Multi-agent reply system based on shared core analysis indexes
By using a multi-agent response system that shares core analytical metrics, and sequentially traversing and integrating the execution results of agents, the problems of token surge and information interference in multi-agent response systems are solved, resulting in cost reduction and improved response accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-03-10
AI Technical Summary
Existing multi-agent response systems require the original question and the execution results of all historical steps as context input to a large model for each task planning or execution step. This leads to a surge in the number of tokens and information interference, affecting the accuracy of judgment and increasing costs.
A multi-agent response system based on shared core analysis metrics is adopted, including a task orchestration module, a multi-agent traversal module, and a context information integration module. By sequentially traversing the target agents and integrating the execution results, the number of input tokens is reduced, thus lowering operating costs.
While ensuring the accuracy of intelligent responses, it reduces operating costs, minimizes token surges and information interference, and improves system reliability and efficiency.
Smart Images
Figure CN121636665A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, and in particular to a multi-agent response system based on shared core analytical indicators. Background Technology
[0002] With the rapid development of large model technology, multi-agent system architecture has become an important paradigm for handling complex primal problems. In this architecture, a large model typically serves as the core decision-maker, responsible for task planning and decomposition of the primal problem, and scheduling multiple agents with specific professional capabilities to execute collaboratively.
[0003] However, existing multi-agent response systems require the original question and the results of all previous steps to be input into the large model for decision-making at each task planning or execution step. This results in a continuous accumulation of input tokens, creating significant cost pressure for large model services that rely on token-based billing. Secondly, the ever-increasing context can introduce redundant information, which not only interferes with the accuracy of subsequent judgments by the large model but also leads to a gradual deterioration in overall performance due to exceeding the effective context processing capacity of the large model, resulting in a reliability decay phenomenon of "initially accurate, later deviating". Summary of the Invention
[0004] The purpose of this application is to provide a multi-agent response system based on shared core analysis indicators, so as to reduce the number of input tokens required for each large model call, reduce operating costs, and improve the accuracy of intelligent responses. The specific technical solution is as follows: In a first aspect of the embodiments of this application, a multi-agent response system based on shared core analysis indicators is provided. The system includes: a task orchestration module, a multi-agent traversal module, a context information integration module, and a large model invocation module. The task orchestration module is used to obtain the original problem and determine the task execution sequence based on the original problem, wherein the task execution sequence contains M target agents, and M is a positive integer; The multi-agent traversal module is used to traverse the target agents in the task execution sequence in sequence, input the original question or the execution result corresponding to the previous target agent into the currently traversed target agent, and obtain the execution result corresponding to the currently traversed target agent. The large model invocation module is used to determine the response result based on the execution results of each of the M target agents.
[0005] In an optional implementation, determining the task execution sequence based on the original problem includes: Retrieve a list of pre-defined intelligent agents; Retrieve at least one task execution suggestion related to the original problem from a pre-defined information knowledge base; The task execution sequence is determined based on the original question, the at least one task execution suggestion, and the preset agent information list.
[0006] In an optional implementation, determining the task execution sequence based on the original problem, the at least one task execution suggestion, and the preset agent information list includes: Based on the original question, determine the target task execution suggestion from the at least one task execution suggestion; Based on the target task execution suggestion, obtain the description information corresponding to the M target intelligent agents from the preset intelligent agent information list; The original problem, the target task execution suggestions, and the description information corresponding to the M target agents are input into a pre-trained large language model to obtain the task execution sequence.
[0007] In an optional implementation, the step of inputting the execution result corresponding to the original problem or the previous target agent into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent includes: Based on the target agent currently being traversed, the original question or the execution result corresponding to the previous target agent is processed to obtain target information; The target information is input into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent.
[0008] In an optional implementation, the step of processing the original problem or the execution result corresponding to the previous target agent based on the currently traversed target agent to obtain target information includes: The original problem is updated based on the execution result of the previous agent; The updated original problem is split into N initial pieces of information, where N is a positive integer; Based on the target agent currently being traversed, K candidate information are selected from the N initial information, where K is a positive integer; The target information is obtained by integrating the K candidate information.
[0009] In an optional implementation, updating the original problem based on the execution result of the previous agent includes: Extract the execution result corresponding to the previous target agent to generate an execution summary; The original problem is updated based on the execution summary.
[0010] In an optional implementation, the step of selecting K candidate pieces of information from the N initial pieces of information based on the currently traversed target agent includes: For any of the initial information, determine the correlation score between the initial information and the currently traversed target agent; If the relevance score exceeds a preset threshold, the initial information is used as candidate information.
[0011] In an optional implementation, after obtaining the execution result corresponding to the currently traversed target agent, the method further includes: Based on the original question and the execution results of all target agents that have been traversed, determine whether the traversal termination condition is met. If the traversal termination condition is met, then the traversal of the target agents that have not yet been traversed in the task execution sequence is stopped, and the response result is determined based on all the execution results obtained so far.
[0012] In an optional implementation, the traversal termination condition is determined in the following manner: Obtain target processing strategies related to the execution results of the original problem and all target agents that have been traversed from a preset information knowledge base; Extract the condition description related to traversal termination from the target processing strategy, and use it as the traversal termination condition.
[0013] The technical solution provided in this application, a multi-agent response system based on shared core analysis indicators, includes: a task orchestration module, a multi-agent traversal module, a context information integration module, and a large model invocation module. The task orchestration module is used to obtain the original question and determine the task execution sequence based on the original question, wherein the task execution sequence contains M target agents, where M is a positive integer. The multi-agent traversal module is used to sequentially traverse the target agents in the task execution sequence. The context information integration module is used to input the original question or the execution result corresponding to the previous target agent into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent. The large model invocation module is used to determine the response result based on the execution results corresponding to each of the M target agents. By integrating context information into the module, the original question or the execution result of the previous target agent is used as the input to the currently traversed agent to obtain the execution result of this traversal. The large model calling module determines the response result based on the execution results of each of the M target agents. This can solve the problem of token surge and information interference caused by using the original question and the execution results of all historical steps as context in every task planning or execution step to input into the large model for decision-making. This reduces operating costs while ensuring the accuracy of intelligent responses. Attached Figure Description
[0014] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.
[0017] Figure 1 A schematic diagram of the structure of a multi-agent response system based on shared core analysis indicators provided in this application embodiment; Figure 2 A schematic diagram illustrating the implementation process of a multi-agent response method based on shared core analysis indicators, provided for an embodiment of this application; Figure 3A schematic diagram illustrating the implementation process of a method for determining a task execution sequence provided in this application embodiment; Figure 4 A schematic diagram illustrating the implementation process of another multi-agent response method based on shared core analysis indicators provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0019] The following disclosure provides numerous different embodiments or examples for implementing various structures of this application. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the scope of this application. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.
[0020] To address the technical problem of token overload and information interference caused by inputting the original question and the execution results of all historical steps as context into a large model for decision-making in each task planning or execution step in existing technologies, this application provides a multi-agent response system based on shared core analysis indicators, comprising: a task orchestration module, a multi-agent traversal module, a context information integration module, and a large model invocation module; the task orchestration module is used to obtain the original question and determine the task execution sequence based on the original question, wherein the task execution sequence contains M target agents, where M is a positive integer; the multi-agent traversal module is used to sequentially traverse the target agents in the task execution sequence; the context information integration module is used to input the original question or the execution result corresponding to the previous target agent into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent; the large model invocation module is used to determine the response result based on the execution results corresponding to each of the M target agents. By integrating contextual information, the module takes the execution result of the original question or the previous target agent as the input of the currently traversed agent to obtain the execution result of this traversal. The large model calling module determines the response result based on the execution results of each of the M target agents, which can reduce operating costs while ensuring the accuracy of intelligent responses.
[0021] Figure 1 A schematic diagram of the structure of a multi-agent response system based on shared core analysis indicators provided in this application embodiment is shown below. Figure 1 As shown, the multi-agent response system based on shared core analysis metrics includes: a task orchestration module, a multi-agent traversal module, a context information integration module, and a large model invocation module.
[0022] The task orchestration module is used to obtain the original problem and determine the task execution sequence based on the original problem. The task execution sequence contains M target agents, where M is a positive integer.
[0023] In this embodiment, after obtaining the original question, the task orchestration module determines a task execution sequence based on the original question. The task execution sequence contains M target agents, where M is a positive integer (e.g., M is 1, 2, 3, etc.). The original question refers to the user-inputted question or service request that triggers the intelligent response process. For example, a user reports "VIP members cannot watch 4K movies." The task execution sequence refers to the planned execution order of a series of agents to solve the original question, and may include structures such as serial, parallel, or conditional branching.
[0024] For example, if the original problem is "VIP members cannot watch 4K movies", based on this original problem, the serial sequence of the device capability query agent, the bitstream status check agent, and the playback error analysis agent is determined as the task execution sequence.
[0025] The multi-agent traversal module is used to sequentially traverse the target agents in the task execution sequence.
[0026] In this embodiment, the multi-agent traversal module sequentially traverses the target agents in the task execution sequence to invoke the corresponding target agents. Here, a target agent refers to an agent with specific responsibilities, capable of independently analyzing and solving a particular type of problem. For example, in video playback issues, this includes an "equipment capability query agent," a "playback error analysis agent," and a "stream status check agent," etc.
[0027] Specifically, each target agent is invoked sequentially according to the task execution sequence. For example, if the task execution sequence specifies that target agent 1, target agent 2, and target agent 3 are invoked in that order, then target agent 1, target agent 2, and target agent 3 are invoked in that order. Alternatively, if the task execution sequence specifies that target agent 1 and target agent 2 are invoked in parallel, and target agent 3 is invoked after target agent 1 and target agent 2 are invoked, then target agent 1 and target agent 2 are invoked simultaneously, followed by target agent 3.
[0028] The context information integration module is used to input the original problem or the execution result of the previous target agent into the currently traversed target agent, and obtain the execution result of the currently traversed target agent.
[0029] In this embodiment, the context information integration module inputs the original question or the execution result corresponding to the previous target agent into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent.
[0030] Specifically, if the target agent currently being called is the first target agent in the task execution sequence, then the original problem is input into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent; if the target agent currently being called is not the first target agent in the task execution sequence (i.e., other target agents have been called before the execution of the currently traversed target agent), then the execution result corresponding to the previous target agent is input into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent.
[0031] For example, in a task execution sequence, target agent 1, target agent 2, and target agent 3 are called sequentially. When target agent 1 is called, since it is the first target agent, the original question is input into target agent 1 to obtain its execution result. When target agent 2 is called, since it is not the first target agent in the task execution sequence (target agent 1 has been called before target agent 2), and target agent 1 is the preceding target agent, its execution result is input into target agent 2 to obtain its execution result. Similarly, when target agent 3 is called, since it is not the first target agent in the task execution sequence (target agent 1 and target agent 2 have been called before target agent 3), and target agent 2 is the preceding target agent, its execution result is input into target agent 3 to obtain its execution result.
[0032] The large model calling module is used to determine the response result based on the execution results of each of the M target agents.
[0033] In this embodiment, the large model invocation module is used to determine the response result based on the execution results of each of the M target agents. Specifically, the execution results of each of the M target agents can be input into the large model for aggregation to determine the response result. The response result may include root cause analysis, solution suggestions, and preventative measures, etc., which are not limited in this embodiment.
[0034] Based on the above description of the technical solution provided in the embodiments of this application, the multi-agent response system based on shared core analysis indicators includes: a task orchestration module, a multi-agent traversal module, a context information integration module, and a large model invocation module; the task orchestration module is used to obtain the original question and determine the task execution sequence based on the original question, wherein the task execution sequence contains M target agents, where M is a positive integer; the multi-agent traversal module is used to traverse the target agents in the task execution sequence in sequence; the context information integration module is used to input the original question or the execution result corresponding to the previous target agent into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent; the large model invocation module is used to determine the response result based on the execution results corresponding to each of the M target agents. By integrating context information into the module, the original question or the execution result of the previous target agent is used as the input to the currently traversed agent to obtain the execution result of this traversal. The large model calling module determines the response result based on the execution results of each of the M target agents. This can solve the problem of token surge and information interference caused by using the original question and the execution results of all historical steps as context in every task planning or execution step to input into the large model for decision-making. This reduces operating costs while ensuring the accuracy of intelligent responses.
[0035] Figure 2 This application provides a schematic diagram of the implementation process of a multi-agent response method based on shared core analysis indicators, applied to a multi-agent response system based on shared core analysis indicators. Figure 2 As shown, it can specifically include the following: S201, Obtain the original question.
[0036] In this embodiment, the original question is obtained. Specifically, user input can be received through a user interface (such as a web form, API interface, or voice interaction system), and the input content is standardized and preprocessed, including text cleaning, semantic parsing, and intent recognition, to form a structured original question. This embodiment does not limit this process.
[0037] S202, Obtain the list of preset intelligent agent information.
[0038] In this embodiment, a preset agent information list is obtained. This preset agent information list refers to a predefined directory of available agents, which may include information such as the unique identifier, responsibility description, capability scope, and invocation parameters of each agent. An agent registry can be loaded from the configuration center as the preset agent information table; this embodiment does not limit this approach.
[0039] S203, retrieve at least one task execution suggestion related to the original problem from a preset information knowledge base.
[0040] In this embodiment, at least one task execution suggestion related to the original problem is retrieved from a preset information knowledge base. The preset information knowledge base is a vector database storing historical intelligent response cases and task execution suggestions. The task execution suggestions are used to provide a validated problem-solving path reference for the current original problem.
[0041] Specifically, the original question can be encoded into an original question vector, and an approximate nearest neighbor search can be performed in a pre-defined information knowledge base to retrieve multiple similar response cases related to the original question vector, thereby obtaining at least one task execution suggestion.
[0042] S204. Based on the original problem, at least one task execution suggestion, and a pre-set list of agent information, determine the task execution sequence, wherein the task execution sequence contains M target agents, and M is a positive integer.
[0043] In this embodiment of the application, a task execution sequence is determined based on the original problem, at least one task execution suggestion, and a preset list of intelligent agents, wherein the task execution sequence contains M target intelligent agents, and M is a positive integer.
[0044] For details on how to determine the task execution sequence based on the original problem, at least one task execution suggestion, and a pre-defined list of agent information, please refer to [reference needed]. Figure 3 The method shown. Figure 3 This is a schematic diagram illustrating the implementation process of a method for determining a task execution sequence provided in an embodiment of this application, as shown below. Figure 3 As shown, the specific steps may include: S301, Based on the original problem, determine the target task execution suggestion from at least one task execution suggestion.
[0045] In this embodiment of the application, a target task execution suggestion is determined from at least one task execution suggestion based on the original problem, wherein the target task execution suggestion refers to the task execution suggestion that best matches the current original problem from the at least one task execution suggestion.
[0046] Specifically, the execution scenario corresponding to the original problem can be determined, and the target task execution suggestion can be determined from at least one task execution suggestion based on the execution scenario corresponding to the original problem.
[0047] For example, if the original problem is the inability to watch 4K videos, and the execution scenario corresponding to the original problem is that the 4K bitrate is not displayed, then the task execution suggestion is to first check the device configuration through agent 2. If the device has turned off 4K display, then the task ends; otherwise, use agent 3 to analyze the bitrate problem. If the execution scenario corresponding to the original problem is that the 4K bitrate is displayed, then use agent 1 to analyze the playback error.
[0048] S302, based on the target task execution suggestion, obtain the description information corresponding to M target intelligent agents from the preset intelligent agent information list.
[0049] In this embodiment, based on the target task execution suggestion, description information corresponding to M target agents is obtained from a preset agent information list. The description information refers to a complete capability description of the agent, which may include metadata such as responsibility description, input parameters, output format, and performance indicators. This information is used to provide accurate agent capability references for task planning, ensuring the executableness of the generated execution sequence; however, this embodiment does not limit this aspect.
[0050] Specifically, the system can analyze the agent types explicitly mentioned or implicitly required in the target task execution suggestions. From a pre-defined agent information list, it can extract complete descriptions of the M successfully matched target agents through precise name matching (directly searching by agent name), semantic similarity matching (calculating semantic similarity of responsibility descriptions when agent names are not identical), and functional requirement matching (reversely searching for agents with corresponding capabilities based on the functional requirements in the target task execution suggestions). This description information can include the agent's unique identifier, detailed description of its responsibilities, input parameter specifications and requirements, output result format and meaning, execution constraints, and prerequisite dependencies.
[0051] For example, if the target task execution suggestion is "first check the device configuration through agent 2, and then use agent 3 to analyze the bitstream problem", then the description information of "agent 2" and "agent 3" will be accurately searched from the preset agent information list.
[0052] S303: Input the original problem, the target task execution suggestions, and the description information corresponding to the M target agents into the pre-trained large language model to obtain the task execution sequence.
[0053] In this embodiment, the original problem, the target task execution suggestions, and the description information corresponding to M target agents are input into a pre-trained large language model to obtain a task execution sequence. The pre-trained large language model can be a pre-trained artificial intelligence model with natural language understanding and generation capabilities, such as the GPT series or ChatGLM, used for reasoning and analysis based on the input context information to generate a task execution sequence that conforms to business logic. This embodiment does not limit the specific model used in this application.
[0054] S205: Iterate through the target agents in the task execution sequence in order, and process the execution results of the original problem or the previous target agent based on the currently traversed target agent to obtain the target information.
[0055] In this embodiment, the target agents in the task execution sequence are traversed sequentially. Based on the target agent currently being traversed, the execution result corresponding to the original problem or the previous target agent is processed to obtain target information. The target information refers to the input content that has been filtered and refined and is specially customized for the current target agent. This ensures that each agent only receives information that is highly relevant to its responsibilities, avoids irrelevant information from interfering with the analysis and judgment, and significantly reduces the number of input tokens.
[0056] Specifically, for the target agent currently being traversed, processing the original problem or the execution result corresponding to the previous target agent to obtain target information may include the following steps: Step 1: Update the original problem based on the execution result of the previous agent.
[0057] In this embodiment of the application, if the target agent currently being traversed is not the first target agent, the original problem is updated based on the execution result of the previous agent.
[0058] Specifically, a summary can be extracted from the execution result corresponding to the previous target agent to generate an execution summary, and the original problem can be updated based on the execution summary.
[0059] For example, suppose the original problem is "VIP members cannot watch 4K movies". The task execution sequence is: device capability query agent, network stream diagnosis agent, and account and client analysis agent. When iterating to the first agent (device capability query agent), the input is to check if the playback device supports 4K decoding. The execution result of the device capability query agent is "The device hardware supports 4K decoding, but no 4K stream was detected". A summary is extracted from the execution result of the device capability query agent, resulting in "The device supports 4K but there is no stream". This updates the original problem to "The playback device has been confirmed to support 4K decoding. Please diagnose whether there are problems with the network link, CDN status, or the source material itself". If the execution summary of the second agent is "The network link is normal, but the source material has not been released in 4K in this region", based on this, the input generated for the third agent (account and client analysis agent) is "Both the device and network are normal. The root cause has been identified as regional copyright restrictions causing the lack of 4K source material. Please check whether the authorized region settings and client version of this VIP account conflict with this restriction."
[0060] In another embodiment of this application, if the target agent currently being traversed is the first target agent, then there is no need to update the original problem, and the step of splitting the original problem is directly executed, i.e., step 2.
[0061] Step 2: Split the updated original problem to obtain N initial pieces of information, where N is a positive integer.
[0062] In this embodiment of the application, the updated original problem is split into N initial information, where N is a positive integer (e.g., N is 1, 2, 3, etc.). The initial information refers to the basic information unit obtained through semantic segmentation, which can be an independent sentence or a phrase with complete semantics. This embodiment of the application does not limit this.
[0063] Specifically, semantic segmentation algorithms can be used to split the updated original question according to sentence boundaries or semantic integrity. Alternatively, rule-based segmentation (such as segmentation using punctuation marks like periods and semicolons) or deep learning-based segmentation methods can be employed. Simultaneously, the segmented information units are standardized by removing redundant whitespace and special characters, generating a unique identifier for each information unit to obtain N initial information units.
[0064] For example, segmentation can be performed based on punctuation marks and formatting characteristics such as periods, question marks, semicolons, and line breaks. The original problem: Device model is X1. This device does not support 4K decoding. User reports screen stuttering. This is broken down into three initial pieces of information: Device model is X1, This device does not support 4K decoding, and User reports screen stuttering.
[0065] Step 3: Based on the target agent currently being traversed, select K candidate information from N initial information, where K is a positive integer.
[0066] In this embodiment of the application, based on the target agent currently being traversed, K candidate information is selected from N initial information, where K is a positive integer (e.g., K is 1, 2, 3, etc.).
[0067] Specifically, for the target agent currently being traversed, selecting K candidate information from N initial information can include: for any initial information, determining the relevance score between the initial information and the target agent currently being traversed; if the relevance score exceeds a preset threshold, then the initial information is used as candidate information. The preset threshold is used to measure the degree of relevance between the initial information and the target agent currently being traversed, such as 0.5, 0.6, etc.
[0068] Specifically, the initial information can be filtered based on similarity calculation. Then, the reordering model is used to input the text of each initial information after filtering along with the current target agent's duty description (e.g., "query device hardware configuration and decoding capability") into the reordering model to obtain the relevance score between the initial information and the target agent being traversed, so as to determine the candidate information.
[0069] Step 4: Integrate the K candidate information to obtain the target information.
[0070] In this embodiment, K candidate information pieces are integrated to obtain target information. Specifically, the selected K candidate information pieces can be concatenated according to their original order of appearance or logical relationship to form preliminary text. Subsequently, this text is input into a large language model for integration to obtain the target information. This embodiment does not limit the specific steps involved.
[0071] S206, input the target information into the target agent currently being traversed, and obtain the execution result corresponding to the target agent currently being traversed.
[0072] In this embodiment of the application, the target information is input into the target agent that is currently being traversed, and the execution result corresponding to the target agent that is currently being traversed is obtained.
[0073] S207. Determine the response result based on the execution results of each of the M target agents.
[0074] In this embodiment of the application, this step is similar to step S103 above, and will not be described in detail here.
[0075] Figure 4 This is a schematic diagram illustrating the implementation process of another multi-agent response method based on shared core analysis metrics provided in this application embodiment, as follows: Figure 4 As shown, it can specifically include the following: S401, Obtain the original problem and determine the task execution sequence based on the original problem, wherein the task execution sequence contains M target agents, where M is a positive integer.
[0076] In this embodiment of the application, this step is similar to step S101 above, and will not be described in detail here.
[0077] S402, sequentially traverse the target agents in the task execution sequence, input the original problem or the execution result corresponding to the previous target agent into the currently traversed target agent, and obtain the execution result corresponding to the currently traversed target agent.
[0078] In this embodiment of the application, this step is similar to step S102 above, and will not be described in detail here.
[0079] S403, based on the original problem and the execution results of all target agents that have been traversed, determine whether the traversal termination condition is met.
[0080] In this embodiment, based on the original question and the execution results of all target agents that have been traversed, it is determined whether the traversal termination condition is met. The traversal termination condition refers to a set of predefined or dynamically generated logical rules used to determine whether the analysis results obtained so far have fully answered the original question, thereby deciding whether to end the agent traversal process in advance to avoid unnecessary computational overhead.
[0081] Specifically, the target processing strategy related to the execution results of the original problem and all target agents that have been traversed can be obtained from a preset information knowledge base, and the condition description related to the termination of traversal can be extracted from the target processing strategy as the termination condition of traversal.
[0082] First, the complete output of each executed agent can be summarized to obtain a concise result summary. Then, the result summary is concatenated with the original problem and vectorized to obtain a state representation. By obtaining the target processing strategy related to the state representation from the preset information knowledge base, and then extracting the condition description related to the traversal termination from the target processing strategy, it can be used as the traversal termination condition.
[0083] Specifically, the complete output of each executed agent is first summarized to obtain a concise result summary. This result summary is then concatenated with the original question and vectorized to form a state representation vector for the current context. Next, a similarity search is performed in a pre-defined information knowledge base (a vector database storing historical cases and strategies) to find several historical processing strategies that best match the current state representation. The preliminary search results are then finely ranked using a re-ranking model, and the strategy with the highest matching degree is selected as the target processing strategy. From the natural language description of the target processing strategy, a conditional description explicitly indicating the end of the task is extracted. Using a natural language understanding model or a predefined rule template, this description is transformed into structured, executable logical judgment conditions to achieve logical judgments on specific field values in the acquired execution results.
[0084] For example, the original problem is "Unable to watch 4K". The "Device Capability Query Agent" has been executed and the result {"supported":false} has been obtained. After concatenating and encoding "Unable to watch 4K" and "Device does not support 4K", the matching strategy retrieved from the preset information knowledge base is "If the device capability query result shows that the device does not support 4K, then the root cause of the problem is clear and there is no need to perform subsequent bitstream and playback error checks".
[0085] In one optional implementation, the complete output of each executed agent can first be summarized to obtain a concise result summary. This result summary is then concatenated with the original question and vectorized to form a state representation vector for the current context. This state representation vector is directly input into the large language model, requesting it to determine whether a sufficient conclusion to answer the original question has been obtained. Finally, based on the extracted rules or the model's direct judgment, the traversal termination condition is evaluated.
[0086] S404 If the traversal termination condition is met, stop traversing the target agents that have not yet been traversed in the task execution sequence, and determine the response result based on all the execution results obtained so far.
[0087] In this embodiment of the application, if the traversal termination condition is met, the traversal of the target intelligent agents that have not yet been traversed in the task execution sequence is stopped, and the response result is determined based on all the execution results obtained so far.
[0088] In another embodiment of this application, if the traversal termination condition is not met, an independent task replanning can be initiated. This planning uses the execution result of the most recent agent (or the filtered key parts) and the responsibility descriptions of the remaining available agents as core inputs to redetermine the task execution sequence.
[0089] S405, determine the response result based on the execution results of each of the M target agents.
[0090] In this embodiment of the application, this step is similar to step S103 above, and will not be described in detail here.
[0091] This application also provides an electronic device, such as... Figure 5 As shown, it includes a processor 501, a communication interface 502, a memory 503, and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504. Memory 503 is used to store computer programs; In one embodiment of this application, when the processor 501 executes a program stored in the memory 503, it performs the following steps: Obtain the original question and, based on the original question, determine the task execution sequence, where the task execution sequence contains M target agents, where M is a positive integer; traverse the target agents in the task execution sequence in order, inputting the original question or the execution result corresponding to the previous target agent into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent; determine the response result based on the execution results corresponding to each of the M target agents.
[0092] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not indicate that there is only one bus or one type of bus.
[0093] The communication interface is used for communication between the aforementioned electronic devices and other devices.
[0094] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.
[0095] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0096] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0097] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0098] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the protection scope of this application.
Claims
1. A multi-agent response system based on shared core analytics indicators, characterized in that, The system comprises a task arrangement module, a multi-agent traversal module, a context information integration module, and a large model calling module. The task arrangement module is configured to obtain an original question and determine a task execution sequence based on the original question, wherein the task execution sequence comprises M target agents, and M is a positive integer. The multi-agent traversal module is configured to traverse the target agents in the task execution sequence in sequence. The context information integration module is configured to input the original question or an execution result corresponding to a previous target agent into a currently traversed target agent to obtain an execution result corresponding to the currently traversed target agent. The large model calling module is configured to determine a reply result based on the execution results corresponding to the M target agents.
2. The system of claim 1, wherein, The task execution sequence is determined based on the original question, comprising: obtaining a preset agent information list; retrieving at least one task execution suggestion related to the original question from a preset information knowledge base; determining the task execution sequence based on the original question, the at least one task execution suggestion, and the preset agent information list.
3. The system of claim 2, wherein, The task execution sequence is determined based on the original question, the at least one task execution suggestion, and the preset agent information list, comprising: determining a target task execution suggestion from the at least one task execution suggestion based on the original question; obtaining description information corresponding to the M target agents from the preset agent information list based on the target task execution suggestion; inputting the original question, the target task execution suggestion, and the description information corresponding to the M target agents into a pre-trained large language model to obtain the task execution sequence.
4. The system of claim 1, wherein, The original question or the execution result corresponding to the previous target agent is input into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent, comprising: processing the original question or the execution result corresponding to the previous target agent based on the currently traversed target agent to obtain target information; inputting the target information into the currently traversed target agent to obtain the execution result corresponding to the currently traversed target agent.
5. The system of claim 4, wherein, The original question or the execution result corresponding to the previous target agent is processed based on the currently traversed target agent to obtain target information, comprising: updating the original question based on the execution result corresponding to the previous agent; splitting the updated original question to obtain N initial information, and N is a positive integer; filtering K candidate information from the N initial information based on the currently traversed target agent, and K is a positive integer; integrating the K candidate information to obtain the target information.
6. The system of claim 5, wherein, The original question is updated based on the execution result corresponding to the previous agent, comprising: generating an execution summary by abstracting the execution result corresponding to the previous target agent; updating the original question based on the execution summary.
7. The system of claim 5, wherein, The K candidate information is filtered from the N initial information based on the target agent currently traversed. For any initial information, a relevance score of the initial information and the target agent currently traversed is determined. If the relevance score exceeds a preset threshold, the initial information is taken as the candidate information.
8. The system of claim 1, wherein, After the execution result corresponding to the target agent currently traversed is obtained, the method further includes: According to the original question and the execution results corresponding to all the target agents currently traversed, it is determined whether a traversal termination condition is met. If the traversal termination condition is met, the target agents in the task execution sequence that have not been traversed are stopped from being traversed, and the reply result is determined based on all the execution results currently obtained.
9. The system of claim 8, wherein, The traversal termination condition is determined by: From a preset information knowledge base, a target processing strategy related to the original question and the execution results corresponding to all the target agents currently traversed is obtained. From the target processing strategy, a condition description related to the traversal termination is extracted as the traversal termination condition.
10. A multi-agent answering method based on a shared core analysis indicator, characterized in that, The method is applied to the system of any one of claims 1-9, and the method includes: An original question is obtained, and a task execution sequence is determined based on the original question, wherein the task execution sequence includes M target agents, and the M is a positive integer. Target agents in the task execution sequence are sequentially traversed, the original question or an execution result corresponding to a previous target agent is input to a target agent currently traversed, and an execution result corresponding to the target agent currently traversed is obtained. A reply result is determined according to the execution results corresponding to the M target agents respectively.