Knowledge enhancement generative model implementation method based on multi-agent filtering mechanism

Abstract

The invention discloses a knowledge enhancement generative model implementation method based on a multi-agent filtering mechanism. The knowledge enhancement generative model implementation method comprises the steps of document retrieval and subject-based agent initialization; performing multi-agent parallel generation and viewpoint expression; a multi-round reflection and evaluation filtering mechanism; and dynamic agent merging and knowledge refining, and final answer generation. According to the method, task specific fine tuning is not needed, information filtering and integration are carried out through multi-agent cooperation, and the method aims at improving the accuracy, reliability and cross-domain adaptability of a large language model when the large language model processes knowledge-intensive tasks. According to the method, a large number of retrieved documents are regarded as an information market, high-quality information stands out through a multi-agent competition and cooperation mechanism, and meanwhile interference of noise and error information is restrained. According to the method, a bottom layer large language model does not need to be finely adjusted aiming at a specific task, and strong cross-task and cross-domain adaptability can be realized only by utilizing the inherent language understanding and generating capability of the bottom layer large language model.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, and refers to a knowledge enhancement generation model technology, specifically a knowledge enhancement generation model based on a multi-agent filtering mechanism and its implementation method. Background Technology

[0002] Large-scale pre-trained language models have demonstrated exceptional capabilities in natural language understanding and generation tasks. However, inherent limitations of these models, such as knowledge deadlines, the potential for "illusions," and their inability to handle highly specialized or up-to-date information, restrict their direct application in critical scenarios. To alleviate these issues, Retrieval-Augmented Generation (RAG) techniques have emerged. The core idea is to retrieve relevant information from external knowledge sources and provide it as context to the LLM (Language Modeling) to generate more factually grounded answers.

[0003] While the traditional RAG framework has achieved some success, its effectiveness heavily relies on the quality of the search results. When faced with a complex query, the retrieval system often returns a large number of relevant but inconsistent documents with potentially conflicting viewpoints. The traditional method of directly cramming all search results into a limited context window has significant drawbacks: First, noisy or irrelevant information dilutes the concentration of key evidence, interfering with the model's judgment; second, the model may fail to effectively weigh the credibility of information from different sources, leading to the generated content being biased by low-quality information; finally, for complex problems requiring the integration of multi-perspective information, simple context splicing is insufficient to achieve deep information fusion and reasoning.

[0004] Therefore, existing technologies urgently need a more intelligent and robust intermediate processing layer that can effectively denoise, filter, weigh, and integrate the massive amounts of retrieved information, and then provide the refined high-quality knowledge to the generative model, thereby significantly improving the quality of the final output. Summary of the Invention

[0005] The purpose of this invention is to overcome the aforementioned shortcomings of existing retrieval enhancement generation technologies and provide a knowledge enhancement generation model and its implementation method based on a multi-agent filtering mechanism. This invention introduces a multi-agent collaborative framework that simulates the working mode of an "expert team." Through steps such as division of labor, debate, evaluation, and merging, it achieves automated and refined processing of retrieved information, ultimately generating accurate, comprehensive, and reliable content.

[0006] The core idea of ​​this invention is to treat the large number of retrieved documents as an information market. Through a multi-agent competition and cooperation mechanism, high-quality information is allowed to stand out, while noise and erroneous information are suppressed. This method does not require fine-tuning of the underlying large language model for specific tasks; it can achieve strong cross-task and cross-domain adaptability simply by utilizing its inherent language understanding and generation capabilities.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for implementing a knowledge-enhanced generative model based on a multi-agent filtering mechanism includes the following steps: Step 1: Document retrieval and topic-based agent initialization; First, in response to user queries, a set of relevant documents is retrieved from one or more external knowledge sources. Then, based on the semantic content of these documents, the K-means clustering algorithm is used to divide them into several topic clusters. Each topic cluster represents a relatively independent knowledge perspective or subtopic. Next, a dedicated agent, namely the domain agent (the core of the agent is a finely tuned Qwen3-8b), is initialized for each topic cluster. Each domain agent is given all the documents of the corresponding topic cluster as its exclusive knowledge background.

[0008] Step 2: Parallel generation of multiple agents and expression of viewpoints; After initialization, the system submits user queries to each domain agent in parallel. Each domain agent independently generates a local answer to the query based on its assigned subset of topic-filtered documents. The local answer represents a preliminary viewpoint or answer derived from the agent's specific perspective and knowledge background.

[0009] Step 3: Multi-round reflection and evaluation filtering mechanism. The system introduces an independent, neutral "evaluator" (with Qwen2.5-72b as the core) to conduct multiple rounds of evaluation on all local responses generated in step two. The evaluator does not directly generate answers, but is responsible for critically questioning, comparing and evaluating the outputs of agents in each domain. Based on the evaluation results, the evaluator will identify domain agents with low credibility, high noise or obvious errors in their answers and mark them as weak agents.

[0010] The evaluation criteria include: consistency with the evidence in the relevant document, logical rigor, degree of conformity with generally accepted facts, and the absence of speculative content. The evaluation results are presented using a scoring system, with evaluators rating from 0 to 10. Higher scores indicate that the domain agent's response is closer to the referenced document fragment and therefore more credible. Step 4: Dynamic agent merging and knowledge refinement. Based on the evaluator's assessment, the system will not directly delete weak agents. Instead, it will merge them with one or more strong agents with high credibility. The merging process is essentially a reorganization and refinement of knowledge: the strong agent absorbs the document set of the weak agent and, based on this expanded but filtered knowledge set, regenerates a more comprehensive and robust local answer. This "merge-regenerate" process will undergo multiple iterations, much like a workshop that repeatedly refines viewpoints. Each round eliminates the least credible viewpoints and strengthens the most reliable ones.

[0011] Step 5: Final answer generation. When the preset number of rounds is reached, or when the evaluator deems the quality of the remaining agents' responses to be satisfactory and stable, the system integrates the partial responses from one or more (usually merged into a very small number) "super agents." This integration can be summarized by a master agent, or the final merged answer can be directly output. The result is a high-quality final answer that has undergone multiple rounds of critical review and optimization.

[0012] The beneficial effects of this invention are as follows: 1. Significantly improves generation quality: Through competition and cooperation among multiple agents, noise and unreliable information in the retrieval results are effectively filtered out, resulting in a significant improvement in the factual accuracy and logical consistency of the generated answers.

[0013] 2. Strong anti-interference capability: Even if the search results contain some errors or irrelevant information, the system can minimize their impact through evaluation and merging mechanisms.

[0014] 3. No task fine-tuning required: The entire process is built on the natural language capabilities of the pre-trained language model, eliminating the need for costly fine-tuning for specific tasks. It is highly versatile and easy to deploy.

[0015] 4. Good interpretability: The multi-agent workflow provides a record similar to a "decision trajectory", which helps to understand the formation process of the final answer and enhances the interpretability of the model. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. This embodiment takes answering complex questions in a professional question-and-answer system as an example.

[0018] As shown in the attached diagram, the system receives user queries; Documents are retrieved from external knowledge sources, and then clustered based on semantics, grouping documents with similar content into the same cluster.

[0019] Initialize the agent: The system initializes a domain agent for each document cluster, and each agent is specifically responsible for processing information of a document cluster on a particular topic. Each domain agent independently generates local responses based on its document clusters; An independent evaluator agent assesses the quality of all local responses; The process enters a critical decision point to determine whether to terminate, assessing whether the termination conditions are met. If "yes" (condition met): then directly integrate the remaining agent responses and output the final answer; If "No" (condition not met): then proceed to the iterative optimization loop: a. Merging Agents: The system will identify and merge weak agents into strong agents, which is equivalent to reorganizing and simplifying the agents and the document clusters they represent.

[0020] b. Regenerating the answer: Next, the merged strong agent regenerates the answer based on the new document set, that is, it generates a new local answer based on the merged, more comprehensive document cluster.

[0021] c. Return to Evaluation: The process returns to step 5, where the evaluator agent evaluates all local responses and performs the judgment in step 6 again. This loop continues until the evaluation result meets the termination condition.

[0022] End: When the termination condition is met, the system integrates all answers and outputs the final answer, and the process ends.

[0023] Example: The system's question-and-answer interface displays the question, "What indirect impacts might quantitative easing monetary policy have on global climate change?". The system selects the multi-agent mode and begins the following reasoning: 1. Document retrieval and clustering: After receiving a user query, the system retrieves dozens of relevant documents from external knowledge sources such as financial databases, academic paper databases, and news archives. It then uses a text embedding model to vectorize the documents and performs cluster analysis. The clustering is assumed to result in four relatively clear topic clusters: Cluster 1 (Financial Mechanisms): This section mainly discusses how quantitative easing leads to capital flows and affects the financing costs of green industries.

[0024] Cluster 2 (Macroeconomics): Focuses on the relationship between inflation, economic growth, and energy consumption.

[0025] Cluster 3 (Policy Coordination): This involves the interaction and conflict between central bank policies and climate change policies.

[0026] Cluster 4 (Social Impact): This includes some highly speculative viewpoints, such as changes in public consumption behavior.

[0027] The system initializes four domain agents accordingly: Agent_Finance, Agent_Macro, Agent_Policy, and Agent_Society.

[0028] 2. Parallel generation of local answers: The four agents each generated preliminary answers based on documents related to their respective topics. Agent_Finance's answer focused on capital channels; Agent_Macro emphasized the emission pressures brought about by economic growth; Agent_Policy discussed policy effectiveness; and Agent_Society's answer contained more uncertain expressions such as "maybe" and "perhaps".

[0029] 3. Multiple rounds of reflection and filtering: The first round of evaluation: The evaluator agent (Qwen2.5-72b) evaluated the four answers. It scored them based on four aspects: consistency with the evidence in the relevant documents, logical rigor, adherence to accepted facts, and the presence of speculative content. It indicated that Agent_Society's answer had weak evidence, contained too much speculative content, and had the lowest credibility.

[0030] The first round of merging: The system merged Agent_Society into Agent_Macro, which has a relatively neutral perspective and solid evidence. Agent_Macro now has the documents from Agent_Society and, based on this larger but more relevant set of documents, regenerates a more comprehensive answer that includes discussions of social impacts but removes the speculative parts of the original Agent_Society answer.

[0031] Second round of evaluation: The evaluator re-evaluates the three remaining answers (Agent_Finance, Agent_Macro (new), Agent_Policy). It may find that Agent_Policy and Agent_Macro have some overlap in their arguments but different perspectives.

[0032] Second round of merging: To create a more unified perspective, the evaluators recommended merging Agent_Policy into Agent_Macro. Agent_Macro further integrates knowledge, generating a high-level answer that incorporates macroeconomic, social impact, and policy dimensions.

[0033] 4. Generate the final answer: After two rounds of iteration, the system was left with two powerful agents: Agent_Finance and Agent_Macro (super). The evaluator considered their answers to be of high quality and complementary, thus satisfying the termination condition.

[0034] The system instructs a generator agent to write a comprehensive report as the final output, based on the final local responses from Agent_Finance and Agent_Macro (super). This report systematically elucidates the indirect impacts of quantitative easing on climate change through financial, macroeconomic, and policy interaction channels, providing detailed information and reliable evidence.

[0035] As can be seen from the above embodiments, the present invention successfully simulates an efficient expert discussion process, automatically completes the transformation from information overload to knowledge refinement, and finally outputs high-quality results, fully demonstrating the effectiveness and advancement of the method.

Claims

1. A method for implementing a knowledge-enhanced generative model based on a multi-agent filtering mechanism, characterized in that, Includes the following steps: Step 1: Document retrieval and topic-based agent initialization; Step 2: Parallel generation of multiple agents and expression of viewpoints; Step 3: Multi-round reflection and evaluation filtering mechanism; Step 4: Dynamic agent merging and knowledge refinement; Step 5: Answer generation.

2. The method for implementing a knowledge-enhanced generative model based on a multi-agent filtering mechanism according to claim 1, characterized in that, The specific steps for document retrieval and topic-based agent initialization are as follows: First, in response to user queries, a set of relevant documents is retrieved from one or more external knowledge sources. Then, based on the semantic content of these documents, the K-means clustering algorithm is used to divide them into several topic clusters. Each topic cluster represents a relatively independent knowledge perspective or subtopic. Next, an agent is initialized for each topic cluster. The core of the agent is a fine-tuned Qwen3-8b, i.e., a domain agent. Each domain agent is given all the documents of the corresponding topic cluster as its exclusive knowledge background.

3. The method for implementing a knowledge-enhanced generative model based on a multi-agent filtering mechanism according to claim 1, characterized in that, The specific steps for parallel generation of multiple agents and expression of opinions are as follows; After initialization, the system submits user queries to each domain agent in parallel. Each domain agent independently generates a local answer to the query based on its assigned subset of topic-filtered documents. The local answer represents a preliminary viewpoint or answer derived from the agent's specific perspective and knowledge background.

4. The method for implementing a knowledge-enhanced generative model based on a multi-agent filtering mechanism according to claim 1, characterized in that, The specific steps of the multi-round reflection and evaluation filtering mechanism are as follows: The system introduces a core evaluator, Qwen2.5-72b, to evaluate all local responses generated in step two in multiple rounds. The evaluator does not directly generate answers, but is responsible for critically questioning, comparing and evaluating the outputs of agents in each domain. Based on the evaluation results, the evaluator will identify domain agents with low credibility, high noise or obvious errors in their answers and mark them as weak agents.

5. The method for implementing a knowledge-enhanced generative model based on a multi-agent filtering mechanism according to claim 4, characterized in that, The evaluation criteria include: consistency with the evidence in the relevant document, logical rigor, degree of conformity with accepted facts, and the presence of speculative content. The evaluation results are presented in the form of a score, which is scored by the evaluator from 0 to 10. The higher the score, the closer the domain agent's answer is to the referenced document fragment, and the more credible it is.

6. The method for implementing a knowledge-enhanced generative model based on a multi-agent filtering mechanism according to claim 1, characterized in that, The specific steps of dynamic agent merging and knowledge refinement are as follows: Based on the evaluator's judgment, the system will not directly delete the weak agent, but will merge it with one or more strong agents with high credibility. The merging process is essentially a reorganization and refinement of knowledge: the strong agent absorbs the document set of the weak agent, and based on this expanded but filtered knowledge set, regenerates a more comprehensive and robust local answer.

7. The method for implementing a knowledge-enhanced generative model based on a multi-agent filtering mechanism according to claim 1, characterized in that, The specific steps for generating the final answer are as follows: When the preset number of rounds is reached or the evaluator deems the quality of the remaining agents' responses to be up to standard and stable, the system integrates the remaining one or more, usually merged into a very small number of super agents' local responses, to obtain a high-quality final answer that has undergone multiple rounds of critical testing and optimization.

8. The method for implementing a knowledge-enhanced generative model based on a multi-agent filtering mechanism according to claim 7, characterized in that, The integration can be summarized by a master agent, or the final merged answer can be directly output.

Images