Report generation method, system and equipment based on multiple agents and storage medium

Abstract

The embodiment of the invention discloses a report generation method, system and device based on multiple agents and a storage medium. The method comprises the steps of receiving a research task; configuring at least one agent corresponding to the research task; driving the intelligent agent to execute a research task; on-line real-time calibration is carried out on at least one configuration parameter of the intelligent agent according to a cooperation log and evidence pool content generated when the intelligent agent executes the research task, and the step of driving the intelligent agent to execute the research task is executed again; wherein the configuration parameters comprise at least one of the following parameters: a prompt template, a tool permission, an evidence weight and an output structure constraint; and generating and displaying a research report generated when the intelligent agent executes the research task. By means of the mode, the configuration parameters of the agents are calibrated online in real time, real-time self-adaptive adjustment and optimization of the multi-agent role parameters are achieved, and timely adjustment and deviation correction of the configuration parameters of the agents can be achieved.

Description

Technical Field

[0001] This invention relates to the field of AI collaborative decision support technology, and in particular to a report generation method, system, device and storage medium based on multi-agent systems. Background Technology

[0002] With the development of large-scale artificial intelligence modeling technology, tools for automatically generating research reports are constantly emerging. Existing technologies can be broadly categorized as follows: One type of system uses a single large model as its core, generating reports around a preset template. This lacks role division and collaboration mechanisms, making it difficult to cover multiple professional perspectives. Another type uses a static pipeline to connect different sub-tasks, but information exchange between sub-tasks is limited, making it impossible to dynamically adjust tasks and roles based on discussion progress. Still other systems introduce multiple agents, but the collaboration process is invisible to the user; the user only sees the final result, making it difficult to understand the conclusion formation path and to correct deviations in a timely manner. Furthermore, existing solutions cannot achieve real-time calibration of agent configuration parameters. Summary of the Invention

[0003] The purpose of this invention is to propose a report generation method, system, device and storage medium based on multiple agents, which aims to solve the problem that the configuration parameters of agents cannot be calibrated in real time in the prior art.

[0004] To address the aforementioned technical problems, a first aspect of this application provides a multi-agent-based report generation method. The method includes: receiving a research task; configuring at least one agent corresponding to the research task; driving the agent to execute the research task; performing online real-time calibration of at least one configuration parameter of the agent based on the collaboration log and evidence pool content generated by the agent executing the research task; and returning to the step of driving the agent to execute the research task. The configuration parameters include at least one of the following: prompt template, tool permissions, evidence weight, and output structure constraints. The method also generates and displays a research report generated by the agent executing the research task.

[0005] To address the aforementioned technical problems, a second aspect of this application provides a multi-agent-based report generation system, comprising: a receiving module for receiving research tasks; an agent configuration module for configuring at least one agent corresponding to the research task; an agent collaboration engine for driving the agent to execute the research task; a role configuration calibration module for online real-time calibration of at least one configuration parameter of the agent based on the collaboration logs and evidence pool content generated by the agent executing the research task; wherein the configuration parameters include at least one of the following: prompt template, tool permissions, evidence weight, and output structure constraints; and a report generation and display module for generating and displaying research reports generated by the agent executing the research task.

[0006] To address the aforementioned technical problems, a third aspect of this application provides a multi-agent-based report generation device, comprising a processor and a memory coupled to each other; the memory stores a computer program, and the processor executes the computer program to implement the steps of the method provided in the first aspect above.

[0007] To address the aforementioned technical problems, a fourth aspect of this application provides a computer-readable storage medium storing program data, which, when executed by a processor, implements the steps of the method provided in the first aspect above.

[0008] The embodiments of this invention offer the following advantages: Unlike existing technologies, this application configures at least one agent corresponding to a research task, drives the agent to execute the research task, and performs online real-time calibration of at least one configuration parameter of the agent based on the collaboration logs and evidence pool content generated during the task execution. This process returns to drive the agent to execute the research task. The configuration parameters include at least one of the following: prompt template, tool permissions, evidence weight, and output structure constraints. A research report generated by the agent during the task execution is then generated and displayed. This method enhances the depth and robustness of research conclusions through multi-agent dynamic collaboration in multiple rounds. Furthermore, by using the collaboration logs and evidence pool content generated during the task execution to perform online real-time calibration of the agent's configuration parameters, real-time adaptive optimization of the multi-agent role parameters is achieved. This allows for timely adjustment and correction of agent configuration parameters, ensuring the accuracy of research task execution and preventing non-compliant report structures or conflicting viewpoints. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] in: Figure 1 This is a flowchart illustrating an embodiment of the multi-agent report generation method of this application; Figure 2 This is a flowchart illustrating an embodiment of step S12 of this application; Figure 3 This is a flowchart illustrating another embodiment of the multi-agent report generation method of this application; Figure 4 This is a flowchart illustrating an embodiment of step S13 of this application; Figure 5 This is a flowchart illustrating another embodiment of the multi-agent report generation method of this application; Figure 6 This is a schematic block diagram of an embodiment of the multi-agent report generation system of this application; Figures 7-13 This is a schematic diagram of the visual interface operation of an embodiment of the multi-agent report generation method of this application; Figure 14 This is a schematic block diagram of an embodiment of the multi-agent report generation device of this application; Figure 15 This is a schematic block diagram illustrating the structure of an embodiment of a computer-readable storage medium according to this application. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] The terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0013] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0014] Please see Figure 1 , Figure 1 This is a schematic flowchart of an embodiment of the multi-agent report generation method of this application. It should be noted that if substantially the same result is obtained, this embodiment does not necessarily replace it with a similar method. Figure 1 The illustrated process sequence is limited. This embodiment of the multi-agent report generation method includes the following steps S11-S15: S11: Receive research assignments.

[0015] Users can input research tasks through the human-computer interaction interface. This step receives the research tasks input by users through the human-computer interaction interface.

[0016] Research tasks may include one or more of the following: research topic, research scope, and time frame.

[0017] S12: Configure at least one agent corresponding to the research task.

[0018] The configuration of an agent can include settings for at least one of the following agent fields: role name, job description, role backstory, prompt template parameters, tool permissions, and output structure constraints. Here, the role name refers to the agent's name, its virtual identity, such as "Senior Data Researcher," "Market Data Analyst," or "Product Manager"; the job description refers to the core requirements the agent's role must achieve in the research task, guiding all actions during the research process; the role backstory refers to the background settings assigned to the agent's role, including identity, professional skills, and service experience, defining its behavioral boundaries and professional tone; prompt template parameters are the key variables that the agent needs to input when performing tasks; tool permissions refer to the scope and operational permissions of the tools the agent can access; output structure constraints refer to the requirements for the format, length, and specifications of the agent's output; and evidence weight refers to the credibility weight of the evidence allocated to the agent's decisions.

[0019] In one embodiment, at least one agent may be initially configured by the user when inputting a research task through a human-computer interaction interface. For example, the user may initially input the role intent information of the agent to be configured. This step completes the configuration of the agent based on the initially configured role intent information and the research task. Please refer to [link to relevant documentation]. Figure 2 Step S12 includes the following steps S121~S122: S121: Receive role intent information input by the user.

[0020] Users can input role intent information in natural language. Role intent information may include tasks undertaken during the research process, goals to be achieved, etc.

[0021] S122: Based on the role intent information, generate at least one agent configuration. The agent configuration includes the configuration of at least one of the following parameters: role name, job description, background story, prompt template parameters, tool permissions, evidence weight, and output structure constraints.

[0022] The semantic analysis and classification of the role intent information are performed by combining the pre-set domain knowledge base to generate the agent configuration. The matching degree between the role target information in the agent configuration and the research task is not less than a pre-set threshold. The agent configuration includes at least one or more of the following information to complete the agent: role name, job description, background story, prompt template parameters, tool permissions, evidence weight, and output structure constraints.

[0023] In one embodiment, at least one agent can be configured based on a user's selection instructions from a set of preset agents. Specifically, multiple agents can be pre-set with their role names, goals, backstories, prompt template parameters, tool permissions, evidence weights, and output structure constraints. When a user needs to generate a report using the report generation method of this embodiment, they can input the research task into the human-computer interaction interface and select at least one agent from the pre-set agents to perform the task. This method allows users to select the desired agent from a set of preset agents to perform the research task, eliminating the hassle of manually configuring agents each time.

[0024] This embodiment automatically completes the intelligent agent by using role intent information, eliminating the need for users to manually fill in a large number of configuration parameters, reducing the difficulty for ordinary users to configure multi-agent teams, and improving practicality and convenience.

[0025] S13: Drive the intelligent agent to perform research tasks.

[0026] This step drives the agent to conduct multiple rounds of dialogue around the research task, including task delegation, question and answer, information sharing, conflict detection, conflict arbitration, and calling external real-time data sources through the data integration module.

[0027] In one embodiment, prior to this step, the method further includes: real-time dynamic task orchestration of the research task, defining the flow of the research task, dividing each agent into sub-tasks, and setting the execution order, dependencies, and termination conditions of each sub-task. The sub-tasks include discussion tasks and report generation tasks.

[0028] This step can drive agents to perform research tasks according to preset collaboration rules. These preset collaboration rules may include: actively inviting all agents to express their initial views in the early stages of the research task; relying on real-time dynamic task orchestration rules to drive multiple agents to perform progressive collaboration based on previous statements and a shared memory pool; and cross-validating key conclusions by at least two types of agents with different functions. When the user inputs the "follow-up question" command, a clarification subtask for the specified viewpoint is automatically generated and delegated to the matching agent for execution. The task change is pushed to the front-end visualization interface as an event stream and recorded in the collaboration log.

[0029] S14: Based on the collaboration logs and evidence pool content generated by the agent in performing the research task, determine whether to calibrate at least one configuration parameter of the agent.

[0030] At least one configuration parameter includes at least one of the following: prompt template parameters, tool permissions, evidence weights, and output structure constraints. The prompt template parameters refer to a pre-defined standardized instruction framework, containing fixed modules such as roles, tasks, constraints, and formats, which drives consistent agent responses by populating variables. Tool permissions refer to the scope and operational permissions of tools that the agent can invoke. Evidence weights refer to the credibility weights assigned to the evidence upon which the agent's decisions are based. Output structure constraints refer to the specifications for the format, modules, fields, and length of the agent's output.

[0031] Among them, the collaboration log refers to the collection of operation records, dialogue content and event streams generated by the agent in the process of performing research tasks; the evidence pool refers to the storage collection of data sources, evidence, and arbitration results used to support the agent's viewpoint.

[0032] Specifically, when each configured agent performs a research task, a collaboration log is generated synchronously. This step determines, based on the collaboration log and the evidence pool content, whether to calibrate at least one configuration parameter of the agent. If yes, proceed to step S15; otherwise, proceed to step S16.

[0033] S15: Perform online real-time calibration of at least one configuration parameter of the agent.

[0034] If necessary, this step calibrates the agent configuration parameters based on the collaboration logs and evidence pool content.

[0035] This step calibrates the agent's configuration parameters online, and the changes take effect immediately.

[0036] After adjusting the configuration parameters of the agent, return to step S13 to continue driving the agent to perform the research task.

[0037] In one embodiment, after configuring the intelligent agent, information such as the reason for calibration and the differences before and after calibration can be recorded as calibration events and pushed to the front-end visual interface for users to view and understand.

[0038] S16: Generate and display the research report generated by the agent performing the research task.

[0039] The system generates a report after each agent's discussion task concludes. It then summarizes and analyzes the structured records and external data generated during the collaboration process, producing a Markdown-formatted research report. This report is then converted to PDF, Word, and other formats and displayed interactively on a web platform. The PDF and Word reports, along with corresponding discussion records and event streams, are stored in a database for users to view, review, and reuse on the historical report page. In subsequent applications, users can select any report from the historical reports to view the text and replay the task execution process. They can also create a new task based on the report with a single click, using it as a template for a new research task. The report automatically loads the original agent combination, configuration parameters, execution constraints, and other key parameters, requiring only adjustments to the research topic, scope, and timeframe. This allows for the rapid reuse of successful research solutions, improving efficiency.

[0040] In one embodiment, the synchronous conclusion jointly confirmed by all agents participating in the research task and the alternative conclusion that is arbitrated as the second best are marked as the main conclusion paragraph and the supplementary explanation paragraph, respectively. The main reasons for the rejected viewpoint and the cited evidence number are given in the supplementary explanation paragraph.

[0041] This embodiment enhances the depth and robustness of research conclusions through multi-agent dynamic collaboration in multiple rounds. Furthermore, by using the collaboration logs and evidence pool content generated during the execution of research tasks, the configuration parameters of the agents are calibrated online in real time, enabling real-time adaptive optimization of the parameters of the multi-agent roles. This allows for timely adjustment and correction of agent configuration parameters, ensuring the accuracy of research task execution and preventing non-compliant report structures or conflicting viewpoints.

[0042] In one embodiment, please refer to Figure 3 After step S13, the following steps S31-S32 may also be included: S31: When it is detected that different agents generate conflicting opinions on the same issue, trigger the agents corresponding to the conflicting opinions to conduct multiple rounds of debate on the conflicting opinions and obtain the conflict resolution result; or, calculate the weight of the evidence corresponding to each conflicting opinion, select the opinion corresponding to the weight that meets the preset conditions as the adopted opinion, and obtain the conflict resolution result; or, trigger the agents corresponding to the conflicting opinions to supplement the evidence for the conflicting opinions and obtain the conflict resolution result.

[0043] During agent collaboration, this step detects whether there are conflicting viewpoints output by agents on the same problem. If different agents output conflicting viewpoints on the same problem, this step handles the conflicting viewpoints.

[0044] In this process, the agent that triggers the conflicting viewpoint engages in multiple rounds of debate on the conflicting viewpoint. Specifically, the agent that outputs the conflicting viewpoint further debates the conflicting viewpoint to eliminate the conflicting parts of the viewpoint and arrive at a more consistent viewpoint.

[0045] The process involves using an evidence evaluation algorithm to calculate the weight of the evidence corresponding to each viewpoint. Viewpoints corresponding to evidence with weights that meet preset conditions are selected as the adopted viewpoints, resulting in conflict resolution. The weights are calculated based on at least one of the following: the timeliness of the evidence, the credibility of the evidence's source, and the completeness of the evidence's argument. The timeliness of the evidence refers to the effective scope of the evidence supporting the viewpoint over time, matching the task's time requirements, or referring to the relative age of the evidence. The credibility of the evidence's source refers to the authority and reliability of the evidence acquisition channel. The completeness of the evidence's argument refers to whether the evidence chain logically supports the core viewpoint in a closed loop.

[0046] In this process, the agent corresponding to the conflicting viewpoint supplements the evidence for that viewpoint, thus obtaining the conflict resolution result. After supplementing the evidence, the evidence evaluation algorithm can be called to calculate the weight of the evidence corresponding to each viewpoint. The viewpoint corresponding to the evidence with the weight that meets the preset conditions is selected as the adopted viewpoint, thereby obtaining the conflict resolution result.

[0047] S32: Send the conflict resolution result to the front-end visualization interface in the form of an event message, and write the arbitration reasons, adopted viewpoints, rejected viewpoints and cited evidence numbers corresponding to the conflict resolution result into the evidence pool.

[0048] The conflict resolution outcome may include the presentation of evidence corresponding to the conflicting viewpoints, the further debate process regarding the conflicting viewpoints, the presentation of supplementary evidence for the conflicting viewpoints, the arbitration reasons, the accepted viewpoints, and the rejected viewpoints. This embodiment enables real-time arbitration of the collaborative process and visualizes the arbitration results.

[0049] Please see Figure 4 , Figure 4 This is a schematic flowchart of an embodiment of step S13 of this application. This embodiment includes the following steps: S131: Calculate calibration metrics based on the collaborative logs and evidence pool content generated by the agent performing research tasks.

[0050] The calibration metrics may include at least one of structural compliance, conflict frequency, and conflict severity. Structural compliance measures whether the output content meets the report structure and chapter requirements. Conflict frequency measures the number of conflicts between the output viewpoints of each agent or between the output viewpoint and the evidence. Conflict severity measures the severity of the conflict between the evidence upon which each agent's output viewpoint depends. For example, if low-weight evidence contradicts the core conclusion, the conflict severity is "low"; if core evidence contradicts supporting evidence, it may affect the accuracy of the conclusion, and the conflict severity is "moderate"; if core evidence contradicts the task instructions, it may directly prevent the research task from progressing, and the conflict severity is "severe".

[0051] The calibration indicators can be obtained using a preset algorithm. Specifically, the collaboration logs are collected and analyzed in real time, and the structural compliance of the output content of each agent is calculated according to a preset structural compliance algorithm, or the number and severity of conflicts between the output viewpoints of each agent are detected according to a preset conflict detection algorithm.

[0052] S132: When the calibration index meets the preset triggering conditions, perform online real-time calibration on at least one configuration parameter of the agent.

[0053] The calibration indicators are compared with preset trigger conditions, and the results are used to determine whether the calibration indicators meet the trigger conditions. Specifically, the trigger conditions are met when the structural compliance is lower than the preset compliance threshold. The trigger conditions are also met when the number of conflicting opinions exceeds the preset frequency and / or the severity of the conflicting opinions exceeds the preset severity threshold.

[0054] This embodiment calculates calibration metrics based on the collaborative logs and evidence pool content generated by the agent performing research tasks. The configuration parameters of the agent are dynamically calibrated in real time according to whether the calibration metrics meet the trigger conditions. This reduces conflicts in the execution of tasks by the calibrated agent, ensures structural compliance, and significantly improves the structural compliance of the generated report and the consistency of viewpoints across agents.

[0055] Please see Figure 5 , Figure 5 This is a schematic flowchart of another embodiment of the multi-agent report generation method of this application. It should be noted that if substantially the same result is obtained, this embodiment does not necessarily reflect that result. Figure 5 The illustrated process sequence is limited. This embodiment of the multi-agent report generation method includes the following steps S21-S29: S21: Receive research assignment.

[0056] The same as step S11 above, will not be repeated here.

[0057] S22: Configure at least one agent corresponding to the research task.

[0058] The same as step S12 above, so it will not be repeated here.

[0059] S23: Generate execution constraints for each agent to perform the research task.

[0060] Execution constraints include the subtasks corresponding to each agent, as well as the execution order, dependencies, and termination conditions of each subtask. Subtasks may include, for example, discussion tasks and report generation tasks.

[0061] S24: Drive the intelligent agent to perform research tasks.

[0062] The same applies to step S13, so it will not be repeated here.

[0063] S25: Push the entire process of the research task execution to the front-end visualization interface and receive intervention instructions input by the user during the process of pushing the entire process of the event flow.

[0064] In this embodiment, multi-party real-time communication is established based on the WebSocket protocol. The entire process of the intelligent agent collaboration process, including logs, dialogue messages, evidence increments, arbitration results, and other event streams, is pushed to the front-end visualization interface. Users can view the task execution process in real time on the front-end visualization interface and can input intervention commands at any time to modify the execution constraints and configuration parameters of each intelligent agent, so as to output a report that is more in line with the user's research objectives.

[0065] Among them, real-time communication supports the transmission of multiple types of messages, including text, structured data, and visual charts. The front-end visual interface displays messages one by one by intelligent agent roles in the form of team group chat, and marks the virtual identity of the intelligent agent from which the message comes. It also marks the preceding speech / evidence number cited by each message to form a connection link.

[0066] Intervention commands include at least one of the following: follow-up questioning, pausing execution, resuming execution, specifying a data source, rerunning a subtask, replacing an agent role, activating an agent role, suspending an agent role, adjusting evidence weights, and adjusting output structure constraints.

[0067] An event flow includes at least speaking events, evidence writing events, conflict detection events, arbitration events, calibration events, and user intervention events.

[0068] S26: Parse the intervention instructions as task orchestration constraints and / or role modification requests.

[0069] This step involves parsing the intervention instruction after receiving it, and obtaining task scheduling constraints and / or role modification requests.

[0070] S27: Adjust the execution constraints of the corresponding agent based on task orchestration constraints; and / or, calibrate the configuration parameters of the corresponding agent online based on role modification requests.

[0071] If the intervention command is parsed to obtain task orchestration constraints, the execution constraints of the corresponding agent are adjusted according to the execution constraint adjustment request; if the intervention command is parsed to obtain a role modification request, the configuration parameters of the corresponding agent are calibrated online according to the role modification request.

[0072] This step, after adjusting the execution constraints of the corresponding agent and calibrating the agent configuration parameters, outputs the execution constraint adjustment results and the agent configuration parameter adjustment event receipt, and the adjustments to the execution constraints and agent configuration parameters take effect immediately.

[0073] Return to step S24 to drive each agent to continue performing the research task.

[0074] S28: Based on the collaboration logs and evidence pool content generated by the agent in performing the research task, determine whether to calibrate at least one configuration parameter of the agent.

[0075] The same as step S14 described above will not be repeated here. If yes, proceed to step S29; otherwise, proceed to step S30. S29: Perform online real-time calibration of at least one configuration parameter of the agent.

[0076] The process is the same as step S16 described above, and will not be repeated here. After adjusting the agent's configuration parameters, return to step S24 to continue driving the agent to execute the research task.

[0077] S30: Generate and display the research report generated by the agent performing the research task.

[0078] The same as step S15 above, so it will not be repeated here.

[0079] This embodiment uses the WebSocket protocol to establish a multi-party real-time dialogue channel, pushing the entire process event stream of the intelligent agent collaboration process, including logs, dialogue messages, evidence increments, arbitration results, etc., to the front-end visualization interface. Users can view the task execution process in real time on the front-end visualization interface and can input intervention commands at any time to modify the execution constraints and configuration parameters of each intelligent agent. It supports users to intervene, ask follow-up questions, and dynamically adjust tasks during the collaboration process to output reports that better meet the user's expectations.

[0080] The embodiments described above in this application can realize automatic completion and online calibration of intelligent agent role configuration, visualized playback of the entire process of collaboration and arbitration, and immediate effect of user intervention in collaboration. They are applicable to fields such as e-commerce, finance, and consulting that require highly reliable reasoning and multi-person collaboration.

[0081] Please see Figure 6 , Figure 6 This is a schematic block diagram of an embodiment of the multi-agent report generation system 100 of this application. The multi-agent report generation system 100 includes: a receiving module 110, an agent configuration module 120, an agent collaboration engine 130, a report generation and display module 140, and a role configuration calibration module 150. The receiving module 110 receives research tasks; the agent configuration module 120 configures at least one agent corresponding to the research task; the agent collaboration engine 130 drives the agent to execute the research task; the role configuration calibration module 150 performs online real-time calibration of at least one configuration parameter of the agent based on the collaboration logs and evidence pool content generated by the agent executing the research task; wherein the configuration parameters include at least one of the following: prompt template, tool permissions, evidence weight, and output structure constraints; the report generation and display module 140 generates and displays the research report corresponding to the research task.

[0082] Among them, the intelligent agent collaboration engine 130 is based on shared memory and evidence pool to ensure that all speech content is connected to the previous speech, realizes associative collaboration and progressive reasoning, and automatically triggers arbitration in case of conflict.

[0083] In one embodiment, the role configuration calibration module 150 is further configured to calculate calibration indicators based on the collaborative logs and evidence pool content generated by the agent performing research tasks; and to perform online calibration of the agent's configuration parameters when the calibration indicators meet preset trigger conditions.

[0084] In one embodiment, the report generation system 100 further includes a task orchestration module and a real-time communication module. The task orchestration module generates execution constraints for each agent to perform research tasks; the execution constraints include the execution order, dependencies, and termination conditions of each subtask. The real-time communication module pushes the entire process event flow of the research task execution to the front-end visualization interface and receives intervention instructions input by the user during the push of the entire process event flow, parsing the intervention instructions into task orchestration constraints and / or role modification requests. The task orchestration module adjusts the execution constraints of the corresponding agents based on the task orchestration constraints; and / or, the role configuration calibration module 150 calibrates the configuration parameters of the corresponding agents online based on the role modification requests.

[0085] The task orchestration module is also used to uniformly save each speech, evidence and its citation relationship among multiple agents, providing retrieval and progressive context for subsequent speeches.

[0086] The real-time communication module supports the transmission of multiple message types, including text, structured data, and visual charts. The front-end visual interface displays messages one by one by role in a group chat-like format, and marks the virtual identity of the intelligent agent from which the message originates. It also marks the preceding statements / evidence numbers cited by each message to form a connection link.

[0087] In one embodiment, the agent configuration module 120 further includes an agent role AI completion unit, which is used to receive role intent information input by the user; and generate at least one agent configuration based on the research task and role intent information. The agent configuration includes at least one of the following: role name, job description, background story, prompt template parameters, tool permissions, evidence weight, and output structure constraints.

[0088] In one embodiment, the agent configuration module 120 further includes a role management unit, which is used to create, edit, delete and query agents.

[0089] In one embodiment, the agent configuration module 120 further includes a configuration storage unit for persistently storing agent role configurations in a structured form.

[0090] In one embodiment, the agent collaboration engine 130 further includes an arbitration unit. When it detects that different agents generate conflicting opinions on the same issue, the arbitration unit performs the following operations: triggering the agents corresponding to the conflicting opinions to conduct multiple rounds of debate on the conflicting opinions to obtain a conflict resolution result; or, calculating the weight of the evidence corresponding to each opinion, selecting the opinion corresponding to the evidence with the weight that meets the preset conditions as the adopted opinion, and obtaining a conflict resolution result; wherein, the weight is calculated based on at least one of the following: the timeliness of the evidence, the credibility of the source of the evidence, and the completeness of the argumentation of the evidence; or, triggering the agents corresponding to the conflicting opinions to supplement the evidence for the conflicting opinions, and obtaining a conflict resolution result; sending the conflict resolution result to the front end in the form of an event message, and writing the arbitration reasons, the adopted opinion, the rejected opinion, and the cited evidence number corresponding to the conflict resolution result into the evidence pool.

[0091] In one embodiment, the report generation system 100 may further include a data integration module for integrating multiple real-time external data sources and providing a unified data call interface to the agent collaboration engine 130. The real-time external data sources can provide real-time data to the agents.

[0092] The data integration module exposes an external data call interface to the agent collaboration engine 130 through a tool registration mechanism. When the agent calls real-time external data, it verifies the relevance of the returned data to the research topic of the current research task and issues a verification alarm when the relevance is lower than the threshold. The verification alarm is written as an event to the event stream and collaboration log.

[0093] In one embodiment, the report generation system 100 may further include a user management module, which is used to implement user registration, login, access control and team sharing settings.

[0094] In one embodiment, the report generation system 100 may further include a data persistence module, which is used to store user information, agent configuration parameters, task records, session messages and report content based on a relational database.

[0095] In one embodiment, the report generation system 100 may further include a visualization interface module. The visualization interface module is used to provide users with a web interface including a login and registration page, a multi-agent console, an agent management page, and a report display page. The interface displays the multi-agent collaboration process and research report in real time. The visualization interface displays at least the speaking succession relationship, evidence citation relationship, conflict detection and arbitration process, user intervention events, and differences before and after role configuration calibration in an event flow manner, and supports the replay of event sequences.

[0096] The report generation system 100 uses FastAPI or a similar framework to implement REST interfaces and WebSocket services on the backend, MySQL as the database, and React or other web frontend frameworks for the frontend.

[0097] Please see Figures 7-13 This is a schematic diagram of the visual interface operation of an embodiment of the multi-agent report generation method of this application. The operation of this embodiment is as follows: a. When a user accesses the system homepage in their browser, they first see... Figure 7 The welcome pop-up window shown allows you to choose "Log in" or "Register".

[0098] b. When selecting "Register", proceed to... Figure 8 On the interface, enter your username, optional email address, and password to complete the registration.

[0099] c. After successful registration, the user... Figure 9 Enter your username and password to log in. After successful login, you will be redirected to... Figure 10 The multi-agent console page is shown.

[0100] d. Users in Figure 10 Enter "How can a beginner succeed in cross-border e-commerce? Taking TikTok as an example" in the "Research Topic" input box.

[0101] e. in Figure 11On the intelligent agent role management page, the system has several pre-set intelligent agent role entries, such as "Senior Data Researcher", "Market Data Analyst", and "Product Manager". Users can select several intelligent agent roles, such as "Senior Data Researcher", "Market Data Analyst", "TikTok Operations Expert" and "Team Coordinator", and check options such as "Automatically export PDF and Word reports after execution" before clicking "Start Execution".

[0102] e. Users can also click the "+Add Character" button to enter. Figure 12 In the role editing interface shown, enter a brief intent in the "Role Name" input box, such as "TikTok Operations Expert," and then click the "AI Completion" button. The AI ​​completion unit for agent roles generates candidate configurations based on the current research topic and the intent, automatically filling in fields such as "Role Objective," "Role Background Story," and "Task Description," while simultaneously generating default prompt template parameters, tool permissions, evidence weights, and output structure constraints. Users can fine-tune these settings and then click "Save." The new role will then be added to the role list and can participate in subsequent tasks. Check options such as "Automatically export PDF and Word reports after execution" and click "Start Execution."

[0103] Before scheduling any target agent to generate a speech, the agent collaboration engine 130 retrieves previous speeches, evidence entries, and citation numbers associated with the current subtask under the same topic from the shared memory and evidence pool, using them as input for generation. The output automatically records the cited speech / evidence numbers for subsequent visualization and tracing. All agent speeches are not parallel or independent, but are generated progressively based on previous content, achieving "progressive succession" and conflict convergence in topic collaboration.

[0104] The task orchestration module creates new discussion tasks and notifies the agent collaboration engine 130. The agent collaboration engine 130 instantiates the corresponding agent and initiates a multi-round dialogue. During the dialogue: the team coordinator is responsible for breaking down the problem and assigning sub-tasks; the market data analyst uses the data integration module to call the OpenBB API to obtain relevant market data for TikTok; the senior data researcher proposes an analytical framework from a methodological perspective; and the TikTok operations expert provides practical suggestions on content strategy, campaign timing, etc.

[0105] In this scenario, if market data analysts and operations experts reach opposing conclusions regarding whether a high advertising budget is suitable for a novice, the dynamic arbitration unit marks these two viewpoints as conflicting, retrieves the data sources and timestamps cited by each, calculates the weight of evidence, and may request the two agents to supplement their arguments. The arbitration result and reasons are pushed to the front end via a real-time communication module, allowing users to... Figure 10 The arbitration process can be seen in the runtime log area.

[0106] During the collaboration process, the role configuration calibration module 150 continuously or through event-triggered calculations of structural compliance and conflict frequency: if the structural compliance is below the threshold (e.g., missing chapters in the report analyst output or format not conforming to the template), the output structural constraints and prompt template parameters for that role are automatically calibrated; if the conflict frequency continues to increase (e.g., mutually contradictory viewpoints appear repeatedly on the same sub-issue), the conservatism of the evidence weighting strategy is automatically increased, the requirements for evidence citation are strengthened, or the queue of roles participating in arbitration is adjusted; each calibration generates a calibration event, including the triggering reason, the target, and the differences before and after calibration, which are then visualized in the event flow of the front-end visualization interface.

[0107] Users can also input commands such as "follow up / rerun / specify data source / adjust weight / replace role" in the collaboration. The system will convert the commands into task orchestration constraints or role parameter update requests and make them effective immediately. The system will also notify the user of "effective / ineffective and reason" through a receipt event.

[0108] Specifically, when the discussion reaches a preset number of rounds or the user clicks the "Generate Report" button on the front end, the task orchestration module triggers the report generation task. The report generation unit in the report generation and display module 140 reads all discussion records, event streams, and arbitration results, integrates the verified data and viewpoints, and generates a Markdown format report according to a preset template. The report content is... Figure 13 The report display page is presented in rich text format, including sections such as executive summary, problem breakdown, data analysis, implementation path, and risk warnings. Users can scroll through and copy parts of the content on this page, and can click the export button to generate PDF or Word files with one click.

[0109] The system records the metadata (topic, time, participating roles, etc.) and storage path of each report in the reports table, saves the corresponding collaboration process in the messages table, and persists the event stream (including arbitration and calibration events) and reference links.

[0110] Users can select any report on the history report page to view the main text and process replay; based on this, users can create a new task based on the report with one click. The system automatically loads the original role combination and key parameters (including initial values ​​of calibration parameters and task arrangement constraints). Only the topic or time interval needs to be adjusted to quickly iterate new research.

[0111] Please see Figure 14 , Figure 14This is a schematic block diagram of an embodiment of the multi-agent-based report generation device of this application. The multi-agent-based report generation device 900 includes a processor 910 and a memory 920 coupled to each other. The memory 920 stores a computer program, and the processor 910 is used to execute the computer program to implement the multi-agent-based report generation method described in the above embodiments.

[0112] For a description of each step of the processing, please refer to the description of each step in the above embodiment of the multi-agent report generation method of this application, and it will not be repeated here.

[0113] The memory 920 can be used to store program data and modules. The processor 910 executes various functional applications and data processing by running the program data and modules stored in the memory 920. The memory 920 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as configuration parameter calibration function, agent configuration function, etc.), etc.; the data storage area may store data created based on the use of the multi-agent-based report generation device 900 (such as agent configuration parameters, cooperation logs, etc.). In addition, the memory 920 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 920 may also include a memory controller to provide the processor 910 with access to the memory 920.

[0114] In the various embodiments of this application, the disclosed methods, apparatuses, and devices can be implemented in other ways. For example, the embodiments of the multi-agent-based report generation device described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be indirect couplings or communication connections through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0115] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0116] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0117] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium.

[0118] In addition, the term “coupling” in this article refers to any means of connection, both direct and indirect.

[0119] See Figure 15 , Figure 15 This is a schematic block diagram of a computer-readable storage medium according to an embodiment of the present application. The computer-readable storage medium 700 stores program data 710. When the program data 710 is executed, it implements the steps of the above embodiments of the multi-agent-based report generation method.

[0120] For a description of each step of the processing, please refer to the description of each step in the above embodiment of the multi-agent report generation method of this application, and it will not be repeated here.

[0121] The computer-readable storage medium 700 can be any medium capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0122] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A report generation method based on multi-agent systems, characterized in that, The method includes: Receive research assignments; Configure at least one agent corresponding to the research task; Drive the intelligent agent to perform the research task; Based on the collaboration logs and evidence pool content generated by the agent in performing the research task, at least one configuration parameter of the agent is calibrated online in real time, and the steps for driving the agent to perform the research task are returned; wherein, the configuration parameter includes at least one of the following: prompt template, tool permissions, evidence weight, and output structure constraints; Generate and display the research report generated by the agent in performing the research task.

2. The method according to claim 1, characterized in that, The step of performing online real-time calibration of at least one configuration parameter of the agent based on the collaborative logs and evidence pool content generated by the agent in performing the research task includes: Calculate calibration metrics based on the collaborative logs and evidence pool content generated by the agent performing the research task; When the calibration index meets the preset triggering conditions, at least one of the configuration parameters of the agent is calibrated online in real time.

3. The method according to claim 2, characterized in that, The calibration metrics include at least one of compliance, conflict frequency, and conflict severity.

4. The method according to claim 1, characterized in that, Before driving the agent to perform the research task, the method further includes: Execution constraints are generated for each of the aforementioned agents to perform the research task; the execution constraints include the sub-tasks corresponding to each agent and the execution order, dependencies, and termination conditions of each sub-task; After driving the agent to perform the research task, the method further includes: The entire process of the research task execution is pushed to the front-end visualization interface, and intervention instructions input by the user are received during the process of pushing the entire process event flow. The intervention instructions are parsed into task scheduling constraints and / or role modification requests; Adjust the execution constraints of the corresponding agent based on the task orchestration constraints; and / or calibrate the configuration parameters of the corresponding agent online based on the role modification request.

5. The method according to claim 1, characterized in that, The configuration of at least one agent corresponding to the research task includes: Receive role intent information input by the user; Based on the role intent information, generate at least one agent configuration, which includes the configuration of at least one of the following parameters: role name, responsibility description, background story, prompt template parameters, tool permissions, evidence weight, and output structure constraints.

6. The method according to claim 1, characterized in that, The method further includes performing the following operation when conflicting viewpoints generated by different agents on the same issue are detected: The agent corresponding to the triggering conflicting viewpoint conducts multiple rounds of debate on the conflicting viewpoint to obtain a conflict resolution result; or... Calculate the weight of the evidence corresponding to each conflicting viewpoint, select the viewpoint corresponding to the evidence with the weight that meets the preset conditions as the adopted viewpoint, and obtain the conflict resolution result; wherein, the weight is calculated based on at least one of the following: the timeliness of the evidence, the credibility of the source of the evidence, and the completeness of the argument of the evidence; or, The agent corresponding to the conflicting viewpoint is triggered to supplement evidence for the conflicting viewpoint, and the conflict resolution result is obtained; The conflict resolution result is sent to the front end in the form of an event message, and the arbitration reasons, adopted viewpoints, rejected viewpoints, and cited evidence numbers corresponding to the conflict resolution result are written into the evidence pool.

7. A multi-agent-based report generation system, characterized in that, The system includes: The receiving module is used to receive research tasks. An agent configuration module is used to configure at least one agent corresponding to the research task; An agent collaboration engine is used to drive the agent to perform the research task; The role configuration calibration module is used to perform online real-time calibration of at least one configuration parameter of the agent based on the collaboration log and evidence pool content generated by the agent in performing the research task; wherein, the configuration parameter includes at least one of the following: prompt template, tool permissions, evidence weight, and output structure constraints; The report generation and display module is used to generate and display the research report generated by the intelligent agent in performing the research task.

8. The system according to claim 7, characterized in that, The role configuration calibration module is also used to calculate calibration indicators based on the collaboration logs and evidence pool content generated by the agent when performing the research task; and to perform online calibration of the agent's configuration parameters when the calibration indicators meet preset trigger conditions.

9. A report generation device based on multi-agent intelligence, characterized in that, The multi-agent-based report generation device includes a processor and a memory coupled to each other; the memory stores a computer program, and the processor executes the computer program to implement the steps of the method as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program data that, when executed by a processor, implements the steps of the method as described in any one of claims 1-6.

Images