needs.

[0097] As user demands continue to increase, a single agent is no longer sufficient to complete user tasks. Therefore, it is possible to consider implementing user interaction based on a multi-agent system. A multi-agent system can include multiple agents, each with different functions and expertise, thus providing better user service based on the capabilities of multiple agents. However, how to coordinate multiple agents is a challenge that needs to be addressed.

[0098] Specifically, when a user is communicating with an agent in a multi-agent system, he or she may raise a request that the agent cannot handle. At this time, the user can jump to other agents for processing. However, when the jump is needed, the current agent may not know which agent can handle this request, resulting in difficulties in the jump process.

[0099] For example, the user is currently asking questions based on a document, and the intelligent agent responsible for document question and answering interacts with the user. After asking multiple questions, a summary needs to be generated based on the previous conversation information. Traditional document question and answering agents are only good at finding answers from documents and are not good at processing summaries. At this time, it is necessary to call the intelligent agent responsible for generating summaries. However, how to call it is a difficult point.

[0100] To address these issues, consider using the MoE (Mixture of Experts) approach. MoE is a deep learning architecture whose core idea is to combine many smaller, area-specific learning models (called expert models) to learn and reason on a larger dataset. In other words, MoE is a divide-and-conquer strategy for building larger models that contain many small, specialized models.

[0101] MoE uses a gating model to effectively utilize multiple expert models. Specifically, the gating model can dynamically determine which expert model or models will be used for processing based on the characteristics of the input information to obtain the final output result.

[0102] However, in the MoE solution, the design and training of the gating mechanism is a challenge. It is necessary to effectively balance the contributions of different expert models by training the gating model and multiple expert models. This leads to problems such as complex processing and high computational overhead.

[0103] In view of this, the embodiments of the present disclosure provide a method that can use multiple agents to build a multi-agent system, use the natural language understanding capabilities of the information processing model in the agent to guide the interaction between the agents, and thus effectively realize the dialogue with the user.

[0104] Figure 1 is a schematic diagram of an application scenario provided by an embodiment of the present disclosure. As shown in Figure 1, a multi-agent system includes multiple agents, each of which may include an information processing model and a description information list, which includes description information of the functions corresponding to each agent in the system.

[0105] Exemplarily, the description information list may include:

[0106] Description of Agent A: This is an agent focused on document question answering. Given a document and a question, it can find the knowledge corresponding to the question from the document and generate the corresponding answer.

[0107] Agent B's description: This agent is good at generating summaries and key points. It can summarize and generalize a given text, generate a corresponding summary, and extract the key points.

[0108] Description of Agent C: This is an agent used to handle code-related requirements. It has certain code understanding and generation capabilities and can implement code generation, code completion, code checking, programming suggestions, etc.

[0109] The agents included in the multi-agent system and the corresponding description information of each agent can be configured by the customer.

[0110] The "customer" in the embodiments of the present disclosure refers to the role that configures the multi-agent system, for example, it can be a corporate customer or other type of customer that provides intelligent customer service services, and the "user" refers to the role that interacts with the multi-agent system in actual applications.

[0111] When a user interacts with an agent in a multi-agent system, if a question arises that the current agent cannot answer, the information processing model can be used to understand the descriptive information, determine which agent can answer the question, and transfer the question to the corresponding agent for processing.

[0112] For example, when a user interacts with agent A, he or she can ask questions based on the document and agent A will answer them. If, during the interaction, the user enters a question that requires code generation, the information processing model of agent A determines that it cannot solve the problem. Moreover, by analyzing the descriptive information of each agent, it can be known that agent C is able to handle the problem, and the problem can be transferred to agent C for processing.

[0113] Through this approach, agents can leverage the model's natural language understanding capabilities to fully understand the areas of expertise of each other agent. During user interactions, they can delegate questions they can't resolve to agents with expertise in those areas. Overall, intelligent customer service implemented using a multi-agent system allows user questions to flow between multiple agents, facilitating the quick and accurate selection of the appropriate agent to answer the user, thereby enhancing the effectiveness of user interaction.

[0114] Moreover, the entire process does not require the intervention of a gating model, avoiding the design and training of a gating mechanism. It can organize multiple intelligent agents without retraining. The processing process is simple and clear, and does not require additional computing overhead.

[0115] Some embodiments of the present disclosure are described in detail below with reference to the accompanying drawings. The following embodiments and features thereof may be combined with one another unless they conflict with each other. Furthermore, the sequence of steps in the following method embodiments is provided for illustrative purposes only and is not intended to be a strict limitation.

[0116] Figure 2 is a flow chart of an information processing method based on a multi-agent system provided in an embodiment of the present disclosure. In this embodiment, the deployment method of the multi-agent system is not limited, for example, it can be deployed locally, implemented in the cloud, implemented on a client, implemented on an IOT (Internet of Things) device, etc.

[0117] The multi-agent system includes multiple agents. Optionally, each of the multiple agents may correspond to different functions. Alternatively, some of the multiple agents may correspond to the same function. The functions may be configured based on actual needs. For example, functions may include solving math problems, translating, generating summaries, processing code-related requirements, processing images, etc. Different agents may correspond to different functions.

[0118] The method is applied to a target agent, which is any agent among the multiple agents. As shown in FIG2 , the method includes:

[0119] Step 201: Determine the information to be processed corresponding to the target agent.

[0120] Optionally, the target intelligent agent can realize corresponding functions based on an information processing model, and the information processing model can be a trained model.

[0121] One optional architecture is for the target agent to include an information processing model, which then performs the corresponding function. For example, if the target agent is an agent that excels at solving math problems, its corresponding information processing model could be an information processing model that excels at solving math problems. By inputting the math problem to the information processing model, the corresponding result can be obtained.

[0122] The target agent may correspond to one or more functions. The functions corresponding to the target agent may be consistent with the functions corresponding to the information processing model, and may not be limited to the functions corresponding to the information processing model.

[0123] For example, in addition to the information processing model, the target agent may also include one or more plug-ins or other modules to enable more capabilities. For example, plug-ins can enable operations such as logging into a browser, downloading documents, and calling APIs. Therefore, the target agent can have more capabilities than the information processing model.

[0124] In addition to the architecture in which the target intelligent agent includes the information processing model, the information processing model and the target intelligent agent can also be two separate parts, as long as the capabilities of the information processing model can be utilized to realize the corresponding functions of the target intelligent agent.

[0125] The target agent can be used to interact with the environment, and the information to be processed can include information obtained by the target agent from the environment. The environment can be anything outside the target agent, for example, the environment can include users, API systems, document systems, other agents, etc.

[0126] Among them, the target intelligent agent can simulate the interaction process between humans and the environment, obtain some information from the environment, make decisions based on this information, and determine the next behavior to be performed (output information). After the behavior acts on the environment, it can obtain information from the environment and make further decisions and behaviors. In this way, task processing can be achieved through interaction with the environment.

[0127] Optionally, if the environment includes a user, the target agent can interact with the user, and the information obtained from the environment may include the user's input information, and the information output to the environment may include the target agent's output information to the user. For example, the user can input a question and the target agent can answer it, or the target agent can ask the user certain questions and the user can answer them.

[0128] If the environment includes an API system, the target agent can implement API calls through interaction with the API system. Specifically, the API system may include multiple APIs, and the agent's output information may include the name and input parameter values ​​of the API to be called, so that the API system can return the output parameter values ​​of the corresponding API based on the API name and input parameter values. Accordingly, the information obtained by the target agent can be the output parameter value of the API.

[0129] If the environment includes a document system, the target agent can realize document query through interaction with the document system. Specifically, the document system may include multiple documents, and the information output by the agent may include information to be queried, so that the document system can determine the corresponding document based on the information to be queried. Accordingly, the information obtained by the target agent may be a document that matches the information to be queried.

[0130] If the environment includes other agents, the information output by the target agent to the environment may include information instructing other agents to perform processing. For example, if the target agent determines that agent A can handle a certain requirement, it can send the requirement to agent A and obtain the corresponding processing result from agent A. The other agents can be other agents in the multi-agent system or agents outside the system, without limitation.

[0131] Optionally, the environment can include multiple types of objects. For example, the environment can include both an API system and a user. The target agent can output a greeting message to the user and obtain the user's input requirements. After analyzing the requirements, if it finds that a certain API can be called to assist in fulfilling the requirements, it can interact with the API system and output the corresponding results to the user based on the information returned by the API system.

[0132] In this embodiment, the pending information may be information corresponding to a pending task, and one task may correspond to one or more pending information. The tasks may be set according to actual needs. For example, the tasks may include answering user questions, guiding users to complete a procedure, chatting with users, or fulfilling a certain need.

[0133] For example, the task to be processed may be answering user questions. The user may have one or more questions, and each question may serve as information to be processed.

[0134] When there are multiple pieces of information to be processed, the method described in this embodiment can be executed for each piece of information to be processed.

[0135] Step 202: Based on the information processing model, when the information to be processed does not match the function corresponding to the target agent, determine an agent whose function matches the information to be processed.

[0136] The matching of the information to be processed and the function means that the function possessed by the target agent is suitable for processing the information to be processed, or is capable of processing the information to be processed. There are multiple implementation options for how the model can determine whether this is a match.

[0137] Alternatively, the model's own capabilities can be used to determine whether a match exists, or the client can constrain whether a match exists based on actual conditions. The former can be achieved through prompt information, while the latter can be achieved through the training process. These are explained below.

[0138] In an optional implementation, the model has a certain degree of natural language understanding and analysis capabilities, and can determine whether the two match based on its own understanding of the information to be processed and the function.

[0139] Specifically, given a piece of information to be processed and the function of an agent, the model can determine whether the information to be processed matches the function of the agent. Given a piece of information to be processed and the functions of multiple agents, the model can determine which of the multiple agents has a function that matches the information to be processed.

[0140] Optionally, prompt information can be used to allow the model to determine the agent whose function matches the information to be processed. For example, when the information to be processed is user input, the information input to the model can be:

[0141] "The user's input information is XXXXX. You are an intelligent agent with XXX capabilities. Please determine whether the user's input information is within your capabilities (whether it matches your capabilities):

[0142] If it is within your capabilities (if it matches), process the user's input information;

[0143] If it is beyond your capabilities (if it does not match), please select a suitable agent to handle it (determine the matching agent) based on the functions of other agents. The functions of other agents are as follows:

[0144] Agent A, function is XXXX; Agent B: function is XXXX; Agent C: function is XXXX."

[0145] Based on the above prompts, the model can analyze the user's input information and the functions of the target agent, and process the user's input information when the user's input information matches the functions. If they do not match, the model determines which agent can process the user's input information based on the functions of other agents.

[0146] For example, the information input to the model can specifically be:

[0147] "The user input is: 'Based on the following, write a Python code: Given n numbers, arrange them in descending order. If there are any repeated numbers, display the number of repetitions.'

[0148] You are an agent that translates Chinese into English. Please determine whether the user's input information is within your capabilities:

[0149] If it is within your capabilities, process the user's input information;

[0150] If it is beyond your capabilities, please answer "I am not good at processing" and determine which agent is suitable for processing the user's input information based on the functions of other agents. The functions of other agents are as follows:

[0151] Agent A, used to generate corresponding answers based on documents and questions;

[0152] Agent B, used to generate images based on text;

[0153] Agent C is used to generate code in multiple languages ​​such as C, C++, Java, and Python.

[0154] Because the model has a certain level of understanding and analysis capabilities, it can determine whether the user's input information and the target agent's capabilities match. If they don't match, it can identify and output an agent with capabilities that match the input information. For example, after inputting the above information into the model, the model might output "I am not good at processing this information, and Agent C is suitable for processing this information." Based on the model's output, it can be determined that Agent C's capabilities match the user's input information.

[0155] Another alternative implementation is to have the agent learn to accurately respond to input information during the training phase, and to train the agent to output the expected information when the input information exceeds the agent's functional scope. For example, if the target agent is a translation agent, the corresponding training samples may include:

[0156] Training sample 1: The input information is "Please translate the following content into English: XXXXX", and the output information is the corresponding English text;

[0157] Training sample 2: The input information is "Please generate a piece of code based on the following content: XXXXXX", and the output information is "I am not good at processing, but the code generation agent can handle it";

[0158] Training sample 3: The input information is "Please generate the corresponding image based on the following content: XXXXX", and the output information is "I am not good at processing, but the image generation agent can handle it."

[0159] The model is trained to acquire the required capabilities by feeding the input information of each training sample into the model with the goal of obtaining the corresponding output information. After training, prompts can be removed or weakened in actual applications, and the model can directly output the desired response based on the content learned during the training phase. For example, if the input information instructs the agent to translate a text, the agent can process it normally and output the translated text. If the input information instructs the agent to generate a piece of software code, the agent can output "I am not good at processing, but the code generation agent can handle it."

[0160] Through the above approach, a large language model can be used as an information processing model. By adding prompt information, the model can be equipped with the ability to determine whether the information to be processed matches a specific function. Alternatively, the large language model may not necessarily be used. The model can be trained in a targeted manner to acquire the corresponding capabilities.

[0161] Step 203: Send the information to be processed to a matching agent for processing.

[0162] Optionally, the information processing model is used to: generate corresponding reply information when the information to be processed matches the function corresponding to the target intelligent agent; and output the identifier of the intelligent agent that matches the information to be processed when the information to be processed does not match the function corresponding to the target intelligent agent.

[0163] Optionally, when the information to be processed is information obtained from the environment, if corresponding reply information is generated according to the information processing model, the reply information can be output to the environment.

[0164] For example, in the field of intelligent customer service, the environment may include users, the information to be processed corresponding to the target intelligent agent may include the user's input information, and the corresponding reply information may be the answer output to the user.

[0165] Figure 3 is a schematic diagram of the interaction principle of an agent provided by an embodiment of the present disclosure. As shown in Figure 3, the target agent and other agents can both include an information processing model and a system module. The system module can be implemented as software, hardware, or a combination of software and hardware. The above steps provided in this embodiment can be specifically performed by the system module in the target agent.

[0166] Specifically, the system module can interact with the environment, for example, obtain the user's input information as information to be processed, and input the information to be processed into the information processing model. If the information to be processed matches the function corresponding to the target intelligent agent, the information processing model can generate corresponding reply information, and the system module outputs the reply information to the user.

[0167] If the information to be processed does not match the function of the target agent, and the information processing model is unable to respond to the information to be processed, it can output the identifier of the agent that matches the information to be processed. After obtaining the identifier of the matching agent, the system module can forward the information to the matching agent for processing.

[0168] In the matching intelligent agent, after the system module obtains the information to be processed, it can input the information to be processed into the information processing model of the intelligent agent, and the information processing model outputs the corresponding reply information, and the system module of the intelligent agent outputs the reply information to the user.

[0169] Through the above method, the ability of the information processing model can be utilized to give corresponding reply information when it is able to process the information to be processed. When it is difficult to process the information correctly, the intelligent agent that can process the information is determined and transferred to the corresponding intelligent agent for processing. The ability of the information processing model is fully utilized to generate a reply or transfer the information to be processed, thereby improving the processing effect of the target intelligent agent on the information to be processed.

[0170] In addition to the scheme in which the information processing model directly generates reply information based on the information to be processed, the generation of reply information can also be achieved with the help of other modules. For example, you can first query related documents based on the information to be processed, and input the documents and the information to be processed into the information processing model together to obtain the corresponding reply information. Alternatively, you can first input the information to be processed into the information processing model to obtain an intermediate result, and then perform other operations based on the intermediate result, such as querying a database, to obtain the final reply information.

[0171] In other optional implementation schemes, the intelligent agent may not need to output reply information. For example, if the information to be processed indicates that the relevant information of the order should be calculated and updated, after obtaining the calculation results based on the information processing model, the relevant information of the order in the database can be directly updated according to the calculation results without outputting reply information to the user.

[0172] In practical applications, each agent in a multi-agent system can include an information processing model, and each agent can execute the aforementioned scheme. Alternatively, some agents can include information processing models, while the remaining agents only include smaller, specialized models or other decision-making modules. As long as the corresponding functions can be achieved, the aforementioned scheme can be applied to the agents that include information processing models, while the other agents may not be involved in the handover operation, or the handover operation may be achieved through other means.

[0173] The information processing method based on a multi-agent system provided in this embodiment can use multiple agents to build a multi-agent system, use the multi-agent system to implement task processing, and better complete the task through the capabilities of the information processing model corresponding to the agent and the coordination between the functions of each agent. In addition, any target agent in the multi-agent system receives information to be processed. If the information to be processed is not within the functional scope of the target agent, the information processing model corresponding to the target agent can be used to determine the agent that matches the information to be processed, and the information to be processed can be transferred to the matching agent for processing, thereby utilizing the analysis capability of the information processing model to quickly and accurately realize the flow of information to be processed between agents, and having the appropriate agent process it, effectively improving the processing effect of the task.

[0174] There are various ways to determine matching agents. For example, the information to be processed and the names of the agents can be input into the information processing model. The names can reflect the functions of the agents, such as "code agent," "translation agent," or "math problem agent." The information processing model can analyze the information to be processed and the names of the agents to determine matching agents.

[0175] Alternatively, description information can be configured for each agent, and matching agents can be determined based on the description information.

[0176] Optionally, based on the information processing model, when the information to be processed does not match the function corresponding to the target agent, determining an agent whose function matches the information to be processed includes:

[0177] The information to be processed and the description information of the functions corresponding to each of the multiple agents are input into the information processing model so that when the information to be processed does not match the function corresponding to the target agent, the information processing model determines the agent whose function matches the information to be processed.

[0178] Optionally, the identifications and description information corresponding to multiple agents may be input into the information processing model, and the information processing model may output the identification of the matching agent based on the description information.

[0179] Alternatively, the information processing model can first output a conclusion on whether the information to be processed matches the function corresponding to the target agent, and then output the identifier of the matching agent if there is no match. Alternatively, the information processing model can directly output the identifier of the matching agent, and based on this identifier, the target agent can determine that the information to be processed does not match the function corresponding to the target agent and send the information to the matching agent for processing.

[0180] There are many ways to input the description information of the functions corresponding to the multiple agents into the information processing model, which can be a one-time input or a step-by-step input.

[0181] In an optional implementation method, the description information of the function corresponding to the target intelligent agent can be first input into the information processing model. If it matches the information to be processed, the corresponding reply information is directly generated. If the information processing model indicates that it cannot answer, the description information of the functions corresponding to other intelligent agents can be further input into the information processing model, so that the information processing model can select a suitable intelligent agent based on the description information of other intelligent agents.

[0182] Specifically, inputting the information to be processed and the description information of the function corresponding to each of the multiple agents into the information processing model includes:

[0183] Inputting the information to be processed, the description information of the function corresponding to the target agent, and first prompt information into the information processing model, wherein the first prompt information is used to prompt the information processing model to generate corresponding reply information when it is determined based on the information to be processed and the description information that the information to be processed matches the function corresponding to the target agent, and output a preset message indicating that a reply is impossible when the information does not match;

[0184] If the output of the information processing model is the preset information, the information to be processed, the descriptive information of the functions corresponding to the remaining intelligent agents and the second prompt information are input into the information processing model, wherein the remaining intelligent agents are the intelligent agents other than the target intelligent agent among the multiple intelligent agents, and the second prompt information is used to prompt the information processing model to determine the intelligent agent whose function matches the information to be processed based on the information to be processed and the descriptive information corresponding to the remaining intelligent agents.

[0185] For example, assuming that the information to be processed is a user's question, the first prompt information may include: "Please determine whether the user's question is within your ability. If it is, please answer directly. If it is not, please answer: I am not good at this."

[0186] Accordingly, the input to the information processing model may include:

[0187] "You are an intelligent agent with [functional description]. Please determine whether the user's question is within your ability. If it is, please answer directly. If it is not, please answer: I am not good at this; User: [question]".

[0188] Among them, the description information of the function corresponding to the target intelligent agent can be inserted in [Function Description], and the user's question can be inserted in [Question].

[0189] Optionally, since the information processing model itself has the ability to answer questions, the following prompt content can be omitted from the first prompt: When the information to be processed and the description information determine that the information to be processed matches the function corresponding to the target agent, a corresponding reply message is generated. For example, the following content can be omitted from the information input to the information processing model: "If it is within your ability, please answer directly."

[0190] Based on the above input to the information processing model, the preset information is "I'm not good at this." If it is detected that the information processing model outputs the preset information, the information to be processed, descriptions of the functions corresponding to the remaining agents, and a second prompt information can be further input to the information processing model to enable the information processing model to determine an agent whose function matches the information to be processed.

[0191] For example, the second prompt information may be: "Please determine the agent that can answer the user's question based on the description information of the functions corresponding to the following agents."

[0192] Through the above scheme, the information processing model can be guided to output reply information or matching agents using step-by-step prompts. Since the target agent often has the ability to directly answer user questions, there is no need to further select a matching agent from other agents in these cases. Therefore, when the information to be processed is obtained, only the description information of the information to be processed and the target agent is input into the information processing model. There is no need to input the description information of other agents. The description information of other agents is introduced only when necessary, which can effectively reduce the amount of data that needs to be processed. It is especially suitable for scenarios where the information to be processed does not need to be transferred frequently.

[0193] In another optional implementation, the generation of reply information or the determination of a matching intelligent agent can be achieved through a single interaction with the information processing model.

[0194] Specifically, inputting the information to be processed and the description information of the function corresponding to each of the multiple agents into the information processing model includes:

[0195] Inputting the information to be processed, description information of the function corresponding to each of the multiple agents, and third prompt information into the information processing model;

[0196] Among them, the third prompt information is used to prompt the information processing model to determine whether the information to be processed matches the function corresponding to the target intelligent agent based on the description information of the information to be processed and the functions corresponding to each intelligent agent, and generate corresponding reply information when they match, and determine the intelligent agent whose function matches the information to be processed when they do not match.

[0197] For example, the third prompt information may be: "If the user's question is within your ability, please reply to the user's question. If you cannot reply, please select a suitable agent to handle the user's question. When selecting a suitable agent, you can refer to the following description information."

[0198] Through the above method, the generation of reply information or the determination of the matching intelligent agent can be achieved in one go, without the need for multiple interactions with the information processing model, effectively reducing the number of interactions. It is especially suitable for scenarios where information to be processed needs to be transferred more frequently.

[0199] The client can configure the prompt method and choose whether the description information of multiple agents is input into the information processing model at one time or in steps.

[0200] By setting the descriptive information of the functions corresponding to the agent, the natural language understanding ability of the information processing model can be used to analyze and match the information to be processed and the descriptive information, and select the appropriate agent to process the information to be processed. The descriptive information can reflect the functions that the information processing model can achieve in as much detail as possible, effectively improving the accuracy of determining the matching agent.

[0201] Alternatively, the target agent can obtain the description information of other agents in a variety of ways. For example, the description information can be input by the client. Alternatively, each agent maintains its own corresponding description information and can forward it to each other.

[0202] Specifically, the target agent includes a memory module, and the method further includes:

[0203] After the multi-agent system is constructed, sending description information of the function corresponding to the target agent to the remaining agents in the multi-agent system, and receiving description information of the corresponding functions sent by the remaining agents;

[0204] The received description information is stored in the memory module, so that when the information to be processed is obtained, the description information corresponding to each agent is read from the memory module and input into the information processing model for processing.

[0205] Figure 4 is a schematic diagram of the architecture of an agent provided by an embodiment of the present disclosure. As shown in Figure 4, the target agent and other agents can include an information processing model, a system module and a memory module. The memory module can be used to store description information.

[0206] Before the multi-agent system is built, each agent maintains its own description information. After the multi-agent system is built, agents can forward description information to each other. For example, for any target agent in the multi-agent system, it can send its description information to other agents, or receive description information sent by other agents and store it in the memory module.

[0207] In this way, after multiple agents forward description information to each other, each agent's memory module can store the description information of multiple agents. When any target agent obtains information to be processed, it can use the description information of multiple agents stored in the memory module to determine a matching agent, and the matching agent will process the information to be processed.

[0208] By forwarding description information to each other by multiple agents, the target agent can pre-store the description information of multiple agents. When obtaining information to be processed, it can directly process it according to the pre-stored description information, thereby improving the processing efficiency of the information to be processed.

[0209] Based on the technical solution provided in the above embodiment, optionally, the information input into the information processing model also includes: fourth prompt information; wherein, the fourth prompt information is used to prompt the information processing model to reply to the information to be processed based on the knowledge learned during the training process when it is determined that there is no intelligent agent among the multiple intelligent agents whose functions match the information to be processed.

[0210] Exemplarily, the fourth prompt information may include: "If there is no suitable intelligent agent that can handle the user's question, please output a corresponding reply based on the known information."

[0211] In practical applications, information processing models will learn a lot of knowledge during the training process. Even if the current information to be processed does not belong to the field of expertise of the information processing model, the information processing model can still respond based on the knowledge learned in the past.

[0212] In this way, when there is a suitable agent that can handle the information to be processed, the information to be processed will be transferred to the appropriate agent for processing. When there is no suitable agent that can handle it, a reply will be made based on the existing knowledge of the information processing model to avoid invalid transfers and improve the overall processing efficiency of the multi-agent system.

[0213] In one or more embodiments of the present disclosure, optionally, the information to be processed includes user input information; the target agent is an agent marked as an entry agent in the multi-agent system; the information to be processed input into the multi-agent system is received and processed by the entry agent.

[0214] Optionally, the entry agent can be set by the customer. When the information to be processed includes user input information, an agent in the multi-agent system that can communicate with the user in a more friendly manner can be selected as the entry agent.

[0215] When the multi-agent system interacts with the user as a whole, the information input into the multi-agent system will be sent to the entry agent for processing. If the entry agent determines that it is not good at processing, it can be forwarded to a matching agent for processing.

[0216] Alternatively, marking the entry agent can be implemented in a variety of ways, such as adding a flag to a particular agent among multiple agents to indicate that it is the entry agent, or setting a field corresponding to the entry agent and updating the corresponding value in the field to the ID of the agent to mark the agent as the entry agent. When a user enters information to be processed, the presence of a corresponding flag for each agent or the value of the corresponding field for the entry agent can determine which agent is the entry agent, and the information to be processed can be sent to the entry agent.

[0217] Optionally, the multi-agent system is used to conduct at least one round of dialogue with the user; in any round, the information to be processed input into the multi-agent system includes the input information of the user in that round.

[0218] The information to be processed is sent to a matching agent for processing, including: sending the information to be processed to a matching agent for processing, and transferring the mark of the entry agent to the matching agent, so as to receive the next round of information to be processed input into the multi-agent system through the matching agent and process it.

[0219] Specifically, the multi-agent system can conduct at least one round of dialogue with the user as a whole, and at least one round of dialogue can be recorded as a session. A session can usually include multiple rounds of interaction between the multi-agent system and the user, and each round of interaction can include one input from the user and one output from the multi-agent system.

[0220] In each round, within the multi-agent system, the current entry agent can process the user's input information and obtain corresponding reply information. If the current entry agent cannot process it, it can be transferred to another agent in the multi-agent system for processing and obtain corresponding reply information. At the same time, the entry agent's mark can be transferred to the other agent.

[0221] After the reply information of the current round is output to the user, the user can input the input information of the next round. In the multi-agent system, the other agent acts as the entry agent to process the user's input information. After multiple rounds of interaction, the session can be ended.

[0222] Exemplarily, the multi-agent system interacts with users as an intelligent customer service. The initial entry agent is an ordinary question-and-answer agent, which answers questions raised by users. After one or more rounds of interaction, users begin to input code-related requirements. If the question-and-answer agent determines that it is not good at answering, it can send the code-related requirements to the code agent and transfer the entry agent's tag to the code agent. The code agent will serve as the entry agent corresponding to subsequent rounds to receive and respond to the code-related requirements input by the user.

[0223] Optionally, the information to be processed in each round may include not only the user's input information in the current round, but also the interaction information of previous rounds, for example, the user's input information in each previous round and the output information of the multi-agent system, so that the historical context can be referred to, allowing the multi-agent system to process the user's needs more accurately.

[0224] Optionally, the entry agent corresponding to the first round is a pre-selected agent in the multi-agent system. The method further includes: if the information to be processed matches the function of the target agent, the entry agent remains unchanged.

[0225] Specifically, the client can pre-select a default entry agent from multiple agents. After a session begins, the entry agent corresponding to the first round can be the default entry agent. During the interaction with the user, the entry agent may change. After the session ends, the default entry agent can be restored.

[0226] In one processing method, when the session ends, the tag of the entry agent can be transferred to the pre-selected agent, so that when the next session starts, the pre-selected agent is used as the entry agent to process the first round of information to be processed.

[0227] In another processing method, the step of loading the entry agent can be executed each time a session is opened. For example, when a session is opened, a pre-selected agent in the multi-agent system can be determined and set as the current entry agent.

[0228] Optionally, a control module can be configured to maintain the entry agent corresponding to the current session. After the multi-agent system is released, multiple users can interact with the system simultaneously. This means that multiple sessions may be active at the same time. The control module can maintain information corresponding to each session, such as the entry agent corresponding to the session. This allows the control module to process each user's input information based on the entry agent corresponding to the session to which the input information belongs.

[0229] Through the above scheme, the entry agent in the multi-agent system can receive and process the information to be processed, so as to achieve a fast and orderly response to the information to be processed. Moreover, the identity of the entry agent can be transferred while transferring the information to be processed. Since users often ask questions in different fields separately and continuously, the overall processing efficiency of the multi-agent system can be effectively improved, and the user experience can be enhanced.

[0230] In one or more embodiments of the present disclosure, optionally, the information to be processed includes user input information; and sending the information to be processed to a matching agent for processing includes:

[0231] Acquiring perception configuration information, where the perception configuration information is used to indicate whether to inform the user to transfer the information to be processed to a matching agent for processing;

[0232] If so, the transfer action is notified to the user, and the information to be processed is transferred to the matching agent for processing.

[0233] The perception configuration information can be set by the client or by default. The perception configuration information is mainly used to control whether to inform the user of the handover behavior between agents.

[0234] Optionally, the multiple agents can interact with the user through group chat. In the group chat interface, the entry agent responds to the questions input by the user. When the entry agent cannot answer the user's question, a transfer operation is performed. If the perception configuration information indicates that the transfer behavior needs to be informed to the user, the transfer behavior is first notified to the user, and then the question is transferred to the matching agent for processing, and the matching agent responds to the user's question in the group chat interface.

[0235] For example, the following information can be output to the user to inform the user of the transfer behavior: "Sorry, I cannot answer this question, the code agent can answer it, I will transfer the question to the code agent for processing."

[0236] Through the above method, the transfer behavior between intelligent agents can be notified to the user, thereby improving the user experience. Moreover, whether the transfer behavior is notified can be configured according to actual needs, which can meet the customer's personalized configuration needs in different scenarios.

[0237] On the basis of the information processing method based on the multi-agent system provided in the above embodiment, the embodiment of the present disclosure also provides a method for constructing a multi-agent system.

[0238] Figure 5 is a flow chart of a method for constructing a multi-agent system according to an embodiment of the present disclosure. The method can be applied to an application platform that can maintain a large number of agents. Clients can create applications on the platform and configure multiple agents for the application to implement the application through the corresponding functions of the multiple agents.

[0239] As shown in FIG5 , the method may include:

[0240] Step 501: Output a list of agents contained in an agent resource library, wherein the agent resource library includes at least two trained agents.

[0241] The trained agent can be a pre-trained agent or an agent that has been pre-trained and fine-tuned. The agent can be trained using supervised fine-tuning (SFT) training samples or reinforcement learning.

[0242] Optionally, an agent trained or uploaded by a customer may be obtained and added to the agent resource library. Descriptive information added by the customer for one or more agents in the agent resource library may also be obtained.

[0243] Among them, the descriptive information can be the descriptive information corresponding to the function of the intelligent agent, which is used to describe the function in detail and can be configured by the customer. For example, for an intelligent agent that focuses on math problems, the corresponding descriptive information can be: This is an intelligent agent that focuses on math problems, can solve math problems in primary and secondary schools, and has certain natural language understanding and generation capabilities.

[0244] In addition, the platform can also provide a number of general intelligent agents. The intelligent agents provided by the platform and the intelligent agents trained by customers are all stored in the intelligent agent resource library for customers to use when creating applications later.

[0245] When a customer has a need to create an application, the platform can output to the customer a list of agents contained in the agent resource library, which may include the agent's identification, name, description information, etc.

[0246] Step 502: Obtain multiple agents selected by the client from the list for the target application.

[0247] Specifically, based on the list of agents contained in the current agent resource library, customers can create target applications on the platform and select multiple agents for the target applications. Based on the capabilities of the platform, customers can build their own applications by combining agents.

[0248] The target application may be any application that the customer desires to create, for example, it may be an application for implementing intelligent customer service functions.

[0249] Step 503: Construct a multi-agent system corresponding to the target application based on the agent selected by the customer.

[0250] Among them, any agent in the multi-agent system is used to execute the information processing method described in any embodiment of the present disclosure.

[0251] Specifically, after the customer selects multiple agents for the target application, the target application can be bound to the multiple agents to obtain a multi-agent system corresponding to the target application. The target application can then be used to implement corresponding tasks, such as interacting with users. When implementing the corresponding tasks, the information to be processed is responded to by the multi-agent system corresponding to the target application.

[0252] Figure 6 is a schematic diagram illustrating the construction principles of a multi-agent system for a target application provided by an embodiment of the present disclosure. As shown in Figure 6 , the client may be an enterprise. In the first phase, the enterprise may train a group of agents with different functions, such as Agent 1, ..., Agent 1, ..., Agent N in the figure.

[0253] For example, in the first stage, enterprises can use SFT and other methods to create exclusive intelligent agents with specific functions, such as document question-and-answer agents, reception agents, code agents, recommendation marketing agents, agents focusing on math problems, etc., and provide each intelligent agent with descriptive information of the corresponding function.

[0254] In the second stage, the enterprise can create a target application and select multiple agents from the agent resource library. As shown in Figure 6, the enterprise selects agent i, agent j, and agent k to build an agent system corresponding to the target application.

[0255] Optionally, the entry agent and / or entry configuration information specified by the customer for the multi-agent system can also be obtained. The entry configuration information is used to indicate whether to transfer the entry agent's tag to a matching agent when the information to be processed does not match the function of the current entry agent.

[0256] In addition, the perception configuration information configured by the customer for the multi-agent system can also be obtained; wherein the perception configuration information is used to indicate whether to inform the user to transfer the information to be processed to the matching agent for processing.

[0257] For example, an enterprise can select a math problem agent, a reception agent, and a document question-and-answer agent to build a multi-agent system. They can then designate the reception agent as the entry agent, creating their own intelligent customer service application. Once deployed on the platform, this application can provide online intelligent customer service capabilities. The entry and perception configuration information can be set by default or adjusted at any time.

[0258] Figure 7 is a schematic diagram of an interactive interface provided by an embodiment of the present disclosure. As shown in Figure 7, the interactive interface can display a list of agents in the agent resource library, allowing the user to create and enter the name of a target application and select multiple agents for the target application from the list. It can also provide functions such as setting an entry agent.

[0259] Optionally, after constructing the multi-agent system corresponding to the target application, a notification message may be sent to each agent in the multi-agent system, wherein the notification message is used to instruct the multiple agents to forward description information of their corresponding functions to each other.

[0260] Specifically, in the third stage, after the multi-agent system is built, the platform can instruct multiple agents to forward description information to each other. Optionally, each agent can be allowed to send its ID (identification) and description information to other agents simultaneously through a multi-agent group chat.

[0261] For example, agent i can send its ID and description information to agents j and k. After receiving the information, agents j and k can store it in their own memory modules. Specifically, the memory module can include an information queue, in which the ID and description information of the agent can be stored.

[0262] Agent j and agent k can also send their own ID and description information to other agents respectively. After forwarding each other, each agent in the multi-agent system maintains its own ID and description information and that of each other agent.

[0263] In this way, multiple agents send their own IDs and corresponding function description information to each other in the form of traversal, so that multiple agents can inform each other of their own description information through natural language communication, thereby accurately and comprehensively obtaining the information of each agent in the multi-agent system, facilitating the subsequent accurate transfer of information to be processed.

[0264] In other optional implementations, after building the multi-agent system, the platform can also directly send the description information of each agent to other agents.

[0265] In the fourth stage, each agent can configure its own prompt, which, along with the information to be processed, is fed into the agent's information processing model for processing. Optionally, for any target agent, the prompt can include a system prompt and a memory component, one for displaying the target agent's description and one for displaying the description of other agents, respectively.

[0266] Optionally, in intelligent customer service scenarios, the information to be processed can include user input. When a user makes their first online request, their input is sent to a portal agent, which responds to the user's input. The portal agent can be an agent with better conversational terminology and user guidance.

[0267] In the case where there are multiple rounds of interaction with the user, optionally, the information to be processed in each round may include not only the user's input information in the current round, but also the interaction information of the previous round (the user's input information and the agent's response information in each previous round).

[0268] For example, in each round, the content input to the information processing model of the entry agent may include:

[0269] "System prompt: You are an agent with [functional description]. If you encounter a problem beyond your capabilities, please refer to the IDs and functional descriptions of the following agents. Based on the user's historical context, select an appropriate agent and output its corresponding ID. If no suitable agent can handle the task, output a corresponding answer based on the known information."

[0270] memory: Agent ID and its function description:

[0271] [Agent ID1: Functional description];

[0272] [Agent ID2: Functional description];

[0273] [Agent ID3: Functional description];

[0274] User history: [user history].

[0275] Among them, the description information of the target agent is inserted in [Function Description], the ID and description information of other agents are inserted in memory, and the interaction information of the previous round and the input information of the current round are inserted in [User History].

[0276] The information processing model can output the ID of the matching agent based on the above content, and the entry agent can send the user history to the matching agent. At the same time, the tag of the entry agent will be transferred to the matching agent.

[0277] If the information processing model does not output the agent's ID but directly outputs the corresponding reply information, the entry agent directly outputs the reply information to the user and continues to communicate with the user until the interaction is completed.

[0278] The method for constructing a multi-agent system provided in this embodiment allows customers to create an agent resource library with exclusive functions through their own training, and to construct their own target applications by combining agents. Multiple agents communicate with each other through natural language to inform each other of their own description information. During the use of the target application, when user input information comes in, the input information can flow between multiple agents, thereby utilizing the natural language understanding ability of the information processing model to allow multiple agents to understand the functions of other agents through dialogue, thereby improving the task processing effect in actual applications. In addition, multiple agents can be organized without retraining, without the need for additional computing overhead or gate selection, thus avoiding the design of the gate model. The natural language understanding ability of the information processing model is directly utilized to achieve rapid semantic understanding and calling, reduce system complexity, and improve the overall processing effect of the system.

[0279] In the above embodiments, the response to the information to be processed is achieved by configuring the entry agent. In other optional implementation methods, the operation of configuring the entry agent can also be canceled. After obtaining the information to be processed, the information to be processed is randomly sent to an agent, or the agent of the previous round is used to process the information to be processed.

[0280] On the basis of the technical solution provided in the above embodiment, optionally, the function and / or capability requirement information input by the customer for one or more agents in the agent resource library can be obtained; according to the function and / or capability requirement information of each agent, the corresponding capability enhancement method of each agent is determined, so that each agent is processed according to the corresponding capability enhancement method.

[0281] The capability enhancement method includes at least one of the following: fine-tuning the information processing model within the agent, reinforcement learning for a single agent, joint reinforcement learning with other agents, and configuring prompt information for the agent. Fine-tuning can include supervised fine-tuning, unsupervised fine-tuning, or other types of fine-tuning.

[0282] Optionally, capability requirement information can be used to express the client's requirements for the capabilities of the information processing model. If the requirements are high, targeted capability enhancement operations can be performed on the agent.

[0283] Specifically, the information processing model in the initial intelligent agent can be a pre-trained information processing model with certain natural language processing capabilities. On this basis, in order to improve the capabilities of the intelligent agent so that the intelligent agent can better realize the corresponding functions, the intelligent agent can be enhanced. The specific cases of capability enhancement are as follows:

[0284] In the first case, if the corresponding function of the intelligent agent requires strong specific capabilities, for example, code processing, the traditional information processing model pre-trained in natural language may be difficult to cope with. At this time, the pre-trained information processing model can be supervised fine-tuned, and supervised fine-tuning is achieved using code data.

[0285] Alternatively, although the function of the intelligent agent is also to process natural language, the customer has higher capability requirements for the intelligent agent. For example, the customer hopes to further improve the performance of the intelligent agent in processing natural language, or the customer can perform supervised fine-tuning of the information processing model based on specific tasks.

[0286] In the second case, if the agent's capabilities are relatively common or the required capabilities are relatively low, supervised fine-tuning can be avoided and capabilities can be enhanced directly by configuring descriptive information. After configuring the descriptive information, the descriptive information can be input as a prompt into the information processing model, allowing the information processing model to execute the corresponding function based on the descriptive information.

[0287] In the third case, if we want the intelligent agent to enhance its ability to interact with the environment, we can train the intelligent agent through reinforcement learning and other methods. Reinforcement learning can include reinforcement learning for a single intelligent agent or reinforcement learning between multiple intelligent agents. The choice can be made according to actual needs.

[0288] In the fourth case, the above-mentioned capability enhancement methods can be combined. For example, descriptive information can be configured for the intelligent agent, the information processing model in the intelligent agent can be fine-tuned in a supervised manner, and the intelligent agent can be further trained in combination with reinforcement learning.

[0289] Through the above method, the ability enhancement method of the intelligent agent can be determined according to actual needs. The pre-trained intelligent agent can be used directly, or it can be configured with prompt information or fine-tuned before use. It can also be used after training with reinforcement learning to meet the functional and ability requirements of the intelligent agent in different scenarios and improve the overall processing effect of the multi-agent system in different scenarios.

[0290] The present disclosure also provides an information processing method based on a multi-agent system. The method is applied to the online use phase of the multi-agent system. The execution subject may be an application platform, more specifically, a control module in the application platform. The method includes:

[0291] Obtaining information to be processed corresponding to the multi-agent system;

[0292] determining an entry agent in the multi-agent system;

[0293] The information to be processed is sent to the entry agent so that when the information to be processed matches the function of the entry agent, the entry agent generates corresponding reply information. When the information to be processed does not match the function of the entry agent, the agent in the multi-agent system whose function matches the information to be processed is determined based on the information processing model, and the information to be processed is sent to the matching agent for processing.

[0294] The entry agent can be a target agent that processes the information to be processed. The information to be processed can include user input. After receiving the response information output by the entry agent, the response information can be fed back to the user. If the information to be processed is transferred to another agent for processing, the response information output by the other agent can be fed back to the user after receiving the response information output by the other agent.

[0295] The specific implementation principles, processes and beneficial effects of the method provided in this embodiment can be found in the aforementioned embodiments and will not be repeated here.

[0296] The disclosed embodiments also provide another information processing method based on a multi-agent system, applied to an intelligent customer service scenario, wherein the multi-agent system includes multiple agents, and the method is applied to a target agent among the multiple agents. Optionally, the target agent may include an information processing model.

[0297] FIG8 is a flow chart of another information processing method based on a multi-agent system provided by an embodiment of the present disclosure. As shown in FIG8 , the method includes:

[0298] Step 801: Obtain user input information.

[0299] Step 802: If the input information matches the function corresponding to the target agent, corresponding reply information is generated through the information processing model.

[0300] Specifically, if the information to be processed matches the function corresponding to the target agent, corresponding reply information is directly generated and fed back to the user.

[0301] Step 803: If the input information does not match the function corresponding to the target agent, the input information is sent to an agent whose function matches the input information for processing; wherein the matching agent is determined by the information processing model.

[0302] The specific implementation principles and processes of the method provided in this embodiment can be found in the aforementioned embodiments and will not be repeated here.

[0303] This embodiment can utilize a multi-agent system to interact with users. During the interaction process, the target agent can transfer problems that it cannot solve to an agent that is good at solving the problem, allowing the user's questions to flow between multiple agents, making it easier to quickly and accurately select the appropriate agent to answer the user's questions, thereby improving the effect of interaction with the user.

[0304] Optionally, multiple agents in the multi-agent system and users can interact through group chat, and the input information and the reply information are used to be displayed in the group chat interface.

[0305] Optionally, sending the input information to an agent whose function matches the input information for processing includes:

[0306] Acquiring perception configuration information, where the perception configuration information is used to indicate whether to inform the user to transfer the information to be processed to a matching agent for processing;

[0307] If so, the user is notified of the forwarding behavior through the group chat interface, and the input information is forwarded to the matching agent for processing, so that the matching agent generates corresponding reply information and displays it to the user through the group chat interface.

[0308] Specifically, in the group chat interface, after the user enters a question, the target intelligent agent can answer the user's question. When the target intelligent agent cannot answer the user's question, a transfer operation is performed. If the perception configuration information indicates that the transfer behavior needs to be informed to the user, the transfer behavior will be notified to the user first, and then the question will be transferred to the matching intelligent agent for processing. The matching intelligent agent will answer the user's question in the group chat interface.

[0309] Exemplarily, the multi-agent system is a system corresponding to an intelligent customer service application configured for a certain car company customer. The intelligent customer service application can be used to answer vehicle-related questions for users. The corresponding multi-agent system may include the following agents: Xiaoche (document question-and-answer agent), image processing assistant (image agent), video processing assistant (video agent), etc. Among them, Xiaoche is selected by the enterprise as the default entry agent. Figure 9 is a schematic diagram of a group chat interface provided by an embodiment of the present disclosure. As shown in Figure 9, the interaction process between the multi-agent system and the user can refer to the following example:

[0310] User: Hello.

[0311] Xiaoche: Hello, I am Xiaoche, the smart assistant of a certain car company. I can answer questions related to vehicles.

[0312] User: What are the advantages of a certain car model?

[0313] Small car: A certain car model has the following advantages: ….

[0314] User: Generate a video of a certain car model in a certain scenario.

[0315] Xiaoche: Sorry, I’m not good at this. The video processing assistant can handle it. I’ll send this requirement to it for processing.

[0316] Video Processing Assistant: Video. "

[0317] Through the above method, the transfer behavior between intelligent agents can be notified to the user, thereby improving the user experience. Moreover, whether the transfer behavior is notified can be configured according to actual needs, which can meet the customer's personalized configuration needs in different scenarios.

[0318] The disclosed embodiments also provide an information processing method based on a multi-agent system, which is applied to scenarios where agents autonomously plan task processes. The multi-agent system includes multiple agents, and the method is applied to a target agent among the multiple agents. The method includes:

[0319] Get pending task information;

[0320] Splitting the task information based on an information processing model to obtain multiple subtasks;

[0321] For any subtask, if the subtask does not match the function corresponding to the target agent, the subtask is sent to an agent whose function matches the subtask for processing; wherein the matching agent is determined by the information processing model.

[0322] Among them, the information processing model can plan the tasks to be processed by the intelligent agent, split them into multiple subtasks, and execute them in sequence according to the execution order of the subtasks. After executing multiple subtasks, the processing of the task is completed.

[0323] When executing each subtask, if the subtask matches the function corresponding to the target intelligent agent, the corresponding subtask processing result is directly generated using the information processing model. If it does not match, the information processing model can be used to determine the matching intelligent agent and transfer it to the matching intelligent agent for processing.

[0324] For example, the task information is: according to the demand, produce a bilingual movie in Chinese and English, and design a script for scheduled broadcasting.

[0325] The target agent can plan the above tasks. For example, it can be divided into the following subtasks: 1. Generate a script according to the requirements; 2. Translate the script; 3. Generate multiple pictures based on the script; 4. Make a movie based on the pictures; 5. Write a code script for scheduled playback.

[0326] Among them, the target intelligent agent is good at planning and natural language processing. Therefore, after planning the task, it can execute the first subtask and generate a script according to the needs. After the first subtask is completed, the target intelligent agent will determine that it is not good at translation. At this time, the script can be sent to the translation intelligent agent and the translation intelligent agent can be instructed to translate the script. The processing process of other subtasks is similar. In each step, the appropriate intelligent agent can be selected for the current subtask based on the description information of each intelligent agent until the entire task is completed.

[0327] Optionally, the target agent may be an entry agent, and the entry agent may remain fixed during the task execution process.

[0328] Optionally, the task information may be a piece of text or in the form of a flowchart. The multi-agent system may execute multiple nodes in the flowchart in sequence to complete the corresponding task.

[0329] The implementation principles of each step in this embodiment can refer to the aforementioned embodiment. It is only necessary to replace the information to be processed with a subtask. The specific processing process and beneficial effects can be referred to the aforementioned embodiment and will not be repeated here.

[0330] The embodiment of the present disclosure further provides a multi-agent system, comprising a plurality of agents, wherein the agents are used to execute the methods described in the relevant embodiments of FIG. 2 and FIG. 8 .

[0331] The disclosed embodiment further provides a multi-agent based application platform, including a control module and an agent resource library; the control module is used to execute the method described in the related embodiment of FIG. 5 .

[0332] The present disclosure also provides an information processing method that can be applied to a system implemented based on multiple information processing models. The method includes:

[0333] Get information to be processed;

[0334] Inputting the information to be processed and description information of the functions corresponding to the multiple information processing models into a target information processing model, so that the target information processing model outputs corresponding reply information when the information to be processed matches the function of the target information processing model, and determines the information processing model that matches the information to be processed among the multiple information processing models when there is no match; wherein the target information processing model is any information processing model among the multiple information processing models;

[0335] If the output of the target information processing model is used to represent the matched information processing model, the information to be processed is sent to the matched information processing model for processing.

[0336] The difference between this embodiment and the aforementioned embodiment is that this embodiment does not require the construction of an intelligent agent, but directly utilizes multiple information processing models to process the information to be processed. Although the capabilities of the information processing model may not be as comprehensive as those of the intelligent agent, it can also complete the processing of some tasks, especially tasks that are mainly based on natural language.

[0337] The implementation principles of each step in this embodiment can refer to the aforementioned embodiment. It is only necessary to replace the intelligent agent with an information processing model. The specific processing process and beneficial effects can be referred to the aforementioned embodiment and will not be repeated here.

[0338] Corresponding to the above method, an embodiment of the present disclosure further provides an information processing device based on a multi-agent system, wherein the multi-agent system includes a plurality of agents; the device is applied to a target agent, wherein the target agent is any one of the plurality of agents; the device includes:

[0339] A first determining module, configured to determine the to-be-processed information corresponding to the target agent;

[0340] a second determining module configured to determine, based on the information processing model, an agent whose function matches the information to be processed when the information to be processed does not match the function corresponding to the target agent;

[0341] The sending module is used to send the information to be processed to the matching agent for processing.

[0342] The present disclosure also provides a device for constructing a multi-agent system, including:

[0343] An output module, configured to output a list of agents contained in an agent resource library, wherein the agent resource library includes at least two trained agents;

[0344] An acquisition module, configured to acquire a plurality of agents selected by a client for a target application from the list;

[0345] A construction module is used to construct a multi-agent system corresponding to the target application based on the agent selected by the customer.

[0346] The present disclosure also provides an information processing device based on a multi-agent system, comprising:

[0347] An acquisition module, configured to acquire information to be processed corresponding to the multi-agent system;

[0348] A determination module, configured to determine an entry agent in the multi-agent system;

[0349] A sending module is used to send the information to be processed to the entry agent, so that the entry agent generates corresponding reply information when the information to be processed matches the function of the entry agent. When the information to be processed does not match the function of the entry agent, the agent in the multi-agent system whose function matches the information to be processed is determined based on the information processing model, and the information to be processed is sent to the matching agent for processing.

[0350] The present disclosure also provides an information processing device based on a multi-agent system, wherein the multi-agent system includes a plurality of agents; the device is applied to a target agent among the plurality of agents; the device includes:

[0351] The acquisition module is used to obtain user input information;

[0352] A generation module, configured to generate corresponding reply information through an information processing model when the input information matches the function corresponding to the target agent;

[0353] The sending module is used to send the input information to an agent whose function matches the input information for processing when the input information does not match the function corresponding to the target agent; wherein the matching agent is determined by the information processing model.

[0354] The present disclosure also provides an information processing device, wherein the method includes:

[0355] An acquisition module is used to obtain information to be processed;

[0356] a determination module, configured to input the information to be processed and description information of functions corresponding to the plurality of information processing models into a target information processing model, so that the target information processing model outputs corresponding reply information when the information to be processed matches the functions of the target information processing model, and to determine an information processing model among the plurality of information processing models that matches the information to be processed when there is no match; wherein the target information processing model is any one of the plurality of information processing models;

[0357] The sending module is configured to send the information to be processed to the matching information processing model for processing when the output of the target information processing model is used to represent the matching information processing model.

[0358] The specific implementation principles and effects of the device provided by the embodiments of the present disclosure can be found in the aforementioned embodiments and will not be repeated here.

[0359] FIG10 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present disclosure. As shown in FIG10 , the electronic device of this embodiment may include:

[0360] At least one processor 1001; and a memory 1002 communicatively connected to the at least one processor; wherein the memory 1002 stores instructions executable by the at least one processor 1001, and the instructions are executed by the at least one processor 1001 to cause the electronic device to perform the method as described in any of the above embodiments. Optionally, the memory 1002 can be independent or integrated with the processor 1001.

[0361] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the aforementioned embodiments and will not be described in detail here.

[0362] An embodiment of the present disclosure further provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When a processor executes the computer-executable instructions, the method described in any of the above embodiments is implemented.

[0363] An embodiment of the present disclosure further provides a computer program product, including a computer program, which implements the method described in any of the aforementioned embodiments when executed by a processor.

[0364] In the several embodiments provided in this disclosure, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the module division is merely a logical function division. In actual implementation, other division methods may be used. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not implemented.

[0365] The integrated modules implemented in the form of software function modules can be stored in a computer-readable storage medium. The software function modules stored in a storage medium include several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute some of the steps of the methods described in various embodiments of the present disclosure.

[0366] It should be understood that the above-mentioned processor can be a processing unit (Central Processing Unit, referred to as CPU), or it can be other general-purpose processors, digital signal processors (Digital Signal Processor, referred to as DSP), application-specific integrated circuits (Application Specific Integrated Circuit, referred to as ASIC), etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the application can be directly embodied as being executed by a hardware processor, or can be executed by a combination of hardware and software modules in the processor. The memory may include high-speed RAM memory, and may also include non-volatile storage NVM, such as at least one disk memory, and can also be a USB flash drive, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk, etc.

[0367] The storage medium may be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random-access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium may be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0368] An exemplary storage medium is coupled to a processor, such that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and storage medium can be located in an application-specific integrated circuit. Of course, the processor and storage medium can also exist as discrete components in an electronic device or a host control device.

[0369] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0370] The serial numbers of the above-mentioned embodiments of the present disclosure are for description only and do not represent the advantages or disadvantages of the embodiments.

[0371] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present disclosure.

[0372] The above are only preferred embodiments of the present disclosure and are not intended to limit the patent scope of the present disclosure. Any equivalent structure or equivalent process transformation made using the contents of the present disclosure and the drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present disclosure.

Claims

1. An information processing method based on a multi-agent system, the multi-agent system including a plurality of agents; The method is applied to a target agent, where the target agent is any one of the multiple agents; the method includes: Determine the information to be processed corresponding to the target agent; Based on the information processing model, when the information to be processed does not match the function corresponding to the target agent, determine an agent whose function matches the information to be processed; Send the information to be processed to the matching agent for processing.

2. The method according to claim 1, wherein The information processing model is used to: generate corresponding reply information when the information to be processed matches the function corresponding to the target agent; output the identifier of the agent whose function matches the information to be processed when the information to be processed does not match the function corresponding to the target agent.

3. The method according to claim 1 or 2, wherein Based on the information processing model, when the information to be processed does not match the function corresponding to the target agent, determining an agent whose function matches the information to be processed includes: Input the information to be processed and the description information of the functions corresponding to each agent among the multiple agents into the information processing model, so that when the information to be processed does not match the function corresponding to the target agent, the information processing model determines an agent whose function matches the information to be processed.

4. The method according to claim 3, wherein, Inputting the information to be processed and the description information of the functions corresponding to each agent among the multiple agents into the information processing model includes: Input the information to be processed, the description information of the function corresponding to the target agent, and the first prompt information into the information processing model. The first prompt information is used to prompt the information processing model to generate corresponding reply information when it determines that the information to be processed matches the function corresponding to the target agent according to the information to be processed and the description information, and output the preset information indicating that it cannot reply when they do not match. If the output of the information processing model is the preset information, then input the information to be processed, the description information of the functions corresponding to the remaining agents, and the second prompt information into the information processing model, where the remaining agents are the agents other than the target agent among the multiple agents, and the second prompt information is used to prompt the information processing model to determine an agent whose function matches the information to be processed according to the information to be processed and the description information corresponding to the remaining agents.

5. The method according to claim 3, wherein, Inputting the information to be processed and the description information of the functions corresponding to each agent among the multiple agents into the information processing model includes: Input the information to be processed, the description information of the functions corresponding to each agent among the multiple agents, and the third prompt information into the information processing model; Wherein, the third prompt information is used to prompt the information processing model to determine whether the information to be processed matches the function corresponding to the target agent according to the information to be processed and the description information of the functions corresponding to each agent, and generate corresponding reply information when they match, and determine an agent whose function matches the information to be processed when they do not match.

6. The method according to any one of claims 3-5, wherein, The information input into the information processing model further includes: fourth prompt information; Among them, the fourth prompt information is used to prompt that when the information processing model determines that there is no agent in the multiple agents whose function matches the information to be processed, the information to be processed is replied according to the knowledge learned during the training process.

7. The method according to any one of claims 1-6, wherein, The information to be processed includes the input information of the user; The target agent is the agent marked as the entry agent in the multi-agent system; The information to be processed input into the multi-agent system is received and processed by the entry agent.

8. The method according to claim 7, wherein The multi-agent system is used to have at least one round of conversation with the user; in any round, the information to be processed input into the multi-agent system includes the input information of the user in this round; The sending the information to be processed to the matching agent for processing includes: Sending the information to be processed to the matching agent for processing, and transferring the mark of the entry agent to the matching agent, so that the matching agent receives and processes the information to be processed input into the multi-agent system in the next round; Among them, the entry agent corresponding to the first round is the agent pre-selected in the multi-agent system.

9. The method according to any one of claims 1-8, wherein The information to be processed includes the input information of the user; Sending the information to be processed to the matching agent for processing includes: Obtaining the perception configuration information, where the perception configuration information is used to indicate whether to inform the user that the information to be processed is transferred to the matching agent for processing; If so, notify the user of the transfer behavior and transfer the information to be processed to the matching agent for processing.

10. The method according to any one of claims 3-6, the method further includes: After the multi-agent system is constructed, sending the description information of the function corresponding to the target agent to the remaining agents in the multi-agent system, and receiving the description information of the corresponding function sent by the remaining agents; Storing the received description information in the memory module, so that when the information to be processed is obtained, the description information corresponding to each agent is read from the memory module and input into the information processing model for processing.

11. A method for constructing a multi-agent system, including: Outputting a list of agents included in the agent resource library, where the agent resource library includes at least two trained agents; Obtaining a plurality of agents selected by the customer from the list for the target application; Constructing a multi-agent system corresponding to the target application according to the agents selected by the customer; Among them, any agent in the multi-agent system is used to execute the method according to any one of claims 1-10.

12. The method according to claim 11, further includes: After constructing the multi-agent system corresponding to the target application, sending a notification message to each agent in the multi-agent system, where the notification message is used to instruct the multiple agents to forward the description information of their corresponding functions to each other.

13. The method according to claim 11 or 12, the method further includes at least one of the following: Obtain the agent trained or uploaded by the customer, and add the agent trained or uploaded by the customer to the agent repository; Obtain the description information added by the customer for one or more agents in the agent repository; Obtain the entry agent and / or entry configuration information specified by the customer for the multi-agent system, where the entry configuration information is used to indicate whether to transfer the label of the entry agent to a matching agent when the information to be processed does not match the function of the current entry agent; Obtain the perception configuration information configured by the customer for the multi-agent system; wherein, The perception configuration information is used to indicate whether to inform the user to transfer the information to be processed to a matching agent for processing.

14. The method according to claim 11 or 13, further comprising: Obtain the function and / or capability requirement information input by the customer for one or more agents in the agent repository; Determine the corresponding capability enhancement method for each agent according to the function and / or capability requirement information of each agent, so as to process each agent according to the corresponding capability enhancement method; Wherein, the capability enhancement method includes at least one of the following: fine-tuning the information processing model in the agent, performing reinforcement learning on a single agent, performing reinforcement learning jointly with other agents, and configuring hint information for the agent.

15. An information processing method based on a multi-agent system, comprising: Obtain the information to be processed corresponding to the multi-agent system; Determine the entry agent in the multi-agent system; Send the information to be processed to the entry agent, so that when the information to be processed matches the function of the entry agent, generate a corresponding reply message, and when the information to be processed does not match the function of the entry agent, determine the agent in the multi-agent system whose function matches the information to be processed based on the information processing model, and send the information to be processed to the matching agent for processing.

16. An information processing method based on a multi-agent system, the multi-agent system including a plurality of agents; The method is applied to the target agent among the multiple agents; the method includes: Obtain the input information of the user; If the input information matches the function corresponding to the target agent, generate a corresponding reply message through the information processing model; If the input information does not match the function corresponding to the target agent, send the input information to the agent whose function matches the input information for processing; wherein, the matching agent is determined by the information processing model.

17. The information processing method according to claim 16, wherein, The input information and the reply message are used to be displayed in the group chat interface; Sending the input information to the agent whose function matches the input information for processing includes: Obtain the perception configuration information, where the perception configuration information is used to indicate whether to inform the user to transfer the information to be processed to a matching agent for processing; If so, notify the user of the transfer behavior through the group chat interface, and transfer the input information to the matching agent for processing, so that the matching agent generates a corresponding reply message and displays it to the user through the group chat interface.

18. An information processing method, the method comprising: Obtain the information to be processed; Input the information to be processed and the description information of the functions corresponding to multiple information processing models into the target information processing model, so that when the information to be processed matches the function of the target information processing model, the corresponding reply information is output, and when there is no match, determine the information processing model among the multiple information processing models that matches the information to be processed; wherein, the target information processing model is any one of the multiple information processing models; If the output of the target information processing model is used to represent the matching information processing model, send the information to be processed to the matching information processing model for processing.

19. A multi-agent system, including multiple agents, where the agents are used to execute the method according to any one of claims 1-10, 16, and 17.

20. A multi-agent based application platform, including a control module and an agent resource library; The control module is used to execute the method according to any one of claims 11-15.

21. An electronic device, including: At least one processor; And A memory communicatively connected to the at least one processor; Wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the electronic device executes the method according to any one of claims 1-18.

22. A computer-readable storage medium, in which computer-executable instructions are stored, and when the processor executes the computer-executable instructions, the method according to any one of claims 1-18 is implemented.

23. A computer program product, including a computer program, and when the computer program is executed by the processor, the method according to any one of claims 1-18 is implemented.

Images