Search generation method and device based on multi-agent cooperation, equipment and medium

Abstract

The application provides a search generation method, device and equipment based on multi-agent cooperation and a medium. The method comprises the following steps: receiving a user search demand and calling a planning agent to perform task planning to generate a tree-shaped task planning result; calling a decomposition agent to perform complex task fine-grained decomposition according to the tree-shaped task planning result to obtain corresponding planning tasks and the dependency relationship and parallel relationship between the planning tasks, calling multiple agents, and executing the corresponding planning tasks with the parallel relationship in parallel, and after the corresponding planning tasks are executed, calling an agent to execute the next planning task with the dependency relationship to obtain corresponding search results; calling a verification agent to perform multi-source information cross verification according to the corresponding search results, and generating a corresponding search report based on the verification. The application can ensure that the conclusion of the final report is based on solid and reliable facts, significantly improves the efficiency, and comprehensively guarantees the factual accuracy of the report.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a search generation method, apparatus, device, and medium based on multi-agent collaboration. Background Technology

[0002] As financial markets become increasingly complex and data volumes explode, the generation of financial research reports faces higher demands for professionalism and timeliness. Automating the generation of high-quality, compliant financial research reports has become a critical issue that the fintech sector urgently needs to address. To meet this challenge, artificial intelligence technologies, particularly Large Language Models (LLMs), have been introduced into the automated generation process of financial research reports. This aims to improve report generation efficiency, reduce labor costs, and ensure the professionalism and compliance of the content. Related technological solutions are gradually becoming a hot topic in industry research.

[0003] Currently, most serialization processing schemes employ a single language model (LLM) as the core processing unit, executing tasks step-by-step through a linear, pipelined (Prompt Chaining) prompting process. Specifically, after a user's query is input into the model, the model sequentially performs a series of steps, including information retrieval, content generation, and formatting. The entire process presents an end-to-end linear structure. This approach, with its simple architecture and ease of implementation, is capable of handling basic financial research report content generation tasks.

[0004] However, when dealing with complex financial research tasks, the aforementioned solutions suffer from limitations in context length. The computational complexity of the attention mechanism in the Transformer architecture increases quadratically with the input length, making it difficult for the model to retain all historical information when processing long texts. This hinders the maintenance of logical coherence in long-range dependencies and multi-step reasoning processes, leading to broken logical chains and information loss, resulting in linguistic illusions or reasoning errors. In financial research, a conclusion often requires traversing multiple data sources and analytical steps. If information in intermediate stages is omitted, the logic of the final report will inevitably be biased. Furthermore, although multimodal large language models have made significant progress in language understanding, generation, and reasoning, they may still accumulate errors and struggle with effective self-correction when handling specialized tasks requiring multiple independent and complex steps due to their inability to effectively manage and utilize long-term historical information. Summary of the Invention

[0005] This invention provides a search generation method, apparatus, device, and medium based on multi-agent collaboration to address the shortcomings of existing technologies where logical deviations in the final report often occur due to context length limitations. This ensures that the conclusions of the final report are based on solid and reliable facts, significantly improving efficiency while fully guaranteeing the factual accuracy of the report.

[0006] This invention provides a search generation method based on multi-agent collaboration, comprising: receiving a user's search request and invoking a planning agent to perform task planning, generating a tree-like task planning result; based on the tree-like task planning result, invoking a decomposition agent to perform fine-grained decomposition of complex tasks, obtaining corresponding planning tasks and the dependencies and parallel relationships between planning tasks; based on the planning tasks and the dependencies and parallel relationships between planning tasks, invoking multiple agents to execute corresponding planning tasks with parallel relationships in parallel, and after executing the corresponding planning tasks, invoking agents to execute the next planning task with dependencies, obtaining the corresponding search results; based on the corresponding search results, invoking a verification agent to perform multi-source information cross-validation, and generating a corresponding search report based on successful verification.

[0007] According to the present invention, a search generation method based on multi-agent collaboration provides that the search results include data fragments and the data sources of the data fragments. Based on the corresponding search results, a verification agent is invoked to perform multi-source information cross-validation, and based on successful verification, a corresponding search report is generated. The method includes: inputting the corresponding search results into the invoked verification agent to pair data fragments from different data sources, using the target data source or previously verified data fragments as premise sentences and other data fragments as hypothesis sentences to obtain corresponding pairing results; inputting the pairing results into a consistency recognition model to obtain the verification result output by the consistency recognition model; wherein the consistency recognition model is based on a pre-trained model, fine-tuned using financial domain corpus; and generating a search report based on the successful verification result and the corresponding search results.

[0008] According to a multi-agent collaborative search generation method provided by the present invention, the pairing results are input into a consistency recognition model to obtain the verification results output by the consistency recognition model. The method includes: inputting the pairing results into the consistency recognition model to determine the semantic consistency of the corresponding paired evidence fragments for each pairing result, and mapping this consistency to a preset three-dimensional judgment matrix to obtain a three-dimensional judgment matrix; wherein the preset three-dimensional judgment matrix includes support, contradiction, and relevance, where support characterizes the entailment degree of the premise sentence to the hypothesis sentence, contradiction characterizes the degree of contradiction between the premise sentence and the hypothesis sentence, and relevance characterizes the neutrality of the premise sentence and the hypothesis sentence; the consistency recognition model is trained based on historical pairing results and the corresponding actual matrices; according to the three-dimensional judgment matrix, if the relevance is greater than a first preset threshold, the verification result is determined to be verified successfully; or, if the relevance is less than or equal to the first preset threshold, and the contradiction is greater than a second preset threshold, the verification result is determined to be verified unsuccessfully; or, if the contradiction is less than or equal to the second preset threshold, and the support is less than a third preset threshold, the verification result is determined to be verified unsuccessfully; and if the support is greater than or equal to the third preset threshold, the verification result is determined to be verified successfully.

[0009] According to the multi-agent collaborative search generation method provided by the present invention, after calling the verification agent to perform multi-source information cross-validation based on the corresponding search results, the method further includes: based on the failure of verification, calling the reflection agent according to the corresponding pairing results to determine the semantic similarity of the corresponding pairing results, the difference between the keyword premise sentence and the hypothesis sentence, and the information release time difference, to determine the credibility of the data source, and to determine the planning task corresponding to the hypothesis sentence in the failed pairing results, as well as the retrieval keywords and retrieval timestamps corresponding to the planning task, thereby obtaining feature combinations; based on the feature combinations, using the Bayesian optimization algorithm, defining a multi-dimensional search space, and predicting the probability distribution of potential success rates based on any correction scheme in the search space, thereby obtaining probability distribution prediction results; based on the probability distribution prediction results, selecting the correction scheme with the highest predicted success rate as the correction instruction, and sending it to the decomposition agent.

[0010] According to the multi-agent collaborative search generation method provided by the present invention, after generating a corresponding search report based on successful verification, the method includes: calling a compliance check agent to perform structured detection and semantic detection based on the search report, obtaining structured detection results and semantic detection results; when it is determined that both the structured detection results and the semantic detection results are compliant, outputting the generated search report as the final search report; when it is determined that the structured detection results and / or the semantic detection results are non-compliant, performing a comprehensive evaluation based on the non-compliance of the structured detection results and / or the semantic detection results, determining a non-compliance severity score, and automatically injecting a disclaimer or risk disclosure when the non-compliance severity score is less than or equal to a preset score threshold.

[0011] According to a multi-agent collaborative search generation method provided by the present invention, based on a search report, a compliance inspection agent is invoked to perform structured detection and semantic detection to obtain structured detection results and semantic detection results. The method includes: invoking a compliance inspection agent based on the search report to use regular expressions to match a preset rule base, determine the illegal content in the search report, and obtain the corresponding structured detection results; wherein the preset rule base is constructed based on the rules of different financial regulatory agencies; and performing semantic analysis on the search report to analyze the tone and wording in the search report, and combining this with the preset rule base to determine whether there is any illegal behavior, thereby obtaining the corresponding semantic detection results.

[0012] According to the present invention, a search generation method based on multi-agent cooperation is provided. Based on the tree-structured task planning results, a decomposition agent is invoked to perform fine-grained decomposition of complex tasks, obtaining the planned tasks and the dependencies and parallel relationships between them. The method includes: inputting the tree-structured task planning results into the invoked decomposition agent to convert each task in the tree-structured task planning results into a corresponding task node. Each task node carries semantic information of the corresponding planned task, which is a vector generated based on the corresponding task name and task description; analyzing the semantic information of each task node to determine the dependencies between them and creating directed edges to generate a task dependency graph; for each task node in the task dependency graph, checking all predecessor nodes pointing to the corresponding task node, and fusing the semantic information of all predecessor nodes with the semantic information of the corresponding task node, updating the semantic information of the corresponding task node based on the fused information; and performing task flow detection and loop detection based on the updated task dependency graph to determine that all task flows are unidirectional and that no loops are formed in information transmission, thus obtaining a directed acyclic graph. The directed acyclic graph includes each planned task and the dependencies and parallel relationships between them.

[0013] This invention also provides a search generation device based on multi-agent collaboration, comprising: a task planning module, which receives user search requests and calls a planning agent to perform task planning, generating a tree-like task planning result; a task parsing module, which, based on the tree-like task planning result, calls a decomposition agent to perform fine-grained decomposition of complex tasks, obtaining corresponding planning tasks and the dependencies and parallel relationships between planning tasks; a parallel execution module, which, based on the planning tasks and the dependencies and parallel relationships between planning tasks, calls multiple agents to execute corresponding planning tasks with parallel relationships in parallel, and after executing the corresponding planning tasks, calls agents to execute the next planning task with dependencies, obtaining corresponding search results; and a report generation module, which, based on the corresponding search results, calls a verification agent to perform multi-source information cross-validation, and generates a corresponding search report based on successful verification.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-agent cooperative search generation method as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-agent cooperative search generation method as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a multi-agent collaborative search generation method as described above.

[0017] The present invention provides a search generation method, apparatus, device, and medium based on multi-agent collaboration. By invoking a planning agent, the received user search request is transformed into a structured, hierarchical tree diagram. Combined with a decomposition agent, complex tasks are decomposed into fine-grained parts to generate new dependency graphs, enhancing the system's robustness and adaptability. This ensures that each task can be precisely executed by a dedicated agent, improving the system's executability and feasibility. Based on the planned tasks and the dependencies and parallel relationships between them, multiple agents are simultaneously mobilized to handle different tasks, significantly shortening the total execution time and maximizing efficiency. Furthermore, computing resources are intelligently scheduled and allocated according to task dependencies, allowing more agents to be launched in parallel during task-intensive periods and resources to be released during idle periods, thus optimizing resource utilization. Based on the corresponding search results, a verification agent is invoked to perform multi-source cross-validation, automatically identifying contradictory, inconsistent, or false information, ensuring that the final report's conclusions are based on solid and reliable facts. This significantly improves efficiency while comprehensively guaranteeing the factual accuracy of the report. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts of the search generation method based on multi-agent cooperation provided by the present invention; Figure 2 This is the second flowchart of the search generation method based on multi-agent cooperation provided by the present invention; Figure 3 This is a schematic diagram of the structure of the search generation device based on multi-agent cooperation provided by the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] Figure 1 This is a flowchart illustrating the search generation method based on multi-agent cooperation provided by the present invention, as shown below. Figure 1 As shown, the method includes: S11 receives the user's search request and calls the planning agent to perform task planning, generating a tree-like task planning result; S12, based on the tree-structured task planning results, call the decomposition agent to perform fine-grained decomposition of complex tasks, and obtain the corresponding planning tasks and the dependencies and parallel relationships between each planning task; S13, based on the planning task and the dependencies and parallel relationships between the planning tasks, call multiple agents to execute the corresponding planning tasks with parallel relationships in parallel, and after the corresponding planning task is completed, call the agent to execute the next planning task with dependencies to obtain the corresponding search results. S14. Based on the corresponding search results, call the verification agent to perform cross-validation of multi-source information, and generate the corresponding search report based on the successful verification.

[0022] It should be noted that the step numbers "S1N" in this specification do not represent the order of steps in the multi-agent collaborative search generation method. The following details will explain this in conjunction with... Figure 2 The present invention describes a search generation method based on multi-agent cooperation.

[0023] Step S11: Receive the user's search request and call the planning agent to perform task planning, generating a tree-like task planning result.

[0024] In this embodiment, receiving user search requests and invoking a planning agent to perform task planning and generate a tree-like task planning result includes: inputting the received user search requests into the invoked planning agent to perform request parsing and intent recognition based on the input user search requests; performing top-down hierarchical decomposition based on the recognized intent to generate main tasks, and recursively decomposing subtasks for each main task; and organizing and outputting the decomposed tasks in a tree structure to obtain the tree-like task planning result.

[0025] It should be noted that the planning agent can select the corresponding large language model according to actual design needs. For example, it can use GPT-4, which, based on its powerful multimodal language understanding and generation capabilities, can transform the input user search requirements into a high-level, structured tree-like task planning result. This tree-like task planning result defines the research objectives, analysis dimensions, and expected output format.

[0026] Step S12: Based on the tree-structured task planning results, call the decomposition agent to perform fine-grained decomposition of complex tasks, and obtain the corresponding planned tasks and the dependencies and parallel relationships between each planned task.

[0027] It should be added that the decomposition agent can be selected according to actual design requirements. For example, a graph neural network (GNN) can be used to automatically generate a directed acyclic graph (DAG) for each task node, such as retrieving financial data, calculating valuation, and writing a summary, as well as its potential dependencies. For example, calculating valuation must first complete the retrieval and analysis of financial data. This DAG clarifies the dependencies and parallel relationships between all tasks in the form of a state graph, thereby fundamentally solving the problem that traditional chain-like architectures cannot process in parallel.

[0028] Specifically, based on the tree-structured task planning results, a decomposition agent is invoked to perform fine-grained decomposition of complex tasks, obtaining the planned tasks and the dependencies and parallel relationships between them. This includes: inputting the tree-structured task planning results into the invoked decomposition agent to convert each task in the tree-structured task planning results into a corresponding task node. Each task node carries semantic information of the corresponding planned task, which is a vector generated based on the corresponding task name and task description; analyzing the semantic information of each task node to determine the dependencies between them and creating directed edges to generate a task dependency graph; for each task node in the task dependency graph, checking all predecessor nodes pointing to the corresponding task node, and fusing the semantic information of all predecessor nodes with the semantic information of the corresponding task node, and updating the semantic information of the corresponding task node based on the fused information; based on the updated task dependency graph, performing task flow detection and loop detection to determine that all task flows are unidirectional and that no loops are formed in information transmission, thus obtaining a directed acyclic graph; where the directed acyclic graph includes each planned task and the dependencies and parallel relationships between them.

[0029] For example, a directed non-transformation graph includes N1[node: retrieve financial report] [Data requirements, quality standards], N2 [node: calculation of valuation] Depends on N1 and N3 [nodes: collecting macro data] [Can be parallelized], so that N1 and N3 tasks can be executed in parallel according to the directed non-transformation graph, while N2 task needs to be executed after N1 task is completed.

[0030] It should be added that during the iterative process of generating corresponding fusion information for each task node in the task dependency graph, the information transmission on each edge implicitly carries a strength or importance weight. A loop typically contains a weakest dependency. Therefore, after performing task flow detection and loop detection based on the updated task dependency graph, the process also includes: when a loop is determined to be formed by information transmission, identifying and removing the edge with the lowest weight in the loop, or generating an error report and inputting it into the decomposition agent to regenerate the directed acyclic graph. The error report includes task nodes with logical loops. When a non-unidirectional task flow is determined, and the corresponding task is determined to be executed in parallel, removing the corresponding edge, or analyzing the semantic information of the task nodes corresponding to the non-unidirectional task flow and identifying potential, undeclared strong dependencies to correct the directed edges of the corresponding tasks.

[0031] Step S13: Based on the planning tasks and the dependencies and parallel relationships between them, call multiple agents to execute the corresponding planning tasks with parallel relationships in parallel. After completing the corresponding planning task, call the agent to execute the next planning task with dependencies to obtain the corresponding search results.

[0032] It should be noted that the Orchestrator scheduler can be used as the execution engine for the state graph. Based on a multi-directional acyclic graph and asynchronous message queue, it can distribute parallel tasks without pre-dependencies to a cluster of tool-invoking agents in parallel. The scheduler uses an improved Dijkstra algorithm to optimize task execution paths, minimizing the overall task completion time and thus significantly improving efficiency. After the pre-dependent tasks are completed, the Orchestrator scheduler can invoke multiple agents in parallel to execute the next planned task after the pre-dependent tasks have been completed. Furthermore, after receiving a task, the corresponding agent executes the corresponding task based on the dependencies between tasks and returns the corresponding search results to the Orchestrator scheduler. After the task is completed, the Orchestrator scheduler submits the corresponding search results for subsequent multi-source information cross-validation, as detailed below.

[0033] For example, based on the directed acyclic graph listed above, Agent-SEC and Agent-Reuters execute Task 1 and Task 3 in parallel. Agent-SEC is responsible for retrieving SEC financial statements for Task 1, and Agent-Reuters is responsible for collecting macroeconomic indicators for Task 3. After Task 1 is completed, Agent-Bloomberg is called to execute Task 2 to query historical stock prices.

[0034] Step S14: Based on the corresponding search results, call the verification agent to perform cross-validation of multi-source information, and generate the corresponding search report based on the successful verification.

[0035] In this embodiment, the search results include data fragments and the data sources of those data fragments. Based on the corresponding search results, a verification agent is invoked to perform multi-source information cross-validation. Based on successful verification, a corresponding search report is generated. This includes: inputting the corresponding search results into the invoked verification agent to pair data fragments from different data sources, using the target data source or previously verified data fragments as premise sentences and other data fragments as hypothesis sentences to obtain corresponding pairing results; inputting the pairing results into a consistency recognition model to obtain the verification results output by the consistency recognition model; wherein, the consistency recognition model is based on a pre-trained model, fine-tuned using financial domain corpus; and generating a search report based on the successful verification result and the corresponding search results.

[0036] It should be noted that the premise statement can be confirmed and irrefutable facts (such as raw data from official financial reports), evidence marked as consistent in the preceding verification, and data fragments from highly credible data sources (such as authoritative media or official databases). Additionally, the hypothesis statement can be evidence that needs to be verified for consistency with the premise statement, or a conclusion derived from evidence, such as newly retrieved, unverified evidence fragments, analytical conclusions generated by other agents, or evidence that potentially contradicts known facts; no further limitations are imposed here.

[0037] Specifically, the pairing results are input into the consistency recognition model to obtain the verification results output by the consistency recognition model. This includes: inputting the pairing results into the consistency recognition model to determine the semantic consistency of the corresponding paired evidence fragments for each pairing result, and mapping it to a preset three-dimensional judgment matrix to obtain the three-dimensional judgment matrix; wherein, the preset three-dimensional judgment matrix includes support, contradiction, and relevance. Support is used to characterize the degree of entailment of the premise sentence to the hypothesis sentence, contradiction is used to characterize the degree of contradiction between the premise sentence and the hypothesis sentence, and relevance is used to characterize the neutrality of the premise sentence and the hypothesis sentence. The consistency recognition model is trained based on historical pairing results and the actual matrices corresponding to the historical pairing results; according to the three-dimensional judgment matrix, if the relevance is greater than a first preset threshold, the verification result is determined to be verified successfully; or, if the relevance is less than or equal to the first preset threshold, the verification result is determined to be verified unsuccessfully based on the contradiction being greater than a second preset threshold; or, if the contradiction is less than or equal to the second preset threshold, the verification result is determined to be verified unsuccessfully based on the support being less than a third preset threshold; and if the support is greater than or equal to the third preset threshold, the verification result is determined to be verified successfully.

[0038] It should be added that for verification results regarding whether the contradiction level is greater than the second preset threshold, and whether the support level is less than the third preset threshold, special styles can be set to make them stand out more in the flowchart, thus emphasizing their status as key decision points. Examples include using specific colors for filling and setting border widths; no further limitations are made here. Furthermore, the consistency recognition model can be selected based on actual design requirements. For example, DeBERTa, with its preset financial domain corpus obtained from the FinNLI dataset, can be fine-tuned to ensure it accurately understands financial terminology and complex reasoning patterns, thus guaranteeing the accuracy of the verification.

[0039] Furthermore, based on the verification result being successful, a search report is generated according to the corresponding search results. This includes: aggregating the corresponding data content, verification status, support, contradiction, and relevance obtained from the three-dimensional judgment matrix, data source, related tasks, and data types to generate a knowledge graph or fact database to provide material for report generation; extracting core themes from the aggregated data and generating a narrative outline based on user search needs; searching for the most relevant high-support arguments from the knowledge graph or fact database based on each sub-argument in the narrative outline, connecting multiple high-support evidences to form a logical chain of evidence, and organizing them into corresponding paragraphs using a natural language model to obtain an initial report; identifying data-driven evidence in the initial report, generating bar charts or line graphs, and embedding them into the initial report to obtain the final search report.

[0040] It should be added that search reports can be generated based on actual user needs. For example, if the actual user need is intelligent credit rating, the corresponding search report can be a credit rating report that complies with SEC Rule 17g-7. If the actual user need is merger and acquisition due diligence, the corresponding search report can be a risk analysis and valuation report. If the actual user need is internal analysis, the corresponding search report can be a customized research report for the company's investment department, risk management department, and senior management to help them quickly obtain accurate and reliable data insights and support smarter business decisions.

[0041] In an optional embodiment, after invoking the verification agent to perform multi-source information cross-validation based on the corresponding search results, the method further includes: based on the failed verification, invoking the reflection agent according to the corresponding pairing results to determine the semantic similarity of the corresponding pairing results, the difference between the keyword premise sentence and the hypothesis sentence, and the information release time difference, to determine the credibility of the data source, and to determine the planning task corresponding to the hypothesis sentence in the failed pairing results, as well as the retrieval keywords and retrieval timestamps corresponding to the planning task, to obtain feature combinations; based on the feature combinations, using a Bayesian optimization algorithm, defining a multi-dimensional search space, and based on any correction scheme in the search space, predicting the probability distribution of potential success rates, to obtain probability distribution prediction results; based on the probability distribution prediction results, selecting the correction scheme with the highest predicted success rate as the correction instruction, and sending it to the decomposition agent, thereby avoiding repeated starting from scratch and significantly improving factual accuracy.

[0042] It should be added that the reasons for validation failures may include various factors, such as conflicting evidence, task timeouts, or API errors. The specific reasons will be determined based on the actual situation, and no further limitations are made here. Additionally, the search space can be constructed based on all possible correction schemes determined according to the feature combinations. Correction schemes include query restructuring, data source switching, and execution strategy adjustments. Query restructuring may involve modifying keywords or query syntax; data source switching may involve selecting a more reliable or timely source from alternative data sources; and execution strategy adjustments may involve modifying task priority or the degree of parallelization. No further limitations are made here.

[0043] Furthermore, the Bayesian optimization algorithm does not directly evaluate all possible correction schemes. Instead, it establishes a surrogate model (probabilistic model) to estimate the potential success rate of each correction scheme based on existing historical execution data. Then, it uses a sampling function to balance "exploration" (trying unknown but potentially promising schemes) and "utilization" (selecting the known best-performing scheme) to determine the next most promising correction scheme to try. Thus, through Bayesian optimization, the agent can find the optimal correction scheme in the fewest attempts. Compared to traditional random or grid search, it is much more efficient, enabling the system to recover quickly and intelligently from failures, forming an efficient self-optimization loop, thereby significantly improving task success rate and system robustness.

[0044] Furthermore, in the actual design process, after the correction instruction is sent to the decomposition agent, the following steps are taken: the decomposition agent corrects the corresponding planned task according to the correction instruction and puts it back into the scheduling queue to wait for the Orchestrator to reschedule it; or, the decomposition agent generates a reconstructed query or instruction according to the correction instruction so that the new task can be executed immediately without having to be scheduled again through the scheduling queue.

[0045] In addition, after the correction instruction is sent to the decomposition agent, whether it is rescheduled or a new task is executed immediately, the verification agent is invoked to perform multi-source information cross-validation based on the corresponding search results, so that the corresponding verification results and correction scheme are used as a set of training data to update the Bayesian optimization algorithm online.

[0046] In one optional embodiment, after generating a corresponding search report based on successful verification, the process includes: calling a compliance check agent to perform structured and semantic detection based on the search report, obtaining structured detection results and semantic detection results; when both structured and semantic detection results are found to be compliant, outputting the generated search report as the final search report; when structured and / or semantic detection results are found to be non-compliant, conducting a comprehensive assessment based on the non-compliance of the structured and / or semantic detection results, determining a severity score for the non-compliance, and automatically injecting a disclaimer or risk disclosure when the severity score is less than or equal to a preset score threshold.

[0047] It should be added that the pre-set rule base can include rules from the Financial Industry Regulatory Authority (FIRA), the Securities and Exchange Commission (SEC), and the China Securities Regulatory Commission (CSRC). The specific rule base can be designed and constructed according to actual needs. For example, for FIRA rules, such as Rule 2241 and Rule 2242, research reports are required to disclose analysts' conflicts of interest, the company's investment banking relationship with the target company over the past 12 months, and equity holdings of more than 1%. For SEC rules, such as Rule 17g-7, rating agencies must disclose in detail their rating methodologies, key assumptions, and any potential conflicts of interest when issuing credit ratings. For CSRC rules, such as the "Interim Provisions on the Issuance of Securities Research Reports," reports are required to be objective, professional, and prudent, and are prohibited from providing false information or implying guarantees of investment returns.

[0048] Specifically, based on the search report, a compliance inspection intelligent agent is invoked to perform structured and semantic detection, obtaining structured and semantic detection results. This includes: invoking the compliance inspection intelligent agent to use regular expressions to match a preset rule base to identify illegal content in the search report and obtain corresponding structured detection results; wherein the preset rule base is constructed based on the rules of different financial regulatory agencies; and performing semantic analysis on the search report to analyze the tone and wording in the search report, and combining it with the preset rule base to determine whether there is any illegal behavior and obtain corresponding semantic detection results. In this way, the compliance inspection intelligent agent minimizes regulatory risks.

[0049] It should be noted that determining the content of violations in the search report can include: checking whether there is a risk disclosure statement, whether the filing number is correctly marked, and whether target keywords exist; determining whether there are violations can include: judging whether there are violations such as inducing people without investment intentions to open an account or implying guaranteed investment returns.

[0050] In addition, a comprehensive assessment is conducted based on the structured detection results and / or semantic detection results to determine the severity score of the violation. This includes: identifying the corresponding violating content based on the structured detection results and determining the severity level of each violating content in conjunction with a pre-set rule base; and / or identifying the corresponding violating behavior based on the semantic detection results and determining the severity level of each violating behavior in conjunction with a pre-set rule base; and a comprehensive assessment is conducted based on the severity level of each violating content and / or each violating behavior to determine the severity score of the violation.

[0051] Furthermore, after determining the severity score of the violation, the system includes: blocking report generation when the severity score exceeds a preset threshold, and feeding back the violation details to the responding agent for correction. Through this mandatory built-in check, the compliance violation rate is reduced from 15% in the traditional manual mode to 0.3%, fundamentally improving risk control capabilities.

[0052] In summary, this embodiment of the invention, by invoking a planning agent, transforms the received user search requests into a structured, hierarchical tree diagram. Combined with a decomposition agent, it performs fine-grained decomposition of complex tasks to generate a new dependency graph, enhancing the system's robustness and adaptability. This ensures that each task can be precisely executed by a dedicated agent, improving the system's executability and feasibility. Based on the planned tasks and the dependencies and parallel relationships between them, multiple agents are simultaneously mobilized to handle different tasks, significantly shortening the total execution time and maximizing efficiency. Furthermore, based on task dependencies, computing resources are intelligently scheduled and allocated, allowing more agents to be launched in parallel during task-intensive periods and resources to be released during idle periods, thus optimizing resource utilization. Based on the corresponding search results, a verification agent is invoked to perform multi-source information cross-validation, automatically identifying contradictory, inconsistent, or false information, ensuring that the final report's conclusions are based on solid and reliable facts. This significantly improves efficiency while comprehensively guaranteeing the factual accuracy of the report.

[0053] The search generation apparatus based on multi-agent cooperation provided by the present invention will be described below. The search generation apparatus based on multi-agent cooperation described below can be referred to in correspondence with the search generation method based on multi-agent cooperation described above.

[0054] Figure 3 A schematic diagram of a search generation device based on multi-agent cooperation is shown. The device includes: Task planning module 31 receives user search requests and calls the planning agent to perform task planning, generating a tree-like task planning result. Task parsing module 32, based on the tree-structured task planning results, calls the decomposition agent to perform fine-grained decomposition of complex tasks, and obtains the corresponding planned tasks and the dependencies and parallel relationships between each planned task; The parallel execution module 33 calls multiple agents to execute the corresponding planning tasks with parallel relationships in parallel, based on the planning tasks and the dependencies and parallel relationships between them, and after executing the corresponding planning tasks, calls the agents to execute the next planning task with the dependencies, and obtains the corresponding search results. The report generation module 34 calls the verification agent to perform multi-source information cross-validation based on the corresponding search results, and generates the corresponding search report based on the successful verification.

[0055] In this embodiment, the task planning module 31 is used to: input the received user search requirements into the called planning agent, and perform requirement parsing and intent recognition based on the input user search requirements; perform top-down hierarchical decomposition based on the recognized intent to generate main tasks, and recursively decompose sub-tasks for each main task; and organize and output the decomposed tasks in a tree structure to obtain tree-structured task planning results.

[0056] Additionally, the task parsing module 32 is used to: input the tree-like task planning results into the called decomposition agent, so as to convert each task in the tree-like task planning results into a corresponding task node. Each task node carries the semantic information of the corresponding planned task. The semantic information is a vector generated based on the corresponding task name and task description. The semantic information of each task node is analyzed to determine the dependency relationship between each task node and to create directed edges to generate a task dependency graph. For each task node in the task dependency graph, all predecessor nodes pointing to the corresponding task node are checked, and the semantic information of all predecessor nodes is fused with the semantic information of the corresponding task node. The semantic information of the corresponding task node is updated according to the fused information. Based on the updated task dependency graph, task flow detection and loop detection are performed to determine that all task flows are unidirectional and that no loop is formed in the information transmission, thus obtaining a directed acyclic graph. The directed acyclic graph includes each planned task and the dependency and parallel relationships between each planned task.

[0057] It should be added that, after performing task flow detection and loop detection based on the updated task dependency graph, the task parsing module 32 is also used to: when it is determined that information transmission forms a loop, identify the edge with the lowest weight in the loop and remove it, or generate an error report and input the error report to the decomposition agent to regenerate the directed acyclic graph. The error report includes task nodes with logical loops; when it is determined that there is a non-unidirectional task flow, when it is determined that the corresponding task is a parallel execution task, remove the corresponding edge, or analyze the semantic information of the task nodes corresponding to the non-unidirectional task flow and determine potential, undeclared strong dependencies to correct the directed edges of the corresponding task.

[0058] The report generation module 34 is used to: input the corresponding search results into the invoked verification agent to pair data fragments from different data sources, using the target data source or previously verified data fragments as premise sentences and other data fragments as hypothesis sentences to obtain the corresponding pairing results; input the pairing results into the consistency recognition model to obtain the verification results output by the consistency recognition model; wherein, the consistency recognition model is based on a pre-trained model and fine-tuned using financial domain corpus; based on the verification result being verified as passed, a search report is generated according to the corresponding search results.

[0059] Specifically, the report generation module 34 is further configured to: input the pairing results into the consistency recognition model to determine the semantic consistency of the corresponding paired evidence fragments for the pairing results, and map them to a preset three-dimensional judgment matrix to obtain a three-dimensional judgment matrix; wherein, the preset three-dimensional judgment matrix includes support, contradiction, and relevance, support is used to characterize the degree of implication of the premise sentence to the hypothesis sentence, contradiction is used to characterize the degree of contradiction between the premise sentence and the hypothesis sentence, and relevance is used to characterize the neutrality of the premise sentence and the hypothesis sentence, and the consistency recognition model is trained based on historical pairing results and the actual matrix corresponding to the historical pairing results; according to the three-dimensional judgment matrix, if the relevance is greater than a first preset threshold, the verification result is determined to be verified as passed; or, if the relevance is less than or equal to the first preset threshold, the verification result is determined to be verified as failed based on the contradiction being greater than a second preset threshold; or, if the contradiction is less than or equal to the second preset threshold, the verification result is determined to be verified as failed based on the support being less than a third preset threshold; and if the support is greater than or equal to the third preset threshold, the verification result is determined to be verified as passed.

[0060] Furthermore, the report generation module 34 is also used to: based on the verification result being verified as passed, aggregate the corresponding data content, verification status, support, contradiction, and relevance obtained from the three-dimensional judgment matrix, data source, associated tasks, and data type to generate a knowledge graph or fact database to provide materials for report generation; extract the core theme from the aggregated data and generate a narrative outline based on the user's search needs; based on each sub-argument in the narrative outline, search for the most relevant high-support arguments from the knowledge graph or fact database, and link the multiple high-support evidence found to form a logical chain of evidence, and use a natural language model to organize it into corresponding paragraphs to obtain an initial report; identify the data-type evidence in the initial report, generate a bar chart or line chart, and embed it into the initial report to obtain the final search report.

[0061] In an optional embodiment, the device further includes a correction module, configured to: after invoking a verification agent to perform multi-source information cross-validation based on the corresponding search results, and based on the failure of the verification, invoking a reflection agent based on the corresponding pairing results to determine the semantic similarity of the corresponding pairing results, the difference between the keyword premise sentence and the hypothesis sentence, and the information release time difference, to determine the credibility of the data source, and to determine the planning task corresponding to the hypothesis sentence in the failed pairing results, as well as the retrieval keywords and retrieval timestamps corresponding to the planning task, thereby obtaining a feature combination; based on the feature combination, using a Bayesian optimization algorithm, defining a multi-dimensional search space, and based on any correction scheme in the search space, predicting the probability distribution of the potential success rate, thereby obtaining a probability distribution prediction result; based on the probability distribution prediction result, selecting the correction scheme with the highest predicted success rate as the correction instruction, and sending it to the decomposition agent, thereby avoiding repeated starting from scratch and significantly improving the accuracy of the facts.

[0062] Furthermore, the correction module is also used in the actual design process to: after sending the correction instruction to the decomposition agent, the decomposition agent corrects the corresponding planned task according to the correction instruction and puts it back into the scheduling queue to wait for the Orchestrator to reschedule it; or, the decomposition agent generates a reconstructed query or instruction according to the correction instruction so that the new task can be executed immediately without having to be scheduled again through the scheduling queue.

[0063] In an optional embodiment, the device further includes a report checking module, configured to: after generating a corresponding search report based on successful verification, invoke a compliance checking agent to perform structured detection and semantic detection based on the search report, and obtain structured detection results and semantic detection results; when it is determined that both the structured detection results and the semantic detection results are compliant, output the generated search report as the final search report; when it is determined that the structured detection results and / or the semantic detection results are non-compliant, conduct a comprehensive evaluation based on the non-compliance of the structured detection results and / or the semantic detection results, determine the severity score of the non-compliance, and automatically inject a disclaimer or risk disclosure when the severity score of the non-compliance is less than or equal to a preset score threshold.

[0064] Specifically, the report inspection module is also used to: invoke a compliance inspection intelligent agent based on the search report, and use regular expressions to match a preset rule base to determine the illegal content in the search report and obtain the corresponding structured detection results; wherein the preset rule base is constructed in advance based on the rules of different financial regulatory agencies; and perform semantic analysis on the search report, analyze the tone and wording in the search report, and combine it with the preset rule base to determine whether there is any illegal behavior and obtain the corresponding semantic detection results, thereby minimizing regulatory risks through the compliance inspection intelligent agent.

[0065] In addition, the report inspection module is also used to: determine the corresponding violation content based on the structured detection results, and determine the violation severity level of each violation content in combination with the preset rule base; and / or, determine the corresponding violation behavior based on the semantic detection results, and determine the violation severity level of each violation behavior in combination with the preset rule base; and conduct a comprehensive evaluation based on the severity level of each violation content and / or the severity level of each violation behavior to determine the violation severity score.

[0066] In addition, the report inspection module is also used to: after determining the severity score of the violation, block the generation of the report when the severity score of the violation is greater than the preset score threshold, and feed back the violation details to the responding agent for correction. In this way, through this mandatory built-in inspection, the compliance violation rate is reduced from 15% in the traditional manual mode to 0.3%, which fundamentally improves the risk control capability.

[0067] In summary, this embodiment of the invention utilizes a task planning module to invoke a planning agent, transforming the received user search requests into a structured, hierarchical tree diagram. A task parsing module then invokes a decomposition agent to perform fine-grained decomposition of complex tasks, generating a new dependency graph. This enhances the system's robustness and adaptability, ensuring that each task can be precisely executed by a dedicated agent, thus improving the system's executability and feasibility. Furthermore, a parallel execution module, based on the planned tasks and their dependencies and parallel relationships, simultaneously mobilizes multiple agents to handle different tasks, significantly shortening the total execution time and maximizing efficiency. Based on task dependencies, it intelligently schedules and allocates computing resources, enabling the parallel activation of more agents during task-intensive periods and releasing resources during idle periods, thereby optimizing resource utilization. Finally, based on the corresponding search results, a verification agent is invoked to perform multi-source information cross-validation, automatically identifying contradictory, inconsistent, or false information. This ensures that the report generation module's final conclusions are based on solid and reliable facts, significantly improving efficiency while comprehensively guaranteeing the accuracy of the reported facts.

[0068] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440. The processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a search generation method based on multi-agent cooperation. This method includes: receiving a user search request and calling a planning agent to perform task planning, generating a tree-like task planning result; based on the tree-like task planning result, calling a decomposition agent to perform fine-grained decomposition of complex tasks, obtaining corresponding planning tasks and the dependencies and parallel relationships between these planning tasks; based on the planning tasks and the dependencies and parallel relationships between these planning tasks, calling multiple agents to execute corresponding planning tasks with parallel relationships in parallel; and after executing the corresponding planning task, calling an agent to execute the next planning task with dependencies, obtaining the corresponding search result; based on the corresponding search result, calling a verification agent to perform multi-source information cross-validation, and generating a corresponding search report based on successful verification.

[0069] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the search generation method based on multi-agent cooperation provided by the above methods. The method includes: receiving a user search request and calling a planning agent to perform task planning and generate a tree-like task planning result; based on the tree-like task planning result, calling a decomposition agent to perform fine-grained decomposition of complex tasks to obtain corresponding planning tasks and the dependencies and parallel relationships between planning tasks; based on the planning tasks and the dependencies and parallel relationships between planning tasks, calling multiple agents to execute corresponding planning tasks with parallel relationships in parallel, and after executing the corresponding planning tasks, calling agents to execute the next planning task with dependencies to obtain the corresponding search results; based on the corresponding search results, calling a verification agent to perform multi-source information cross-validation, and generating a corresponding search report based on the successful verification.

[0071] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the search generation method based on multi-agent cooperation provided by the above methods. The method includes: receiving a user search request and invoking a planning agent to perform task planning and generate a tree-like task planning result; based on the tree-like task planning result, invoking a decomposition agent to perform fine-grained decomposition of complex tasks to obtain corresponding planning tasks and the dependencies and parallel relationships between planning tasks; based on the planning tasks and the dependencies and parallel relationships between planning tasks, invoking multiple agents to execute corresponding planning tasks with parallel relationships in parallel, and after executing the corresponding planning tasks, invoking agents to execute the next planning task with dependencies to obtain the corresponding search results; based on the corresponding search results, invoking a verification agent to perform multi-source information cross-validation, and generating a corresponding search report based on the successful verification.

[0072] The device embodiments described above are merely illustrative. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0073] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A search generation method based on multi-agent cooperation, characterized in that, include: Receive user search requests and invoke the planning agent to perform task planning, generating a tree-like task planning result; Based on the tree-like task planning results, the decomposition agent is invoked to perform fine-grained decomposition of complex tasks, thereby obtaining the corresponding planning tasks and the dependencies and parallel relationships between each planning task. Based on the planning tasks and the dependencies and parallel relationships between them, multiple agents are invoked to execute the corresponding planning tasks with parallel relationships in parallel. After the corresponding planning task is completed, the agent is invoked to execute the next planning task with a dependency relationship to obtain the corresponding search results. Based on the corresponding search results, the verification agent is invoked to perform cross-validation of multi-source information, and a corresponding search report is generated based on the successful verification.

2. The search generation method based on multi-agent cooperation according to claim 1, characterized in that, The search results include data fragments and the data sources of those data fragments. Based on the corresponding search results, a verification agent is invoked to perform multi-source information cross-validation. Based on successful verification, a corresponding search report is generated, including: The corresponding search results are input into the called verification agent to pair data fragments from different data sources, and the target data source or previously verified data fragments are used as premise sentences, while other data fragments are used as hypothesis sentences to obtain the corresponding pairing results. The pairing results are input into the consistency recognition model to obtain the verification results output by the consistency recognition model; wherein, the consistency recognition model is obtained by fine-tuning a pre-trained model using financial domain corpus. Based on the verification result being successful, a search report is generated according to the corresponding search results.

3. The search generation method based on multi-agent cooperation according to claim 2, characterized in that, The pairing results are input into the consistency recognition model to obtain the verification results output by the consistency recognition model, including: The pairing results are input into a consistency recognition model to determine the semantic consistency of the corresponding paired evidence fragments and map them to a preset three-dimensional judgment matrix to obtain a three-dimensional judgment matrix. The preset three-dimensional judgment matrix includes support, contradiction, and relevance. Support characterizes the degree of implication of the premise sentence to the hypothesis sentence; contradiction characterizes the degree of contradiction between the premise sentence and the hypothesis sentence; and relevance characterizes the degree of neutrality between the premise sentence and the hypothesis sentence. The consistency recognition model is trained based on historical pairing results and the corresponding actual matrices. Based on the three-dimensional judgment matrix, if the relevance is greater than a first preset threshold, the verification result is determined to be successful; or... When the relevance is determined to be less than or equal to the first preset threshold, and the contradiction is determined to be greater than the second preset threshold, the verification result is determined to be a verification failure; or, When the degree of contradiction is determined to be less than or equal to the second preset threshold, the verification result is determined to be a verification failure based on the degree of support being less than the third preset threshold. When the support level is determined to be greater than or equal to the third preset threshold, the verification result is determined to be a successful verification.

4. The search generation method based on multi-agent cooperation according to claim 2, characterized in that, After invoking the verification agent to perform multi-source information cross-validation based on the corresponding search results, the process also includes: Based on the failed verification, according to the corresponding pairing results, the reflective agent is invoked to determine the semantic similarity of the corresponding pairing results, the difference between the premise sentence and the hypothesis sentence in the keyword, and the information release time difference, to determine the credibility of the data source, and to determine the planning task corresponding to the hypothesis sentence in the failed verification pairing results, as well as the retrieval keywords and retrieval timestamps corresponding to the planning task, to obtain the feature combination; Based on the feature combination, a multi-dimensional search space is defined using the Bayesian optimization algorithm, and the probability distribution of potential success rate is predicted based on any correction scheme in the search space, thus obtaining the probability distribution prediction result. Based on the probability distribution prediction results, the correction scheme with the highest prediction success rate is selected as the correction instruction and sent to the decomposition agent.

5. The search generation method based on multi-agent cooperation according to claim 1, characterized in that, After successful verification and the generation of the corresponding search report, the report includes: Based on the search report, the compliance check agent is invoked to perform structured detection and semantic detection, and the structured detection results and semantic detection results are obtained. When both the structured detection results and the semantic detection results are determined to be compliant, the generated search report is output as the final search report. When the structured detection result and / or the semantic detection result are found to be in violation, a comprehensive evaluation is performed based on the violation of the structured detection result and / or the semantic detection result to determine the severity score of the violation. If the severity score of the violation is less than or equal to a preset score threshold, a disclaimer or risk disclosure is automatically injected.

6. The search generation method based on multi-agent cooperation according to claim 5, characterized in that, Based on the search report, the compliance check agent is invoked to perform structured and semantic detection, yielding structured and semantic detection results, including: Based on the search report, a compliance check agent is invoked to use regular expressions to match a preset rule base, identify the non-compliant content in the search report, and obtain the corresponding structured detection results; wherein, the preset rule base is constructed prior to the rules of different financial regulatory agencies; and, Semantic analysis is performed on the search report to analyze the tone and wording, and in conjunction with the preset rule base, to determine whether there are any violations, thereby obtaining the corresponding semantic detection results.

7. The search generation method based on multi-agent cooperation according to claim 1, characterized in that, Based on the tree-structured task planning results, a decomposition agent is invoked to perform fine-grained decomposition of complex tasks, obtaining the planned tasks and the dependencies and parallel relationships between them, including: The tree-structured task planning results are input into the called decomposition agent to convert each task in the tree-structured task planning results into a corresponding task node. Each task node carries semantic information of the corresponding planned task. The semantic information is a vector generated based on the corresponding task name and task description. The semantic information of each task node is analyzed to determine the dependencies between task nodes and to create directed edges, thereby generating a task dependency graph. For each task node in the task dependency graph, check all predecessor nodes pointing to the corresponding task node, fuse the semantic information of all predecessor nodes with the semantic information of the corresponding task node, and update the semantic information of the corresponding task node according to the fused information. Based on the updated task dependency graph, task flow detection and loop detection are performed to determine that all task flows are unidirectional and that no loops are formed in information transmission, thus obtaining a directed acyclic graph; wherein, the directed acyclic graph includes each planned task and the dependency and parallel relationships between each planned task.

8. A search generation device based on multi-agent cooperation, characterized in that, include: The task planning module receives user search requests and calls the planning agent to perform task planning, generating a tree-like task planning result. The task parsing module, based on the tree-like task planning results, calls the decomposition agent to perform fine-grained decomposition of complex tasks, and obtains the corresponding planned tasks and the dependencies and parallel relationships between the planned tasks. The parallel execution module calls multiple agents to execute corresponding planning tasks with parallel relationships in parallel, based on the planning tasks and the dependencies and parallel relationships between them. After executing the corresponding planning task, it calls an agent to execute the next planning task with the dependency relationship to obtain the corresponding search results. The report generation module calls the verification agent to perform cross-validation of multi-source information based on the corresponding search results, and generates the corresponding search report based on the successful verification.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the multi-agent cooperative search generation method as described in any one of claims 1 to 6.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-agent cooperative search generation method as described in any one of claims 1 to 6.

Images