A multi-agent collaborative generation method for industrial decision-making

Abstract

The application discloses a kind of multi-agent collaborative generation methods for industry decision-making, by main agent and several expert agents collaborative execution, main agent is responsible for task arrangement, and expert agent is responsible for special analysis, comprising: receiving industry decision-making request and associated data and constructing hierarchical task context, hierarchical task context can be dynamically written and expose the minimum necessary information to agent;Determine task path by main agent based on hierarchical task context, generate task processing scheme;Identify information gap and perform data and external knowledge retrieval, obtain structured search results;At least one expert agent stage is constructed;Expert agent stage is executed, and output result is obtained by expert agent analysis and is combined by main agent to generate industry decision-making result.The application realizes context structured management, task path dynamic scheduling, external knowledge reliability introduction and multi-expert agent collaborative arrangement, improves the professionalism, timeliness and explainability of decision result.

Description

Technical Field

[0001] This invention relates to the technical field of artificial intelligence and large language model applications, and in particular to a multi-agent collaborative generation method for industrial decision-making. Background Technology

[0002] In recent years, the rapid development of generative artificial intelligence technology, especially large language models, has made intelligent agents, as key carriers connecting large language models with real-world business scenarios, a mainstream trend in AI applications. Large language models possess strong language understanding, knowledge summarization, and reasoning capabilities. Multi-agent collaboration further breaks down complex tasks and assigns them to agents with different professional capabilities, providing new technical approaches for knowledge-intensive tasks such as industry analysis, comprehensive judgment, and decision support.

[0003] Currently, industrial decision-making scenarios demand extremely high levels of professionalism, timeliness, and interpretability in their analytical results, often involving the comprehensive analysis of multi-source, heterogeneous information such as market data, industry knowledge, and historical experience. However, existing intelligent agent application solutions mostly remain at the level of general question answering or single-agent task execution, failing to delve into the real business processes of industrial decision-making. They struggle to simultaneously ensure the completeness of context management, the reliability of external knowledge introduction, and the depth of multi-domain expert collaboration, and have yet to form a mature system framework to support the intelligent generation of industrial decisions.

[0004] In traditional large language model application interaction modes, existing industrial decision-making intelligent agent solutions suffer from the following main drawbacks. First, context management is rudimentary. Traditional solutions typically concatenate all dialogue and data into a single prompt word input into the model, failing to provide structured management based on information source and lifecycle, and struggling to expose the minimum necessary information according to the agent's actual needs. This results in contextual redundancy, difficulty in effectively extracting key information, and the risk of over-exposure of sensitive information. Second, task scheduling is rigid. Traditional solutions often employ fixed processes or one-time task planning, unable to dynamically adjust task paths based on context updates and information gaps, making it difficult to adapt to the increasingly refined needs in industrial decision-making processes. Third, the introduction of external knowledge is unreliable. When supplementing retrieved data and external knowledge, traditional solutions lack conflict identification and confidence arbitration of multi-source retrieval results, easily incorporating contradictory facts into decision-making. Finally, expert collaboration lacks orchestration. When scheduling multiple expert agents, traditional solutions often rely on manually specifying the execution order or simple serial calls, failing to automatically build dependencies between tasks and implement parallel and topologically ordered execution, leading to low collaboration efficiency and difficulty in merging results. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a multi-agent collaborative generation method for industrial decision-making that can improve the integrity of context management, realize dynamic scheduling of task paths, ensure the reliability of external knowledge, and support multi-expert agent collaborative orchestration.

[0006] To achieve the above objectives, the technical solution provided by this invention is: a multi-agent collaborative generation method for industry decision-making, which is executed collaboratively by a master agent and several expert agents. The master agent is responsible for the overall task arrangement, and the expert agents are responsible for specialized analysis within their respective professional fields. The method includes the following steps:

[0007] S1: Receive industry decision requests and related information, construct a hierarchical task context based on the industry decision requests and related information. The hierarchical task context is a structured text organized according to a preset hierarchy, which can be dynamically written by subsequent intelligent agents and expose the minimum necessary information required to the intelligent agents participating in the execution.

[0008] S2: The main intelligent agent determines the task path based on the hierarchical task context and generates a task processing plan;

[0009] S3: The main intelligent agent identifies information gaps according to the task processing scheme, and performs data and external knowledge retrieval based on the information gaps to obtain retrieval results presented in the form of structured text entries;

[0010] S4: The main agent constructs at least one expert agent stage based on the hierarchical task context, the task processing scheme and the retrieval results. The expert agent stage is a processing link in which several expert agents perform specialized analysis.

[0011] S5: Execute the expert agent stage, where the expert agent analyzes and obtains the output results, and the main agent merges the output results of the expert agent to generate industry decision results.

[0012] Furthermore, in step S1, the industry decision-making request and related data are processed according to the source, lifecycle, and usage boundaries of the information to generate the hierarchical task context, specifically including:

[0013] A large language model is used to extract key elements from the industry decision-making request to form the current request item;

[0014] Historical dialogues are vectorized through an embedded model and written into a vector index. Using the current industry decision-making request as the query, an approximate nearest neighbor search is performed to recall semantically similar historical dialogues, forming user memory items.

[0015] The uploaded related materials are processed by recursive character segmentation to obtain text blocks. Each text block is then vectorized using an embedding model and written into a vector index to form related material items.

[0016] The large language model summarizes and generates industry object items, task target items, collaboration control items, and result form items based on the current request item, the user memory item, and the associated data item.

[0017] Furthermore, in step S1, differential exposure is performed on the hierarchical task context, specifically including:

[0018] For each agent's task, the large language model is invoked to filter out the minimum necessary information from the hierarchical task context and expose it to the agent.

[0019] After the agent responds, its final output is written into the hierarchical task context.

[0020] Furthermore, in step S2, the task path determination specifically includes:

[0021] Organize the hierarchical task context into prompt words;

[0022] The main intelligent agent evaluates the industry decision request based on the prompt words from three dimensions: task complexity, context completeness, and external knowledge dependence, and obtains the task path determination result as follows:

[0023] When the task complexity is below a preset threshold and the context completeness meets the preset requirements, a direct response path for the main agent is generated.

[0024] When the context completeness does not meet the preset requirements or external knowledge dependencies indicate that external facts need to be supplemented, a retrieval verification path is generated.

[0025] When the task complexity exceeds a preset threshold and domain-specific reasoning is required, an expert agent analysis path is generated.

[0026] When a task requires multi-domain collaborative reasoning, a comprehensive judgment path is generated.

[0027] Based on the determination results of the task path, a structured task processing solution is output, and the task processing solution is recorded in the hierarchical task context.

[0028] Furthermore, the task path determination also includes dynamic rerouting, specifically including:

[0029] After the hierarchical task context is updated, the main intelligent agent re-evaluates the three dimensions of task complexity, context completeness, and external knowledge dependency.

[0030] When the re-evaluation result is inconsistent with the original task path determination result, the main intelligent agent generates a task processing plan based on the re-evaluation result and replaces the original task processing plan.

[0031] The main intelligent agent uses the updated task processing scheme as the basis for the next round of scheduling until the task complexity is lower than a preset threshold and the context completeness meets the preset requirements.

[0032] Furthermore, in step S3, the data and external knowledge retrieval generates the retrieval results, specifically including:

[0033] The main intelligent agent identifies the content that is required by the task processing scheme but is currently missing, as an information gap;

[0034] The main intelligent agent registers the features of the three retrieval sources—related information, knowledge base, and online search—based on three dimensions: data domain, knowledge stability, and timeliness requirements.

[0035] For each information gap, the main agent maps its features with those of each retrieval source in the three dimensions of data domain, knowledge stability, and timeliness requirements into vectors through an embedding model. The cosine similarity between the two vectors is used as the matching degree of the retrieval source, and the retrieval source with the highest matching degree is selected for retrieval.

[0036] The search results are presented in the form of structured text entries, with each result including source type, scope of application, publication time, and hit segment.

[0037] Furthermore, the data and external knowledge retrieval also includes a conflict arbitration operation based on confidence scores, specifically including:

[0038] The large language model is used to break down the hit fragments of each search result into independently verifiable factual statements. The factual statements of different search results are paired up to determine whether the relationship is implied, neutral, or contradictory. The search result pairs to which the factual statements with contradictory relationships belong are marked as conflicting entry pairs.

[0039] For each search result, the large language model is invoked to output a confidence score based on the hit fragment, source type, publication time, and applicable scope;

[0040] When the difference between the highest and second-highest confidence scores in a pair of conflicting entries exceeds a preset arbitration threshold, the search result corresponding to the highest score is retained, and the conflicting search result is annotated as a restriction condition.

[0041] When the highest confidence score in a pair of conflicting items is lower than a preset lower limit, the main agent triggers a supplementary search for the corresponding information gap; if the highest confidence score is still lower than the preset lower limit after a preset number of repeated supplementary searches, the corresponding information gap is marked as a matter to be verified.

[0042] Furthermore, in step S4, the expert agent constituting the expert agent stage is defined by the following technical attributes:

[0043] a. Scope of professional capabilities, which describes the professional identity of the expert agent and the types of reasoning and analysis it can perform;

[0044] b. Toolset, declaring the tools that are allowed to be invoked;

[0045] c. Context exposure boundaries limit the types of hierarchical task context entries that can be accessed.

[0046] Furthermore, the construction of the expert agent phase specifically includes:

[0047] The large language model is invoked to identify the tasks to be executed based on the hierarchical task context; for each task, the embedding model is invoked to convert the professional ability range of each expert agent into an ability vector and the task to be executed into a task vector. Expert agents with matching degree exceeding a preset threshold are selected according to cosine similarity. The large language model then assigns the task to be executed to the expert agent with the highest matching degree to obtain the corresponding relationship.

[0048] The large language model is invoked to generate a directed acyclic graph using the aforementioned correspondence and the hierarchical task context as input. ,in It is a set of task nodes, each node corresponding to a task to be executed and an expert agent responsible for that task. Let be the set of directed edges that define dependencies between nodes. Represents a node Must precede node Completed, among which ,and Each node records the task objective, the responsible expert agent, and the expected task result for that node.

[0049] For the directed acyclic graph Topological sorting yields orderly stage division ,in Each stage The nodes within the node execute in parallel, and Satisfying nodes The stage number is not greater than the node number. The relevant stage number;

[0050] After each stage is completed, the main agent determines whether to initiate the next stage based on the hierarchical task context.

[0051] Furthermore, in step S5, the main intelligent agent generates industry decision results, specifically including:

[0052] The expert agent takes the information exposed by the hierarchical task context as input to generate specific judgments;

[0053] The main intelligent agent merges the specific judgments to generate an industry decision result, which is one of the following: a single-point response, a comprehensive assessment opinion, or a complete industry solution.

[0054] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0055] 1. Layered task context enables structured management of information and minimizes necessary exposure.

[0056] In existing technologies, traditional intelligent agent solutions concatenate all dialogue and data into a single prompt word input model. This not only fails to differentiate and manage information according to its source and lifecycle but also exposes the entire context to each agent, leading to context redundancy, difficulty in effectively extracting key information, and the risk of over-exposure of sensitive information. This invention constructs a hierarchical task context, organizing the current request item, user memory item, associated data item, and summarized industry object item, task objective item, collaborative control item, and result form item in layers. Before each agent executes a task, it filters out the minimum necessary information from the hierarchical task context for exposure. After the agent responds, its output is written back to the hierarchical task context. This ensures the integrity and traceability of the context while ensuring that each agent only accesses the information necessary to complete the task, thus significantly reducing the length of the prompt word and the risk of information leakage.

[0057] 2. Task path determination and dynamic rerouting based on 3D evaluation.

[0058] In existing technologies, traditional solutions often employ fixed processes or one-time task planning, failing to dynamically adjust task paths based on context updates and thus struggling to adapt to the increasingly refined needs in industry decision-making processes. This invention addresses this by having a primary intelligent agent evaluate industry decision requests from three dimensions: task complexity, context completeness, and external knowledge dependence. This generates multiple task paths, including direct response from the primary intelligent agent, retrieval and verification, expert agent analysis, or comprehensive judgment. Upon updates to the hierarchical task context, these three dimensions are re-evaluated. When the re-evaluation result differs from the original task path, the task processing scheme is dynamically updated, using the updated scheme as the basis for the next round of scheduling. Therefore, the task processing scheme can adaptively adjust to changes in information completeness, avoiding processing deviations caused by a single fixed path.

[0059] 3. Multi-source retrieval and feature registration for information gaps.

[0060] In existing technologies, traditional solutions tend to be somewhat haphazard in their selection of retrieval sources when supplementing search data and external knowledge. They often mix information with varying degrees of stability and timeliness requirements, leading to mismatches between search results and the actual gaps. This invention identifies the missing content required by the task processing scheme as the information gap. It then registers the features of three types of retrieval sources—related materials, knowledge bases, and online searches—based on three dimensions: data domain, knowledge stability, and timeliness requirements. For each information gap, it performs semantic matching with each retrieval source across these three dimensions, determining the retrieval source based on the degree of matching and implementing the retrieval. Each information gap is thus directed to the most suitable retrieval source, improving the relevance and effectiveness of the search results.

[0061] 4. Multi-source conflict arbitration based on confidence scores.

[0062] In existing technologies, traditional solutions lack conflict identification and arbitration for multi-source search results, easily incorporating contradictory facts into decision-making, thus affecting the reliability of the decision. This invention breaks down the hit fragments of each search result into independently verifiable factual statements. It pairs these factual statements from different search results, classifying them as implicit, neutral, or contradictory, marking conflicting item pairs. Then, it outputs confidence scores based on the hit fragment, source type, publication time, and applicable scope. When the difference between the highest and second-highest confidence scores in a conflicting item pair exceeds a preset arbitration threshold, the highest score result is retained, and the conflict result is noted as a limiting condition. When the highest confidence score falls below a preset lower limit, a supplementary search is triggered. If the score still fails to meet the threshold after a preset number of supplementary searches, it is marked as a matter to be verified. The interference of contradictory facts on decision-making is thus effectively eliminated, and unverifiable matters are not silently adopted.

[0063] 5. Expert agent orchestration and parallel execution based on directed acyclic graphs.

[0064] In existing technologies, traditional solutions often rely on manually specifying the execution order or simple sequential calls when scheduling multiple expert agents, failing to automatically build dependencies between tasks and resulting in low collaborative efficiency. This invention vectorizes the expertise of expert agents and tasks separately using an embedding model, automatically assigning tasks to expert agents based on cosine similarity. A large language model generates a directed acyclic graph with tasks as nodes and dependencies as directed edges, and topological sorting yields ordered stage divisions. This allows tasks within the same stage to execute in parallel, and cross-stage tasks to execute in a topologically ordered manner based on dependencies. By maximizing parallelism while ensuring dependency correctness, the efficiency of multi-expert agent collaboration is significantly improved.

[0065] 6. Generation of industry decision-making results from hierarchical mergers.

[0066] In existing technologies, traditional solutions offer a single output format, making it difficult to adapt to the varying needs for detail in results across different decision-making scenarios. This invention, however, involves an expert agent generating specific judgments based on information exposed by the hierarchical task context. A main agent then merges these specific judgments to generate one of three industry decision-making results: a single-point response, a comprehensive assessment opinion, or a complete industry solution. This unified collaborative process can therefore output industry decision-making results with different granularities based on the task path determination results, enhancing its adaptability to diverse industry decision-making scenarios. Attached Figure Description

[0067] Figure 1 This is the overall logic flowchart of the method of the present invention. Detailed Implementation

[0068] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0069] This embodiment discloses a multi-agent collaborative generation method for industry decision-making, which is executed collaboratively by a master agent and several expert agents. The master agent is responsible for the overall task arrangement, and the expert agents are responsible for specialized analysis within their respective professional fields. The method includes steps S1 to S5, as follows: Figure 1 As shown below, each step will be further explained. This embodiment uses an industry decision-making request to analyze the investment feasibility and development direction of a specific industry as an example. The relevant data are the industry planning documents and industry statistics uploaded by the decision-maker.

[0070] S1: Receive industry decision requests and related information, and construct a hierarchical task context based on the industry decision requests and related information.

[0071] This embodiment uses a structured state that can be concurrently read and written by multiple agents to carry the hierarchical task context, organized into three categories: message stream, execution trajectory, and file list. The message stream carries the current request, the inputs and outputs of each agent, and the summarized results; the execution trajectory is organized hierarchically from session round to execution node, recording the process of each agent and tool call; the file list registers the metadata of the data items associated with the current session, including filename, summary, and size. The hierarchical task context can be dynamically and concurrently written to by subsequent agents, and the consistency of concurrent writes is ensured by using key-based deduplication and write-back semantics that merge by round and node identifier.

[0072] Furthermore, the industry decision-making requests and related data are processed according to the source, lifecycle, and usage boundaries of the information to generate the hierarchical task context. The specific process is as follows:

[0073] First, a large language model is used to extract key elements from the industry decision-making request to form the current request item. In this embodiment, the main agent filters out internal context messages such as file context and expert context from the message stream, takes the most recent external user message as the current request item, and uses the identifier of the external user message as the round identifier of the current session. The industry object extracted from the above industry decision-making request is the target industry, the decision intention is the analysis of investment feasibility and development direction, and the constraints are the investment scale and implementation cycle. Second, historical dialogues are vectorized and written into a vector index using an embedding model. An approximate nearest neighbor search is performed using the current industry decision-making request as the query to recall semantically similar historical dialogues, forming user memory items. To avoid historical dialogues exceeding the upper limit of the large language model's context window, when the cumulative length of historical messages reaches a preset proportion of this upper limit, the lightweight large language model deepseek-v4-flash used in this embodiment performs structured compression on earlier historical dialogues before vectorizing and writing them into the vector index, while retaining the most recent preset proportion of original messages. Next, the uploaded related materials are processed using a recursive character segmentation method to obtain text blocks. Each text block is then vectorized using an embedding model and written into a vector index to form related material items. In this embodiment, the uploaded heterogeneous format documents, such as industry planning documents, are first uniformly converted into Markdown plain text. Then, a recursive character segmenter is used to recursively segment the text according to the hierarchy of paragraphs, line breaks, spaces, and characters. The block size is set to 1024, and the overlap between adjacent blocks is set to 256. Each text block is vectorized using the 1024-dimensional text embedding model text-embedding-v4 used in this embodiment and written into a vector index. This vector index uses pgvector and is organized using cosine distance. Meta-information such as filename, summary, and size is simultaneously registered as a file list. Finally, based on the current request item, user memory item, and related material item, the large language model summarizes and generates industry object items, task target items, collaboration control items, and result form items, which are then hierarchically organized into structured text and written into the hierarchical task context.

[0074] Furthermore, differentiated exposure is performed on the hierarchical task context. For each agent's task, the minimum necessary information is filtered from the hierarchical task context and exposed to the agent. After the agent responds, its final output is written back to the hierarchical task context. In this embodiment, differentiated exposure is implemented through three levels. First, related information is injected into the main agent only as a list of files as the minimum necessary metadata prompts, and the file text is only recalled as needed when the agent actively initiates a data retrieval. Second, for each expert agent's task, only a tailored context containing the user's original question, the expert agent's task description, and the answers from previous expert agents is constructed, without exposing the main agent's internal thinking or the complete dialogue with other unrelated expert agents. Third, after the expert agent responds, only its final conclusion is written back to the hierarchical task context as a message with an attribution tag; intermediate processes such as tool calls are not written back.

[0075] S2: The main intelligent agent determines the task path based on the hierarchical task context and generates a task processing plan.

[0076] In this embodiment, the main agent organizes the hierarchical task context into prompt words and evaluates the industry decision request from three dimensions: task complexity, context completeness, and external knowledge dependence. When the task complexity is below a preset threshold and the context completeness meets the preset requirements, a direct response path for the main agent is generated. When the context completeness does not meet the preset requirements or external facts need to be supplemented, a retrieval and verification path is generated. When the task complexity is above a preset threshold and requires domain-specific reasoning, an expert agent analysis path is generated. When the task requires multi-domain collaborative reasoning, a comprehensive judgment path is generated. The industry decision request in this embodiment requires multi-domain collaborative reasoning, including resources, technology, industry, market, finance, and risk. Based on this, the main agent generates a comprehensive judgment path, synthesizes the structured task processing solution, and records it in the hierarchical task context.

[0077] Furthermore, the task path determination also includes dynamic rerouting. After the hierarchical task context is updated, the main agent re-evaluates the three dimensions of task complexity, context completeness, and external knowledge dependence. When the re-evaluation result is inconsistent with the original task path determination result, the main agent generates a task processing plan based on the re-evaluation result and replaces the original task processing plan. The updated task processing plan is used as the basis for the next round of scheduling until the task complexity is lower than a preset threshold and the context completeness meets the preset requirements. In this embodiment, the maximum number of expert agent stages within the same session round is set to 8, and expert agents are restricted from scheduling other expert agents. The permissions for task path determination and dynamic rerouting are exclusive to the main agent.

[0078] S3: The main intelligent agent identifies information gaps according to the task processing scheme, and performs data and external knowledge retrieval based on the information gaps to obtain retrieval results presented in the form of structured text entries.

[0079] In this embodiment, the main agent identifies currently missing information in the industry, such as market supply and demand and price trends, key technology maturity and process routes, and similar industry development cases. This information is considered a gap. Based on three dimensions—data domain, knowledge stability, and timeliness requirements—features from three types of retrieval sources: related materials, knowledge bases, and online searches are registered. For each information gap, its features along with those of each retrieval source in the aforementioned three dimensions are mapped to vectors using an embedding model. The cosine similarity between these vectors is used as the matching degree of the retrieval source, and the retrieval source with the highest matching degree is selected for retrieval. A hybrid retrieval method combining vector recall and re-ranking is employed for each type of retrieval source. First, the text-embedding-v4 embedding model used in this embodiment vectorizes the information gap and each text block, and recalls several candidates based on cosine similarity. Then, the gte-rerank-v2 re-ranking model used in this embodiment refines the candidates according to their semantic relevance to the gap, and selects the top few with the highest relevance as the hit results. In this embodiment, the knowledge base retrieval ultimately returns 5 hits, with each hit excerpt not exceeding 800 characters, and the total output not exceeding 6000 characters. The retrieval results are presented in the form of structured text entries, with each retrieval result including source type, applicable scope, publication time, and hit fragment.

[0080] Furthermore, the data and external knowledge retrieval also includes a conflict arbitration operation based on confidence scores. First, a large language model is invoked to break down the hit fragments of each search result into independently verifiable factual statements. For each pair of factual statements from different search results, the relationship is determined as implied, neutral, or contradictory. Search result pairs containing contradictory factual statements are marked as conflicting item pairs. For each search result, the large language model is invoked to output a confidence score based on the hit fragment, source type, publication time, and applicable scope. When the difference between the highest and second-highest confidence scores in a conflicting item pair exceeds a preset arbitration threshold, the search result corresponding to the highest score is retained, and the conflicting search result is annotated as a restriction condition. When the highest confidence score in a conflicting item pair is lower than a preset lower limit, the main agent triggers a supplementary search for the corresponding information gap. After a preset number of repeated supplementary searches, if the highest confidence score is still lower than the preset lower limit, the corresponding information gap is marked as a matter to be verified. In this embodiment, the preset arbitration threshold is 0.15, the preset lower limit is 0.6, and the preset number of searches is 3.

[0081] S4: The main agent constructs at least one expert agent stage based on the hierarchical task context, task processing scheme, and retrieval results.

[0082] In this embodiment, the expert agents constituting the expert agent stage are defined by three technical attributes: professional capability scope, toolset, and context exposure boundary. This embodiment configures eight domain expert agents: resource analyst, technical feasibility expert, industry strategy expert, brand building expert, project implementation expert, financial expert, risk assessment expert, and benefit evaluation expert. Specifically, the resource analyst expert's professional capability scope includes inventorying all industry resources and assessing development potential; the technical feasibility expert's professional capability scope includes feasibility of technical solutions, selection of production processes and equipment, and assessment of technology maturity; the industry strategy expert's professional capability scope includes industry direction selection and justification, supply chain design, and differentiated competitive strategies; the brand building expert's professional capability scope includes target market demand research, brand positioning, and marketing channels; the project implementation expert's professional capability scope includes implementation roadmap, organizational division of labor, milestones, and operational quality management; the financial expert's professional capability scope includes investment estimation, cost-benefit analysis, cash flow, and financing; the risk assessment expert's professional capability scope includes identification and prevention of market, technological, operational, and natural risks; and the benefit evaluation expert's professional capability scope includes comprehensive evaluation of economic, social, ecological, and managerial benefits. Each expert agent shares a toolkit consisting of network search, deep search, web page extraction, image understanding, document retrieval, and knowledge base retrieval tools, and is guided to use the appropriate tools according to its professional focus. Each expert agent's context exposure boundary restricts it from scheduling other expert agents, and after execution, it only returns the execution trajectory and its conclusions to the master agent.

[0083] Furthermore, the construction process of the expert agent phase is as follows:

[0084] First, a large language model is invoked to identify the tasks to be executed based on the hierarchical task context. For each task, an embedding model is invoked to convert the professional capabilities of each expert agent into capability vectors and the task to be executed into a task vector. Expert agents with a matching degree exceeding a preset threshold are selected based on cosine similarity. In this embodiment, the preset threshold is set to 0.7. Then, the large language model assigns the task to be executed to the expert agent with the highest matching degree, obtaining a correspondence. For example, the industry market demand analysis task is assigned to the brand building expert based on cosine similarity, and the technical feasibility and process route analysis task is assigned to the technical feasibility expert. Then, the large language model is invoked to generate a directed acyclic graph using the correspondence and the hierarchical task context as input. ,in It is a set of task nodes, each node corresponding to a task to be executed and an expert agent responsible for that task. Let be the set of directed edges that define dependencies between nodes. Represents a node Must precede node Completed, among which ,and Each node records the task objective, the responsible expert agent, and the expected task result for that node; for the directed acyclic graph... Topological sorting yields orderly stage division ,in Each stage Nodes within the same node execute in parallel, and cross-stage execution is topologically ordered based on dependencies. Satisfying nodes The stage number is not greater than the node number. Phase number. In this embodiment, the three tasks undertaken by the resource analyst expert, the technical feasibility expert, and the industry strategy expert are independent of each other and are classified as Phase 1. Parallel execution. The tasks undertaken by financial experts, risk assessment experts, and benefit evaluation experts, which depend on the conclusions of previous stages, are classified as the second stage. Execution. After each stage is completed, the main agent determines whether to initiate the next stage based on the hierarchical task context.

[0085] S5: Execute the expert agent stage, where the expert agent analyzes and obtains the output results, and the main agent merges the output results of the expert agent to generate industry decision results.

[0086] In this embodiment, the expert agent takes the information provided by the differentiated exposure of the hierarchical task context as input, calls the toolset for analysis, generates specific judgments, and writes them back to the hierarchical task context with a message marked with attribution. For example, a brand building expert produces a specific judgment on the market demand and competitive landscape of the industry, a technical feasibility expert produces a specific judgment on technical solutions and process routes, and a financial expert produces a specific judgment on investment estimates and cost-benefit analysis. When merging the specific judgments, the main agent follows the integration principles of conclusion first, avoiding mechanical splicing of expert agent judgments, resolving disagreements between expert agents, and summarizing risk warnings at the end. It integrates the specific judgments of each expert agent into a logically coherent natural language response, generating an industry decision result. The industry decision result can be one of three forms: a single-point response, a comprehensive assessment opinion, or a complete industry plan. In this embodiment, because the task path is a comprehensive assessment path, the main agent merges the specific judgments of each expert agent, ultimately generating an industry decision result in the form of a complete industry plan, including development feasibility conclusions, key direction suggestions, and risk warnings.

[0087] Each step in the aforementioned multi-agent collaborative generation method for industry decision-making can be implemented entirely or partially through software, hardware, or a combination thereof. These steps can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can invoke and execute the operations corresponding to each step.

[0088] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media containing computer-usable program code, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0089] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the patent of the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A multi-agent collaborative generation method for industrial decision-making, characterized in that, The method is executed collaboratively by a main intelligent agent and several expert intelligent agents. The main intelligent agent is responsible for the overall arrangement of tasks, while the expert intelligent agents are responsible for specialized analysis within their respective professional domains. The method includes the following steps: S1: Receive industry decision requests and related information, construct a hierarchical task context based on the industry decision requests and related information. The hierarchical task context is a structured text organized according to a preset hierarchy, which can be dynamically written by subsequent intelligent agents and expose the minimum necessary information required to the intelligent agents participating in the execution. S2: The main intelligent agent determines the task path based on the hierarchical task context and generates a task processing plan; S3: The main intelligent agent identifies information gaps according to the task processing scheme, and performs data and external knowledge retrieval based on the information gaps to obtain retrieval results presented in the form of structured text entries; S4: The main agent constructs at least one expert agent stage based on the hierarchical task context, the task processing scheme and the retrieval results. The expert agent stage is a processing link in which several expert agents perform specialized analysis. S5: Execute the expert agent stage, where the expert agent analyzes and obtains the output results, and the main agent merges the output results of the expert agent to generate industry decision results.

2. The multi-agent collaborative generation method for industrial decision-making according to claim 1, characterized in that, In step S1, the industry decision-making request and related data are processed according to the source, lifecycle, and usage boundaries of the information to generate the hierarchical task context, specifically including: A large language model is used to extract key elements from the industry decision-making request to form the current request item; Historical dialogues are vectorized through an embedded model and written into a vector index. Using the current industry decision-making request as the query, an approximate nearest neighbor search is performed to recall semantically similar historical dialogues, forming user memory items. The uploaded related materials are processed by recursive character segmentation to obtain text blocks. Each text block is then vectorized using an embedding model and written into a vector index to form related material items. The large language model summarizes and generates industry object items, task target items, collaboration control items, and result form items based on the current request item, the user memory item, and the associated data item.

3. The multi-agent collaborative generation method for industrial decision-making according to claim 2, characterized in that, In step S1, differential exposure is performed on the hierarchical task context, specifically including: For each agent's task, the large language model is invoked to filter out the minimum necessary information from the hierarchical task context and expose it to the agent. After the agent responds, its final output is written into the hierarchical task context.

4. The multi-agent collaborative generation method for industrial decision-making according to claim 3, characterized in that, In step S2, the task path determination specifically includes: Organize the hierarchical task context into prompt words; The main intelligent agent evaluates the industry decision request based on the prompt words from three dimensions: task complexity, context completeness, and external knowledge dependence, and obtains the task path determination result as follows: When the task complexity is below a preset threshold and the context completeness meets the preset requirements, a direct response path for the main agent is generated. When the context completeness does not meet the preset requirements or external knowledge dependencies indicate that external facts need to be supplemented, a retrieval verification path is generated. When the task complexity exceeds a preset threshold and domain-specific reasoning is required, an expert agent analysis path is generated. When a task requires multi-domain collaborative reasoning, a comprehensive judgment path is generated. Based on the determination results of the task path, a structured task processing solution is output, and the task processing solution is recorded in the hierarchical task context.

5. The multi-agent collaborative generation method for industrial decision-making according to claim 4, characterized in that, The task path determination also includes dynamic rerouting, specifically including: After the hierarchical task context is updated, the main intelligent agent re-evaluates the three dimensions of task complexity, context completeness, and external knowledge dependency. When the re-evaluation result is inconsistent with the original task path determination result, the main intelligent agent generates a task processing plan based on the re-evaluation result and replaces the original task processing plan. The main intelligent agent uses the updated task processing scheme as the basis for the next round of scheduling until the task complexity is lower than a preset threshold and the context completeness meets the preset requirements.

6. The multi-agent collaborative generation method for industrial decision-making according to claim 5, characterized in that, In step S3, the data and external knowledge retrieval generate the retrieval results, specifically including: The main intelligent agent identifies the content that is required by the task processing scheme but is currently missing, as an information gap; The main intelligent agent registers the features of the three retrieval sources—related information, knowledge base, and online search—based on three dimensions: data domain, knowledge stability, and timeliness requirements. For each information gap, the main agent maps its features with those of each retrieval source in the three dimensions of data domain, knowledge stability, and timeliness requirements into vectors through an embedding model. The cosine similarity between the two vectors is used as the matching degree of the retrieval source, and the retrieval source with the highest matching degree is selected for retrieval. The search results are presented in the form of structured text entries, with each result including source type, scope of application, publication time, and hit segment.

7. The multi-agent collaborative generation method for industrial decision-making according to claim 6, characterized in that, The data and external knowledge retrieval also includes conflict arbitration based on confidence scores, specifically including: The large language model is used to break down the hit fragments of each search result into independently verifiable factual statements. The factual statements of different search results are paired up to determine whether the relationship is implied, neutral, or contradictory. The search result pairs to which the factual statements with contradictory relationships belong are marked as conflicting entry pairs. For each search result, the large language model is invoked to output a confidence score based on the hit fragment, source type, publication time, and applicable scope; When the difference between the highest and second-highest confidence scores in a pair of conflicting entries exceeds a preset arbitration threshold, the search result corresponding to the highest score is retained, and the conflicting search result is annotated as a restriction condition. When the highest confidence score in a pair of conflicting items is lower than a preset lower limit, the main agent triggers a supplementary search for the corresponding information gap; if the highest confidence score is still lower than the preset lower limit after a preset number of repeated supplementary searches, the corresponding information gap is marked as a matter to be verified.

8. The multi-agent collaborative generation method for industrial decision-making according to claim 7, characterized in that, In step S4, the expert agent constituting the expert agent phase is defined by the following technical attributes: a. Scope of professional capabilities, which describes the professional identity of the expert agent and the types of reasoning and analysis it can perform; b. Toolset, declaring the tools that are allowed to be invoked; c. Context exposure boundaries limit the types of hierarchical task context entries that can be accessed.

9. A multi-agent collaborative generation method for industrial decision-making according to claim 8, characterized in that, The construction of the expert agent phase specifically includes: The large language model is invoked to identify the tasks to be executed based on the hierarchical task context; for each task, the embedding model is invoked to convert the professional ability range of each expert agent into an ability vector and the task to be executed into a task vector. Expert agents with matching degree exceeding a preset threshold are selected according to cosine similarity. The large language model then assigns the task to be executed to the expert agent with the highest matching degree to obtain the corresponding relationship. The large language model is invoked to generate a directed acyclic graph using the aforementioned correspondence and the hierarchical task context as input. ,in It is a set of task nodes, each node corresponding to a task to be executed and an expert agent responsible for that task. Let be the set of directed edges that define dependencies between nodes. Represents a node Must precede node Completed, among which ,and Each node records the task objective, the responsible expert agent, and the expected task result for that node. For the directed acyclic graph Topological sorting yields orderly stage division ,in Each stage The nodes within the node execute in parallel, and Satisfying nodes The stage number is not greater than the node number. The relevant stage number; After each stage is completed, the main agent determines whether to initiate the next stage based on the hierarchical task context.

10. A multi-agent collaborative generation method for industrial decision-making according to claim 9, characterized in that, In step S5, the main intelligent agent generates industry decision results, specifically including: The expert agent takes the information exposed by the hierarchical task context as input to generate specific judgments; The main intelligent agent merges the specific judgments to generate an industry decision result, which is one of the following: a single-point response, a comprehensive assessment opinion, or a complete industry solution.

Images