A Multi-Agent Collaborative Retrieval System and Method Based on Large-Model Causal Verification

Abstract

This invention discloses a multi-agent collaborative retrieval system and method based on large-model causal verification, comprising: an intent parsing agent parsing an input natural language query command into an initial causal graph; a retrieval agent performing candidate document retrieval on the initial causal graph, and constructing explanatory power scores and evidence based on causal edge weights to form a causal graph spectrum; an inference agent generating a reverse causal graph based on the causal graph spectrum; a verification agent performing logical checks on the reverse causal graph and the causal graph spectrum, triggering a rollback mechanism if the check result is a logical conflict; a feedback agent acquiring the user's explicit and implicit behaviors towards the retrieval results in real time, and dynamically updating the causal edge weights in the causal graph spectrum based on the explicit and implicit behaviors; and the retrieval agent collaboratively performing retrieval correction on the causal graph spectrum based on replanning instructions and the updated causal edge weights. This invention, through multi-agent collaboration, solves the problems of logical inconsistency and delayed knowledge updates in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, and in particular to a multi-agent collaborative retrieval system and method based on large-scale model causal verification. Background Technology

[0002] With the widespread adoption of Large Language Models (LLMs), Retrieval Augmentation (RAG) technology has become a mainstream paradigm for addressing model illusion and supplementing external knowledge. Early RAG systems primarily relied on vector similarity for document retrieval, calculating the cosine similarity between the query and the document in the semantic space to find relevant information. While these techniques improved retrieval relevance to some extent, in practical industrial applications, the retrieval granularity was coarse, leading to often ambiguous and impractical recommendations. Therefore, improving the accuracy of retrieval results has become a pressing issue.

[0003] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this invention is to provide a multi-agent collaborative retrieval system and method based on large-model causal verification, aiming to address the technical problem of how to improve the accuracy of retrieval results.

[0005] To achieve the above objectives, this invention provides a multi-agent collaborative retrieval system based on large-model causal verification. The system includes an intent-parsing agent, a retrieval agent, a reasoning agent, a verification agent, and a feedback agent. The intent parsing intelligent agent is used to parse the natural language query command input by the user into an initial cause-effect graph, wherein the natural language query command is an IT operation and maintenance knowledge query command; The retrieval agent is used to retrieve candidate documents from the initial causal graph, and to perform explanatory power scoring and evidence construction on the retrieved candidate documents based on causal edge weights to form a causal graph. The reasoning agent is used to generate a reverse causal graph based on the causal graph. The verification agent is used to perform logical detection on the reverse causal graph and the causal graph spectrum. If the detection result is a logical conflict, a rollback mechanism is triggered and a replanning instruction is generated. The feedback agent is used to acquire the user's explicit and implicit behaviors toward the search results in real time, and dynamically update the causal edge weights in the causal graph based on the explicit and implicit behaviors. The retrieval agent is also used to perform retrieval correction on the causal graph based on the replanning instruction and the updated causal edge weights.

[0006] Optionally, the system further includes a central orchestrator; The central orchestrator is used to invoke the intent parsing agent, retrieval agent, reasoning agent, verification agent, and feedback agent.

[0007] Optionally, the user-input natural language query command is parsed into an initial causal graph, including: Identify an entity set from the user's natural speech query command, and classify each entity in the entity set according to variables; Using a large language model as a proxy for a sequence-driven structural equation model, the dependency direction between variables after classification is determined based on conditional probability through the large language model, in order to construct an initial causal graph.

[0008] Optionally, candidate document retrieval is performed on the initial causal graph, including: Determine causal paths based on the initial causal graph; The causal path is converted into a graph database query statement to obtain the entity attributes and associated triplet text on the causal path; A query vector is generated based on the entity attributes and the associated triplet text, and candidate documents are retrieved from the semantic recall search range using the approximate nearest neighbor search method based on the query vector.

[0009] Optionally, the explanatory power score and evidence construction of the retrieved candidate documents are performed based on the causal edge weights, including: Based on causal edge weights, a multi-dimensional scoring formula is used to score the explanatory power of the retrieved candidate documents to obtain the explanatory power score corresponding to each candidate document. Target documents are selected based on the explanatory power scores of each candidate document according to preset filtering rules; The target document is used as an attribute node and attached to the causal path corresponding to the initial causal graph to construct evidence.

[0010] Optionally, generating a reverse causal graph based on the causal graph includes: Based on the causal graph, natural language response text is generated through a controlled generation task; Triples are extracted from the natural language response text using the RE model, and a reverse causal graph is generated based on the triples.

[0011] Optionally, a replanning instruction is generated, including: Based on the aforementioned rollback mechanism, a negative constraint instruction is generated to display the constraint. A replanning instruction is generated based on the negative constraint instruction and the natural language query instruction.

[0012] Furthermore, to achieve the above objectives, this invention also proposes a multi-agent collaborative retrieval method based on large-model causal verification, the method comprising the following steps: The intent parsing agent parses the user's natural language query command into an initial causal graph; The initial causal graph is retrieved by a retrieval agent, and the retrieved candidate documents are scored for explanatory power and evidence is constructed based on the causal edge weights to form a causal graph. The reasoning agent generates a reverse causal graph based on the causal graph; The intelligent agent performs logical checks on the reverse causal graph and the causal graph spectrum. If the check result is a logical conflict, a rollback mechanism is triggered. The system obtains the user's explicit and implicit behaviors toward the search results in real time through a feedback agent, and dynamically updates the causal edge weights in the causal graph based on these behaviors. The retrieval agent performs retrieval and correction of the causal graph based on the replanning instruction and the updated causal edge weights.

[0013] This invention designs a multi-agent collaborative retrieval system based on large-model causal verification. The system includes an intent parsing agent, a retrieval agent, an inference agent, a verification agent, and a feedback agent. The intent parsing agent parses the natural language query command input by the user into an initial causal graph. The retrieval agent retrieves candidate documents from the initial causal graph and, based on the causal edge weights, scores the explanatory power of the retrieved candidate documents and constructs evidence to form a causal graph spectrum. The inference agent generates a reverse causal graph based on the causal graph spectrum. The verification agent performs logical checks on the reverse causal graph and the causal graph spectrum. If the check result is a logical conflict, a rollback mechanism is triggered, and a replanning instruction is generated. The feedback agent acquires the user's explicit and implicit behaviors in response to the retrieval results in real time and dynamically updates the causal edge weights in the causal graph spectrum based on the explicit and implicit behaviors. The retrieval agent collaboratively performs retrieval correction on the causal graph spectrum based on the replanning instruction and the updated causal edge weights. This invention, through multi-agent collaboration, not only enables the instant construction of causal graphs for specific queries, but also introduces a "verification-rollback" closed-loop control flow and a weight update mechanism based on a "causal multi-armed slot machine," thereby fundamentally solving the problems of logical inconsistency and knowledge update lag in existing technologies, and thus improving the accuracy of retrieval results. Attached Figure Description

[0014] Figure 1 This is a structural block diagram of the first embodiment of the multi-agent collaborative retrieval system based on large model causal verification of the present invention; Figure 2 This is a flowchart illustrating the first embodiment of the multi-agent collaborative retrieval method based on large-model causal verification of the present invention.

[0015] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0017] Reference Figure 1 , Figure 1 This is a structural block diagram of the first embodiment of the multi-agent collaborative retrieval system based on large model causal verification of the present invention.

[0018] It is easy to understand that the executing entity of this embodiment can be a multi-agent collaborative retrieval system based on large model causal verification with functions such as data processing, network communication and program execution, or other computer devices with similar functions. This embodiment does not limit it.

[0019] It should also be noted that this system adopts a distributed multi-agent architecture, consisting of a central orchestrator and five functional agents. The components communicate through a standardized JSON bus.

[0020] like Figure 1 As shown, the multi-agent collaborative retrieval system based on large-model causal verification proposed in this embodiment of the invention includes: a central orchestrator, an intent parsing agent 1001, a retrieval agent 1002, a reasoning agent 1003, a verification agent 1004, and a feedback agent 1005.

[0021] In its implementation, the central orchestrator is responsible for task scheduling, state management, and message routing. It maintains a global context object that records the current query state, iteration count, and rollback history.

[0022] State machine definition: IDLE: Waiting for user input.

[0023] PARSING: Invokes the intent parsing agent.

[0024] RETRIEVING: Invokes the retrieval agent.

[0025] GENERATING: Invokes the reasoning agent.

[0026] VERIFYING: Invokes the verification agent.

[0027] ROLLBACK: Triggers rollback logic and updates constraints.

[0028] UPDATING: Invokes the feedback agent to update the weights.

[0029] The intent parsing agent 1001 is used to parse the natural language query command input by the user into an initial causal graph.

[0030] Natural language query commands can be IT operations and maintenance knowledge query commands, or medical knowledge query commands, etc., but this embodiment does not limit them.

[0031] It should be noted that the model selection for the Intent-Causal Agent is: an instruction-tuned LLM (such as Llama-3-70B or GPT-4).

[0032] Prompt Design: Employs Few-Shot Prompting and includes a boot sample of SD-SCM.

[0033] Input: "Server response is slow, and CPU usage fluctuates wildly; restarting has no effect." Task:Identify variables and causal relations. Output in JSON DAGformat. Reasoning: 1.Identify Variables: V1(Response Time), V2(CPU Usage), V3(Reboot),V4(Ineffective). 2.Implicit Causal Inference: Reboot(V3)->Ineffective(V4) implies V3is NOT the cause of V1 / V2. 3.Hypothesis Generation: Hidden_H1(Memory Leak)->V1&V2? Hidden_H2(Deadlock)->V1&V2? Output:Nodes: [V1, V2, V3, V4, H1, H2], Edges: [H1->V1, H1->V2, V3- / ->V1] Furthermore, the process of parsing the user's input natural language query command into an initial causal graph is as follows: identify the entity set from the user's input natural language query command, and classify the variables of each entity in the entity set; use the large language model as a proxy for the sequence-driven structural equation model, and determine the dependency direction between the classified variables based on conditional probability through the large language model in order to construct the initial causal graph.

[0034] In the specific implementation, the query commands are derived from the user's natural language input. Identifying key entity sets It categorizes variables into treatment, outcome, mediator, and confounder, and uses a large language model as a proxy for sequence-driven structural equation modeling (SD-SCM) based on conditional probability. Determine the dependencies between variables and construct an initial query graph (i.e., an initial causal graph). ,in It contains directed edges with initial confidence levels.

[0035] The intent resolution agent needs to send the generated initial causal graph to the central orchestrator so that the central orchestrator can invoke the retrieval agent to perform retrieval processing.

[0036] The retrieval agent 1002 is used to retrieve candidate documents from the initial causal graph, and to perform explanatory power scoring and evidence construction on the retrieved candidate documents based on causal edge weights to form a causal graph.

[0037] The retrieval agent is constructed using a dual-path index structure (semantic index and causal index). The semantic index stores document blocks based on a dense vector (such as OpenAI text-embedding-3). The causal index stores extracted causal triples and their source document IDs based on a graph database (such as Neo4j).

[0038] Before the system executes online retrieval tasks, the domain knowledge base is pre-processed offline to build a dual-path index underlying layer with accurate "content-structure" mapping: The original unstructured document collection (such as operation and maintenance manuals, medical guidelines, etc.) is segmented into semantically complete fine-grained document chunks. A globally unique Chunk_ID is assigned to each document chunk, its vector representation (Embeddings) is calculated, and stored in a semantic index (vector database).

[0039] Using Large Language Models (LLM) or relation extraction models, causal triples (e.g., ...) are extracted from each specific document block. When storing these triples into a causal index (graph database), a "raw position injection" operation is performed: The Chunk_ID from which the triple originates is written as "Source Metadata" into the edge attributes of the graph database. Simultaneously, the extracted original evidence snippet is also stored as an attribute.

[0040] Thus, a "map structure" was established at the physical level. Chunk_ID The rigid mapping of the "original document block" lays the data foundation for subsequent "index penetration" and precise recall.

[0041] Furthermore, the process for retrieving candidate documents from the initial causal graph is as follows: determine the causal path based on the initial causal graph; convert the causal path into a graph database query statement (such as Cypher), obtain the entity attributes and associated triplet texts on the causal path from the graph database; generate a query vector based on the entity attributes and associated triplet texts, and retrieve candidate documents from the semantic recall search range based on the query vector using an approximate nearest neighbor search method.

[0042] Approximate Nearest Neighbor Search (ANN) is an optimization algorithm that finds the most similar (or closest) data point to a given query point in a large dataset. Examples include the Priority Search K-Means Tree Algorithm or the Hierarchical Clustering Tree.

[0043] In this embodiment, a causal index is used to quickly retrieve the graph structure of the initial causal graph and locate the set of parent nodes of the outcome variable. To construct a "root cause diagnostic pathway"; and simultaneously locate the set of child nodes of the treatment variable. The "intervention path" is constructed. The "root cause diagnosis path" and "intervention path" are used as causal paths. Key nodes and edges in the causal path are mapped to unstructured document block IDs in the semantic index to determine the semantic recall search scope. The causal path is transformed into a graph database query (e.g., Cypher) to obtain entity attributes and associated triplet texts along the path. The "source node + relation + target node" along the path is concatenated into a natural language description (e.g., "firewall misconfiguration caused connection reset"), encoded into a query vector, and an approximate nearest neighbor (ANN) search is performed in the semantic index to recall the Top-K candidate document sets. .

[0044] It should also be noted that the parent node of the outcome variable is a confounding variable or a proximal mediator variable, used to trace and locate the cause of the retrieval phenomenon. Its proximal mediator variable is, in this causal graph, the node whose direct parent node is connected to the outcome variable and whose distance is 1. The set of child nodes of the treatment variable are mediator variables, used to predict the outcome after the intervention.

[0045] Furthermore, the processing method for scoring the explanatory power and constructing evidence for the retrieved candidate documents based on causal edge weights is as follows: the explanatory power of the retrieved candidate documents is scored using a multi-dimensional scoring formula based on the causal edge weights to obtain the explanatory power score corresponding to each candidate document; target documents are selected according to preset screening rules based on the explanatory power scores corresponding to each candidate document; and the target documents are used as attribute nodes and mounted onto the causal path corresponding to the initial causal graph for evidence construction.

[0046] The multidimensional scoring formula is as follows:

[0047] In the formula, This is a vector-based semantic similarity measure to assess surface-level relevance. As a causal implication indicator function, use a Natural Language Inference (NLI) model to determine the document. Does it contain a causal boundary? (Note: This function can output a binary determination result.) Alternatively, the continuous implication probability value output by the NLI model can be used. To improve smoothness); This represents the causal weight of the edge in the global graph.

[0048] This step establishes a strict causal entailment relationship between the path and the document (solving the "path-document relationship" problem). The system will then link the path edges... Candidate documents are considered as "logical hypotheses" to be verified. It is considered a potential "premise of fact." The indicator function is only activated when the document content can logically deduce the causal relationship presupposed by the path (i.e., the evidence supports the hypothesis). Only then will it take effect, thus assigning a score to the document's interpretability.

[0049] The preset filtering rules are user-defined settings, such as filtering documents that exceed a preset threshold. This preset threshold is also user-defined.

[0050] In the specific implementation, the selected high-interpretation-score documents are mounted as attribute nodes to... On the corresponding path, an expanded evidence map (i.e., a causal graph) is formed. This provides structured contextual support for subsequent reasoning generation.

[0051] The retrieval agent needs to send the generated causal graph to the central orchestrator so that the central orchestrator can call the inference agent to generate the reverse causal graph.

[0052] The reasoning agent 1003 is used to generate a reverse causal graph based on the causal graph.

[0053] Functional definition of the reasoning agent: responsible for receiving structured evidence graphs provided by the retrieval agent. It performs a "Constrained Generation" task. This requires not only generating natural language text to answer the user's questions, but also explicitly tagging the sources of cited evidence and outputting a structured draft for subsequent verification. Its model selection employs a large language model that supports long context windows (such as GPT-4-Turbo or Claude-3) to ensure the complete reading and understanding of retrieved long text evidence.

[0054] Input / output: Input: User Query + Evidence Graph (Nodes&Documents). Output: (Draft text containing the causal chain derivation process). Furthermore, the process of generating a reverse causal graph based on the causal graph is as follows: a natural language response text is generated based on the causal graph through a controlled generation task; triples are extracted from the natural language response text through the RE model, and a reverse causal graph is generated based on the triples.

[0055] In practical implementation, the natural language response text generated by the reasoning agent... In this process, a Relation Extraction (RE) model is used to extract triples from the generated response text, thereby constructing a causal graph in reverse. .

[0056] The reasoning agent needs to send the reverse causal graph to the central orchestrator so that the central orchestrator can call the verification agent to verify the causal graph and the reverse causal graph.

[0057] The verification agent 1004 is used to perform logical detection on the reverse causal graph and the causal graph spectrum. If the detection result is a logical conflict, a rollback mechanism is triggered to generate a replanning instruction.

[0058] Furthermore, the process for generating replanning instructions is as follows: generate explicit negative constraint instructions based on the rollback mechanism; generate replanning instructions based on explicit negative constraint instructions and natural language query instructions.

[0059] In the specific implementation, a Consistency-Check Agent is used to perform rigorous logical auditing on the generated content to ensure the consistency between the reasoning path and the retrieved evidence.

[0060] The following two-step algorithm is used to determine logical conflicts, specifically addressing the problem of "how to determine if an edge direction is reversed": Node anchor point alignment: First, calculate the generated graph. With evidence diagram Semantic similarity of nodes, and establishment of node mapping function. For example, map "Cache Miss" in the generated graph to "Redis Cache Miss" in the evidence graph.

[0061] Reverse edge detection logic: traversal Each directed edge in : Search for evidence images The corresponding node pair .

[0062] Judgment criterion: If a reverse edge is explicitly present in the evidence diagram. And at the same time, there are no positive edges. Or there exists a clear "blocking / irrelevant" edge. If so, it is determined that "causal direction reversal" or "pseudo-causal relationship" has been detected.

[0063] Global structural similarity: If there are no specific reverse edges, calculate the Weisfeiler-Lehman (WL) graph kernel similarity. If it is below the threshold This is also considered a structural illusion. WL Iteration: Execution The Weisfeiler-Lehman color aggregation update is performed. Similarity calculation: the inner product of the two graphs in the feature space is calculated.

[0064] Negative Constraint Generation and Rollback: Once the above logical conflict is detected, the rollback mechanism will be triggered immediately: Constraint construction: For detected reverse edges (such as those incorrectly claimed) Generate explicit negative constraint instructions. (For example: "Note: Evidence suggests that A is actually a result of B, not a cause. Please rule out the hypothesis that A caused B.")

[0065] Replanning directive: Send back to the central orchestrator.

[0066] Feedback agent 1005 is used to acquire the user's explicit and implicit behaviors toward the search results in real time, and dynamically update the causal edge weights in the causal graph based on the explicit and implicit behaviors.

[0067] By leveraging feedback agents to capture user interactions, group behavior can be transformed into causal knowledge through online learning, enabling the system to self-evolve.

[0068] User actions on search results, whether explicit or implicit (e.g., clicking, copying, liking, dwell time, skipping), are considered interventions in the recommendation causal path. The current context state is recorded. and the actions taken (i.e., the recommended causal explanation path).

[0069] Because user feedback exhibits positional bias (users tend to click on results appearing earlier) and interaction behavior has multimodal characteristics, inverse bias scoring (IPS) is used to estimate unbiased rewards. :

[0070] in, The probability (propensity score) that the path was recommended at that time is used as the biased denominator. This represents the observed combined interaction utility (cumulative benefit). It is no longer a binary judgment of a single action, but rather a weighted sum of all explicit and implicit behaviors in the current session:

[0071] This is a function that indicates the occurrence of the behavior. Preset behavior weights (e.g.: [Click], [Long stay], [copy], [Dislike]

[0072] Note: If a user both clicks and copies, the rewards will be automatically added together, thus more accurately quantifying user satisfaction.

[0073] Connection point definition: calculated (Usually need to be normalized to a reasonable range) will serve as "observational evidence" in subsequent Bayesian updates, used to correct posterior beliefs about causal weights.

[0074] The Thompson Sampling strategy is employed, utilizing the obtained rewards. Update the edges in the causal graph The weight distribution parameters (Beta distribution) are used to establish a mathematical connection between rewards and weights: Parameter update: for the intervened edge According to unbiased reward Update the hyperparameters of its Beta distribution. (representing the accumulation of positive evidence) and (Representing the accumulation of negative / invalid evidence):

[0075]

[0076] Weight sampling: In the next retrieval, the edge weights are no longer fixed values, but are sampled from the updated distribution. .

[0077] Through this mechanism, high-reward paths... Increase the sampled weights The graph tends to become larger, making it easier to be selected in subsequent searches, thus achieving a "survival of the fittest" evolution of the graph.

[0078] In the specific implementation, the feedback agent sends the updated weights to the central orchestrator, so that the central orchestrator calls the retrieval agent and forces it to search for counterfactual evidence again under the constraints of the new logic (updated weights and replanning instructions) to correct the retrieval space.

[0079] The following example demonstrates the data flow at each step of a complete IT operations and maintenance scenario.

[0080] Scenario input: The user (operations engineer) enters in the dialog box: "The business system API response timed out, the database CPU load is normal, but a large number of TCP connection reset errors were found in the application logs." Phase 1: Initialization The Orchestrator receives the Query and forwards it to the Intent-Causal Agent.

[0081] Intent-Causal Agent Analysis: Outcome: API timeout (V1), TCP connection reset (V2).

[0082] Exclusion Item (Evidence): Database CPU is normal (V3). This means that the probability of "database performance issues" causing V1 is extremely low.

[0083] Potential Hypothesis: Based on SD-SCM knowledge, inferences are made. The network firewall is misconfigured. Load balancer (LB) session persistence failure, Application layer connection pool is full.

[0084] Constructing a DAG:

[0085]

[0086] (Negative edge) Output To the central orchestrator.

[0087] Phase Two: Collaborative Retrieval The Orchestrator invokes the Retrieval Agent.

[0088] The Retrieval Agent performs a dual-path retrieval: Path1 (for) ): Search for related cases of "firewall configuration causing TCP Reset".

[0089] Path2 (for) ): Retrieve documents related to "API timeout caused by connection pool full".

[0090] Path3 (Negative Validation): Retrieves rare causes for "database CPU is normal but API timeout".

[0091] Explanatory power score: Document D1 (“Firewall blocking caused RST packet”) and The path is highly matched, with an NLI score of 0.9 and a high global causal weight (a common problem). The total score is... .

[0092] Document D2 (“Database deadlock caused timeout”) contains the keyword, but is not related to... (CPU normal) Contradictory, NLI score is low or negative, total score .

[0093] Output Evidence Set .

[0094] Phase 3: Generation & Verification Reasoning Agent generates draft "According to document D1, the firewall may have blocked the persistent connection, causing the TCP to be reset. It is recommended to check the firewall rules." Consistency-Check Agent Intervention: extract Cause-and-effect diagram: .

[0095] Comparison Consistent.

[0096] Compare with the knowledge base (KB): There is a rule in the KB that "firewall blocking usually logs Drop".

[0097] Potential omission discovered: The user did not mention Drop logs. The validating agent considers the evidence insufficient and sets the confidence level to Medium.

[0098] Attempting a rollback (simulation): Suppose the generated answer is "timeout due to slow database," which is... Conflict. At this point, a rollback is triggered, generating a constraint "excluding database reasons".

[0099] This example passed the validation: the final answer was output, along with the following operation suggestion: "Please check the firewall drop log (Intervention I1)." Phase Four: Feedback and Evolution The user followed the suggestion and clicked "Feedback: No firewall logs found, problem unresolved" in the system.

[0100] The Feedback Agent captures "negative feedback".

[0101] Weight update: reduce Firewall The weight of TCP reset in the current context (Context = API timeout + DB normal): .

[0102] Orchestrator automatically triggered a secondary search, this time because With reduced weights, the Bandit algorithm tends to explore... (Connection pool full) or (LB problem).

[0103] The system generated a new suggestion: "Please check the application server's ulimit settings or the connection pool's max_active parameter." The user clicked "Accept" and provided feedback that the problem had been resolved.

[0104] Evolution: Increase and weight .

[0105] Long-term impact: The next time a similar description is encountered (API timeout + DB normal + TCP Reset), the system will prioritize checking the connection pool rather than the firewall.

[0106] This system can be deployed on a cloud server cluster.

[0107] Compute nodes: Equipped with high-performance GPUs (such as NVIDIA A100) to run LLM (Intent Agent, Reasoning Agent) and Embedding models.

[0108] Storage nodes: Graph databases (such as Neo4j / NebulaGraph) store global causal graphs of evolution.

[0109] Vector databases (such as Milvus / Faiss) store document indexes.

[0110] Redis cache: Stores short-term session states and rollback contexts for users.

[0111] In this embodiment, a multi-agent collaborative retrieval system based on large-model causal verification is designed. The system includes an intent parsing agent, a retrieval agent, an inference agent, a verification agent, and a feedback agent. The intent parsing agent parses the natural language query command input by the user into an initial causal graph. The retrieval agent retrieves candidate documents from the initial causal graph and performs explanatory power scoring and evidence construction on the retrieved candidate documents based on causal edge weights to form a causal graph spectrum. The inference agent generates a reverse causal graph based on the causal graph spectrum. The verification agent performs logical checks on the reverse causal graph and the causal graph spectrum. If the check result is a logical conflict, a rollback mechanism is triggered, and a replanning instruction is generated. The feedback agent obtains the user's explicit and implicit behaviors towards the retrieval results in real time and dynamically updates the causal edge weights in the causal graph spectrum based on the explicit and implicit behaviors. The retrieval agent collaboratively performs retrieval correction on the causal graph spectrum based on the replanning instruction and the updated causal edge weights. This embodiment, through multi-agent collaboration, not only enables the instant construction of causal graphs for specific queries, but also introduces a "verification-rollback" closed-loop control flow and a weight update mechanism based on a "causal multi-armed slot machine," thereby fundamentally solving the problems of logical inconsistency and knowledge update lag in existing technologies, and thus improving the accuracy of search results.

[0112] This invention provides a multi-agent collaborative retrieval method based on large-model causal verification, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the multi-agent collaborative retrieval method based on large-model causal verification of the present invention.

[0113] In this embodiment, the multi-agent collaborative retrieval method based on large-model causal verification includes the following steps: S1, the intention parsing intelligent agent parses the natural language query command input by the user into an initial causal graph, wherein the natural language query command is an IT operation and maintenance knowledge query command; S2, the retrieval agent performs candidate document retrieval on the initial causal graph, and performs explanatory power scoring and evidence construction on the retrieved candidate documents based on the causal edge weights to form a causal graph. S3, the reasoning agent generates a reverse causal graph based on the causal graph; S4, the intelligent agent performs logical checks on the reverse causal graph and the causal graph spectrum. If the check result is a logical conflict, the rollback mechanism is triggered and a replanning instruction is generated. S5, the user's explicit and implicit behaviors toward the search results are obtained in real time through the feedback agent, and the causal edge weights in the causal graph are dynamically updated according to the explicit and implicit behaviors. S6, the retrieval agent performs retrieval and correction of the causal graph based on the replanning instruction and the updated causal edge weights.

[0114] Other embodiments or specific implementations of the multi-agent collaborative retrieval method based on large model causal verification of the present invention can be referred to the above system embodiments, and will not be repeated here.

[0115] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0116] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0118] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A multi-agent collaborative retrieval system based on large-model causal verification, characterized in that, The system includes an intent parsing agent, a retrieval agent, an inference agent, a verification agent, and a feedback agent. The intent parsing intelligent agent is used to parse the natural language query command input by the user into an initial cause-effect graph, wherein the natural language query command is an IT operation and maintenance knowledge query command; The retrieval agent is used to retrieve candidate documents from the initial causal graph, and to perform explanatory power scoring and evidence construction on the retrieved candidate documents based on causal edge weights to form a causal graph. The reasoning agent is used to generate a reverse causal graph based on the causal graph. The verification agent is used to perform logical detection on the reverse causal graph and the causal graph spectrum. If the detection result is a logical conflict, a rollback mechanism is triggered and a replanning instruction is generated. The feedback agent is used to acquire the user's explicit and implicit behaviors toward the search results in real time, and dynamically update the causal edge weights in the causal graph based on the explicit and implicit behaviors. The retrieval agent is also used to perform retrieval correction on the causal graph based on the replanning instruction and the updated causal edge weights. The step of parsing the user-input natural language query command into an initial causal graph includes: Identify an entity set from the natural speech query command input by the user, and classify each entity in the entity set into variables such as treatment variables, outcome variables, mediation variables, and confounding variables; Using a large language model as a proxy for a sequence-driven structural equation model, the dependency direction between variables is determined based on conditional probability in order to construct an initial causal graph; The initial cause-effect graph is as follows: ； In the formula, V is a set of entities containing directed edges with initial confidence levels; The process of scoring the explanatory power of retrieved candidate documents and constructing evidence based on causal edge weights to form a causal graph includes: Based on causal edge weights, a multi-dimensional scoring formula is used to score the explanatory power of the retrieved candidate documents to obtain the explanatory power score corresponding to each candidate document. Target documents are selected based on the explanatory power scores of each candidate document according to preset filtering rules; The target document is used as an attribute node and attached to the causal path corresponding to the initial causal graph to form an expanded causal graph.

2. The system as described in claim 1, characterized in that, The system also includes a central orchestrator; The central orchestrator is used to invoke the intent parsing agent, retrieval agent, reasoning agent, verification agent, and feedback agent.

3. The system as described in claim 1, characterized in that, Candidate document retrieval is performed on the initial causal graph, including: Determine causal paths based on the initial causal graph; The causal path is converted into a graph database query statement to obtain the entity attributes and associated triplet text on the causal path; A query vector is generated based on the entity attributes and the associated triplet text, and candidate documents are retrieved from the semantic recall search range using the approximate nearest neighbor search method based on the query vector.

4. The system as described in claim 1, characterized in that, Generating a reverse causal graph based on the causal graph includes: Based on the causal graph, natural language response text is generated through a controlled generation task; Triples are extracted from the natural language response text using the RE model, and a reverse causal graph is generated based on the triples.

5. The system as described in claim 1, characterized in that, Generate replanning instructions, including: Based on the aforementioned rollback mechanism, a negative constraint instruction is generated to display the constraint. A replanning instruction is generated based on the negative constraint instruction and the natural language query instruction.

6. A multi-agent collaborative retrieval method based on large-model causal verification, applied to the multi-agent collaborative retrieval system based on large-model causal verification as described in claim 1, characterized in that, The method includes the following steps: The intent parsing agent parses the natural language query command input by the user into an initial cause-effect graph, wherein the natural language query command is an IT operation and maintenance knowledge query command; The initial causal graph is retrieved by a retrieval agent, and the retrieved candidate documents are scored for explanatory power and evidence is constructed based on the causal edge weights to form a causal graph. The reasoning agent generates a reverse causal graph based on the causal graph; The intelligent agent performs logical checks on the reverse causal graph and the causal graph spectrum. If the check result is a logical conflict, a rollback mechanism is triggered and a replanning instruction is generated. The system obtains the user's explicit and implicit behaviors toward the search results in real time through a feedback agent, and dynamically updates the causal edge weights in the causal graph based on these behaviors. The retrieval agent performs retrieval and correction of the causal graph based on the replanning instruction and the updated causal edge weights.

Images