Verifiable search extension generation devices, equipment, recording media, and program products based on cooperative multi-agent systems.

Abstract

This invention provides an apparatus, method, program, and recording medium that improve the verifiability of a RAG (Retrieval-Augmented Generation) system by introducing a dynamic verification layer and an audit agent and constructing a non-linear flow. [Solution] The search extension generation method includes the steps of: passing a user query to a query analysis agent, having the query analysis agent perform intent analysis and optimization of the user query to generate an optimized query valid; distributing the optimized query valid to multiple search agents to perform parallel searches and outputting a raw information set; inputting the raw information set to an information verification and criticism agent, having the information verification and criticism agent perform dynamic verification processing of the raw information set; and having an audit agent capture and record the data flow of each agent in real time to generate audit tracking data.

Description

Technical Field

[0001] The present application relates to the field of information retrieval, and in particular to a verifiable retrieval-augmented generation apparatus, device, recording medium, and program product based on a cooperative multi-agent system.

Background Art

[0002] Conventional retrieval-augmented generation technology adopts a linear single-path architecture, and its three stages of indexing, retrieval and generation are connected in series in a chain. This architecture has inherent vulnerabilities and is prone to cascading failures. Errors in the retrieval stage, such as the inclusion of irrelevant documents in retrieval results or the absence of important information, may directly contaminate the subsequent generation stage, resulting in factual hallucinations in the generation model. Typical failure modes include content missing, extraction failure caused by context noise interference, and failure of effective information to be incorporated into the context due to knowledge collision. More importantly, such systems do not have a dynamic verification mechanism, and cannot detect logical contradictions in real time or filter noise interference during the transmission process of information flow, so errors are continuously amplified. Furthermore, the decision-making process of the system exhibits black-box characteristics, can only provide superficial citation indication of sources, and cannot trace the decision-making basis for why important information is discarded. This does not meet the core requirement of auditability required for trustworthy artificial intelligence, and there are also potential security risks.

[0003] While multi-agent systems have already been applied in the fields of task decomposition and collaborative reasoning, conventional technologies lack specialized roles designed specifically for retrieval and augmentation scenarios. General-purpose multi-agent frameworks can improve the parallel processing capabilities of tasks, but they lack dedicated components to resolve cascading failures in retrieval and augmentation. Such systems not only lack dedicated gatekeeper agents to validate retrieval information for inconsistency detection, source reliability assessment, and conflict resolution, but also lack an integrated audit and tracking layer across the entire process to ensure end-to-end decision-making traceability. This lack of functionality prevents even multi-agent architectures from systematically mitigating the error chains specific to retrieval and augmentation, such as output deviations due to knowledge collisions and loss of useful information due to truncation errors, making it difficult to meet the stringent requirements for verifiability in highly sensitive scenarios such as medical diagnosis and financial risk management.

[0004] Therefore, in order to overcome the reliability, transparency, and security deficiencies that still exist in conventional search augmentation generation systems, there is an urgent need to build a collaborative multi-agent framework specifically for search augmentation generation scenarios. [Overview of the project] [Problems that the invention aims to solve]

[0005] This application provides a collaborative multi-agent-based verifiable search augmentation generation method that improves the verifiability of a RAG (Retrieval-Augmented Generation) system by introducing a dynamic verification layer and audit agents and constructing a nonlinear flow. [Means for solving the problem]

[0006] To achieve the above objective, the present invention employs the following technical solutions.

[0007] In a first aspect, the present invention relates to a verifiable search extension generation method based on cooperative multi-agent, This involves receiving user queries via an orchestration agent to retrieve them and perform task decomposition and resource scheduling, The user query is passed to a query analysis agent, which performs intent analysis and optimization of the user query and generates an optimized query validator. The optimized query validator is distributed to multiple search agents to perform parallel searches and output a set of raw information. The process involves inputting the aforementioned raw information set into an information verification and criticism agent, performing dynamic verification processing on the raw information set by the information verification and criticism agent, and generating a verified context, wherein the dynamic verification processing includes inconsistency detection, confidence scoring, and noise filtering of the raw information set. The verified context is passed to the ground generation agent, and the ground generation agent generates a final response based on the verified context. Based on the data flow of the orchestration agent, query analysis agent, search agent, information verification / criticism agent, and ground generation agent, which are captured and recorded in real time by the audit agent, audit tracking data is generated. This provides a method that includes [something].

[0008] A preferred embodiment of this application further includes performing dynamic verification processing of the original information set by the information verification and criticism agent, If the information verification and criticism agent detects that the confidence score of the original information set falls below a predetermined threshold, it sends a re-search command to the orchestration agent. The orchestration agent generates a targeted query valid based on the received re-search command, and transmits the targeted query valid to the multiple heterogeneous search agents. It may be configured to include

[0009] A preferred embodiment of this application further includes performing dynamic verification processing of the original information set by the information verification and criticism agent, In the aforementioned set of raw information, identify and mark logical inconsistencies between information fragments retrieved from different sources. Assigning a source confidence score to each of the aforementioned information fragments, Filtering information fragments from the aforementioned raw information set whose confidence score falls below a predetermined threshold, Integrating filtered information fragments to form the verified context, It may be configured to include

[0010] A preferred embodiment of this application further distributes the optimized query valid to multiple search agents to perform parallel searches and outputs a set of raw information, The dense search agent and the sparse search agent are activated, the dense search agent performs semantic vector matching on the optimized query valid, and the sparse search agent performs keyword matching on the optimized query valid based on a keyword model to obtain first search data. The process involves launching a knowledge graph search agent, scanning nodes and relationships within a structured knowledge graph, and obtaining second-hand search data. This involves calling a real-time network search agent to obtain current events data and then obtaining a third set of search data. The first search data, the second search data, and the third search data are aggregated to obtain the original information set, It may be configured to include

[0011] In a preferred embodiment of the present application, the ground generation agent may further be configured to be built on a large-scale language model.

[0012] In a preferred embodiment of the present application, the data flow may further include at least the user's initial query, the reconstructed result of the query analysis agent, the data source of the search agent, and the retrieved specific text chunk, wherein the audit tracking data has a timestamp.

[0013] One embodiment of this application further relates the computer device to the diagnosis and treatment of a disease, wherein the user query includes disease information, the optimized query validator includes a reconstructed or extended medical ontology language corresponding to the disease information, the raw information set includes graph pathway relationships between drugs and contraindications, or between symptoms and diseases, corresponding to the optimized query validator, and the final response includes disease diagnosis results and treatment suggestions corresponding to the validated context. Inputting the aforementioned raw information set into an information verification and criticism agent, and performing dynamic verification processing of the raw information set by the information verification and criticism agent, is: Inputting the aforementioned set of raw information into the aforementioned information verification and critical agent, The aforementioned information verification and criticism agent identifies and marks logical contradictions between information fragments retrieved from different sources in the original information set, wherein the logical contradictions include conflicts of dosage. The aforementioned information verification and criticism agent assigns a source confidence score to the information fragments corresponding to the logical contradictions in accordance with the rules of evidence-based medicine, and sends a re-search command to the orchestration agent. The orchestration agent may be configured to: generate targeted query variants based on the received re-search instruction, and transmit the targeted query variants to the plurality of search agents to cause them to perform re-search.

[0014] In a second aspect, the present application provides a verifiable retrieval-augmented generation apparatus based on cooperative multi-agent, comprising: a data collection module configured to obtain a user query and receive the user query via an orchestration agent for performing task decomposition and resource scheduling; a query analysis module configured to pass the user query to a query analysis agent, perform intent analysis and optimization on the user query by means of the query analysis agent, and generate optimized query variants; a search module configured to distribute the optimized query variants to a plurality of search agents to cause them to perform parallel search and output a raw information set; a verification module configured to input the raw information set into an information verification and critique agent, perform dynamic verification processing including contradiction detection, confidence scoring and noise filtering on the raw information set by means of the information verification and critique agent, and generate a verified context; an output module configured to pass the verified context to a grounded generation agent, and generate a final response based on the verified context by means of the grounded generation agent; an auditing module configured to generate audit trail data based on data flows of the orchestration agent, the query analysis agent, the search agent, the information verification and critique agent, and the grounded generation agent that are captured and recorded in real time by an auditing agent; and the apparatus comprises the above modules.

[0015] In a third aspect, the present application provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the verifiable retrieval-augmented generation method based on a cooperative multi-agent according to any of the above embodiments are implemented.

[0016] In a fourth aspect, the present application provides a computer-readable storage medium having a program stored thereon, wherein when the program is executed by a processor, the verifiable retrieval-augmented generation method based on a cooperative multi-agent according to any of the above embodiments is implemented.

[0017] In a fifth aspect, the present application provides a computer program product comprising computer instructions, wherein when the computer instructions are executed by a processor, the verifiable retrieval-augmented generation method based on a cooperative multi-agent according to any of the above embodiments is implemented. Effects of the Invention

[0018] As described above, compared with the prior art, the beneficial effects brought by the technical solution according to the embodiments of the present application include at least the following.

[0019] The conventional RAG linear architecture has a risk of cascading failure, where errors in the retrieval stage are directly propagated to the generation stage. The present application improves the error blocking rate by introducing an information verification and criticism agent as a dynamic filter.

[0020] The present application introduces an audit agent, generates tamper-proof audit trail logs, and realizes traceability of the end-to-end inference process.

[0021] According to the flow architecture of this application, the generating agent must rely entirely on a pre-verified, high-quality context, thereby ensuring that the output content is based on verifiable factual evidence, improving accuracy, and reducing collision response time. [Brief explanation of the drawing]

[0022] [Figure 1] This is a flowchart of a verifiable search extension generation method based on a cooperative multi-agent system according to one embodiment of this application. [Figure 2] This is a data flow diagram of a verifiable search extension generation method based on cooperative multi-agent systems according to one embodiment of this application. [Figure 3] This is an interaction diagram of each agent module in a verifiable search extension generation method based on a cooperative multi-agent system according to one embodiment of this application. [Figure 4] This is a verification search iteration flowchart for a verifiable search extension generation method based on a cooperative multi-agent system according to one embodiment of this application. [Figure 5] This is a modular diagram of a verifiable search extension generation device based on a cooperative multi-agent system according to one embodiment of this application. [Modes for carrying out the invention]

[0023] The technical solutions of the embodiments of this application will be described clearly and completely below with reference to the drawings of the embodiments of this application, but the embodiments described are only some, not all, embodiments of this application. All other embodiments that can be obtained by a person skilled in the art without creative work based on the embodiments of this application are within the scope of protection of this application.

[0024] The verifiable search extension generation method based on cooperative multi-agents according to embodiments of this application is applicable to terminals, applicable to servers, and may also be software that runs on either the terminal or the server. In some embodiments, the terminal may be a smartphone, tablet PC, notebook PC, desktop computer, etc., and the server side may be configured as an independent physical server, as a server cluster or distributed system composed of multiple physical servers, or as a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain services, security services, CDNs, and big data and AI platforms. The software may be, but is not limited to, an application that implements the verifiable search extension generation method based on cooperative multi-agents.

[0025] In one embodiment of this application, a verifiable search extension generation method based on cooperative multi-agent is provided, referring to Figure 1, the method includes the following steps S100 to S600. S100: Retrieves user queries and receives them via an orchestration agent to perform task decomposition and resource scheduling. S200: The user query is passed to the query analysis agent, which performs intent analysis and optimization of the user query and generates an optimized query validator. S300: The optimized query validator is distributed to multiple search agents to perform parallel searches and output a set of raw information. S400: The raw information set is input to the information verification and criticism agent, which performs dynamic verification processing on the raw information set, including inconsistency detection, confidence scoring, and noise filtering, to generate a verified context. S500: The verified context is passed to the ground generation agent, and the ground generation agent generates a final response based on the verified context. S600: Audit tracking data is generated based on the data flow of the orchestration agent, query analysis agent, search agent, information verification / criticism agent, and ground generation agent, which are captured and recorded in real time by the audit agent.

[0026] In the specific implementation, refer to Figure 2 and execute steps S100 to S500 sequentially to obtain the final response. Step S600 synchronously executes data collection as each step from S100 to S500 is performed to obtain audit tracking data.

[0027] Figure 3 shows the main agents in this embodiment and their main interaction relationships.

[0028] The orchestration agent acts as a central coordinator, similar to a "master agent" in a hierarchical structure, and is responsible for receiving the user's initial query and managing and scheduling the overall task execution flow. The orchestration agent functions as the overall control unit of the system, managing the task lifecycle. Its responsibilities include receiving user queries, breaking them down into smaller subtasks according to their complexity, accurately assigning the broken-down subtasks to corresponding specialist agents (e.g., query analysis agents, search agents), receiving the final verified context from the information validation and criticism agents, receiving the generated response drafts from the generation agents, and integrating and providing the user with a final output package that includes the answer itself, as well as a link to or summary of the complete audit trace.

[0029] The aforementioned query analysis agent is a specialized agent dedicated to understanding and optimizing user input before the search begins. Its purpose is to improve the accuracy of information retrieval from its source. Its function is deep analysis and optimization of user input. Its responsibilities include analyzing user queries to identify key entities, core intent, and contextual information (derived from the design philosophy of "intent-driven RAG"), resolving ambiguity in queries and clarifying and concretizing vague or overly broad queries (this step directly mitigates the failure mode known as FP6 (Incorrect Specificity) by improving input quality), and performing query expansion or rewriting to generate multiple query validators from different angles and pass these validators to the search agent, thereby improving search recall and coverage.

[0030] The aforementioned search agent includes a group of parallel search agents, employing a group of agents that execute search tasks in parallel, rather than a single search agent. Each agent can employ a different search strategy, thereby maximizing the breadth and quality of the information retrieved. This goes beyond simple hybrid search. Its function is to retrieve information in parallel from multiple heterogeneous knowledge sources. Specifically, this includes the following: Dense Search Agent: Utilizing advanced embedding models, it performs semantic searches within vector databases to capture deep semantic similarities. Sparse search agents: These agents employ traditional keyword-based methods (such as BM25) or sparse vector models (such as SPLADE) to ensure accurate keyword matching. Knowledge Graph Agent: Scans nodes and relationships in a structured knowledge graph to accurately retrieve factual entities and their relationships. This is particularly effective for queries that require precise facts. Real-time network agent: Uses API calls or web scraping techniques to retrieve real-time information (news, social media, etc.) from the internet, ensuring the up-to-dateness of knowledge.

[0031] In some embodiments, the specific process of step S300 is The dense search agent and the sparse search agent are activated, the dense search agent performs semantic vector matching on the optimized query valid, and the sparse search agent performs keyword matching on the optimized query valid based on a keyword model to obtain first search data. The process involves launching a knowledge graph search agent, scanning nodes and relationships within a structured knowledge graph, and obtaining second-hand search data. This involves calling a real-time network search agent to obtain current events data and then obtaining a third set of search data. The first search data, the second search data, and the third search data are aggregated to obtain the original information set, Includes.

[0032] The aforementioned information verification and critical agent is positioned between retrieval and generation as a critical reasoning and information filtering layer. Its goal is to "purify" and "verify" all retrieved raw information before generation begins. Its functions include critical evaluation, filtering, and integration of the retrieved raw information. Its responsibilities include the following: Inconsistency detection: Implement specific algorithms to identify and mark logical inconsistencies or factual conflicts that exist between blocks of information retrieved from different sources. Source Reliability Scoring: Evaluates the authority and reliability of each information source and assigns higher weight to data from more reliable sources. Noise and redundancy filtering: Removes irrelevant or redundant information to create a concise and efficient contextual environment. This directly solves the FP4 (extraction failure) problem caused by contextual noise. Bias assessment: Using pre-trained models or rules, the searched text is scanned for known social, cultural, or political biases, ensuring compliance with responsible AI design principles. Verifiability Scoring: Based on all the factors above, an overall confidence level or "verifiability" score is generated for each piece of information. Conflict Resolution: Based on the scoring results, agents can decide to discard less reliable information, mark detected critical conflicts and report them to the orchestration agent, or synthesize each party's perspective to generate a summary of the conflict itself. This feature directly solves the "knowledge conflict" problem in traditional RAG systems.

[0033] The aforementioned grounding generation agent is an agent based on a Large Language Model (LLM) and generates natural language responses that the final user can read.

[0034] Its functionality includes generating a final answer based on verified context. Unlike standard RAGs, this agent only receives purified and verified context from information validation and criticism agents. It is explicitly instructed that its output must rely entirely on the provided context, which significantly reduces the possibility of the model generating hallucinations or relying on outdated parameterized knowledge. Furthermore, it is possible to instruct it to adhere to a specific output format (such as JSON), which can mitigate the FP5: Wrong Format issue.

[0035] The aforementioned audit agent is a continuously running agent that, as the "recorder" of the entire system, creates immutable and verifiable records of all processing processes. The design of this agent directly embodies the core principles of verifiable AI. Its functions include recording all critical operations and decisions within the system, forming a traceable audit trail. Specifically, it records all critical actions, including the user's initial query, the reconstruction results of the query analysis agent, the data sources referenced by each search agent, and the specific text chunks retrieved. It also meticulously records the decision-making process of the information verification and criticism agent, including what information was filtered, the reasons for filtering (e.g., "inconsistency detected," "low source reliability"), and the contextual content ultimately passed to the generating agent. Furthermore, it generates a timestamped final audit trail record, ensuring its integrity and tamper-proofness by using cryptographic signatures or storing it in an immutable ledger (such as a blockchain) where necessary. This record provides complete traceability and interpretability for the system's decisions.

[0036] Table 1 provides a structured overview of the role of each agent. [Table 1]

[0037] The agent operation flow in this embodiment is a dynamic collaborative process rather than a linear pipeline, and an example of the flow is shown below. 1. Query Reception: The orchestration agent receives the query Q submitted by the user. 2. Query Preprocessing: The orchestration agent assigns query Q to the query analysis agent. This agent performs analysis and optimization and returns a set of optimized queries {Q'}. 3. Parallel Search: The orchestration agent assigns the query in {Q'} to a group of parallel search agents. These agents operate concurrently, each searching for information from its assigned knowledge source and returning a raw set of information blocks {C_raw} containing potentially duplicated and conflicting information. 4. Critical Verification: The orchestration agent passes the set {C_raw} to the information validation / critical agent. This agent performs its core analytical functions (inconsistency detection, scoring, filtering, etc.) and outputs a clean, validated context {C_validated} that has metadata (confidence score, etc.). 5. Grounding Generation: The orchestration agent passes both {C_validated} and the primitive query Q to the grounding generation agent. This agent generates a draft response R based on this high-quality context. 6. Synchronized Audit: Throughout the entire process from Step 1 to 5, the audit agent continuously monitors all interactions within the system and records all critical events, data flows, and decisions in real time in a secure audit tracker. 7. Synthesis of the Final Response: The orchestration agent receives the response R and references to the audit track T. It packages these elements and provides them to the user as the final output. The output may include the answer, cited sources (extracted from {C_validated}), and links or summaries to access the complete audit track.

[0038] In this embodiment, the linear architecture of conventional RAGs has the risk of cascading failures, and errors in the search phase may propagate directly to the generation phase. This application improves the error prevention rate by introducing an information verification and criticism agent as a dynamic filter.

[0039] This application enables end-to-end traceability of the inference process by introducing an audit agent and generating tamper-proof audit tracking logs.

[0040] According to the flow architecture of this application, the generating agent must rely entirely on pre-verified, high-quality context, thereby ensuring that the output content is based on verifiable factual evidence, improving accuracy, and reducing collision response time.

[0041] In some embodiments, as shown in Figure 4, the process of performing dynamic verification processing of the original information set by the information verification and criticism agent is as follows: If the information verification and criticism agent detects that the confidence score of the original information set falls below a predetermined threshold, it sends a re-search command to the orchestration agent. The orchestration agent generates a targeted query valid based on the received re-search command, and transmits the targeted query valid to the multiple heterogeneous search agents. Includes.

[0042] In its implementation, the steps of the method described above further include a self-correction loop. If the information verification and criticism agent marks a high degree of conflict or significantly low overall confidence in its analysis, the orchestration agent can generate new, more targeted queries based on the feedback from the initial verification and initiate a second round of the search and verification process.

[0043] In some embodiments, the process of performing dynamic verification processing of the original information set by the information verification and criticism agent is as follows: In the aforementioned set of raw information, identify and mark logical inconsistencies between information fragments retrieved from different sources. Assigning a source confidence score to each of the aforementioned information fragments, Filtering information fragments in the aforementioned raw information set whose confidence score falls below a predetermined threshold, Integrating filtered information fragments to form the verified context, Includes.

[0044] In specific implementation, the aforementioned information verification and criticism agent performs data processing on the raw information set to realize functions such as source confidence scoring, noise and redundancy filtering, verifiability scoring, and conflict resolution.

[0045] In this embodiment, the iterative processing makes it possible to actively solve problems such as insufficient information and a decline in information quality.

[0046] This application further provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor, by executing the computer program, realizes a step of a collaborative multi-agent based verifiable search extension generation method as described in any one of the embodiments described above.

[0047] In practical applications, the verifiable search extension generation method based on the cooperative multi-agent according to the embodiments of this application is applied to the diagnosis and treatment of diseases and specifically includes the following steps S10 to S60. S10: The orchestration agent receives the user query, which includes disease information, and performs task decomposition and resource scheduling. S20: The user query is passed to a query analysis agent, which performs intent analysis and optimization of the user query and generates an optimized query valid. Here, the optimized query valid includes a reconstructed or extended medical ontology language corresponding to the disease information. S30: The optimized query validators are distributed to multiple search agents to perform parallel searches and output a set of raw information. Here, the set of raw information includes graph pathway relationships between drugs and contraindications, or between symptoms and diseases, corresponding to the optimized query validators. S40: The raw information set is input to the information verification and criticism agent, which performs dynamic verification processing on the raw information set to generate a verified context. Here, the dynamic verification processing includes inconsistency detection, confidence scoring, and noise filtering of the raw information set. S50: The verified context is passed to the ground generation agent, which generates a final response based on the verified context. Here, the final response includes a disease diagnosis result and treatment suggestions corresponding to the disease information. S60: Audit tracking data is generated based on the data flows of the orchestration agent, query analysis agent, search agent, information verification / criticism agent, and ground generation agent, which are captured and recorded in real time by the audit agent.

[0048] The verifiable search extension generation method based on the cooperative multi-agent according to the embodiments of this application is applicable to the diagnosis and treatment of diseases and can be applied, for example, to humans or non-human animals.

[0049] In some embodiments, step S40, "inputting the raw information set into the information verification / critique agent and performing dynamic verification processing of the raw information set by the information verification / critique agent," is performed by: Inputting the aforementioned set of raw information into the aforementioned information verification and critical agent, The aforementioned information verification and criticism agent identifies and marks logical contradictions, including conflicts in medication, between information fragments retrieved from different sources in the original information set. The aforementioned information verification and criticism agent assigns a source confidence score to the information fragments corresponding to the logical contradictions in accordance with the rules of evidence-based medicine, and sends a re-search command to the orchestration agent. The orchestration agent generates targeted query validators based on the received re-search command, and sends the targeted query validators to the multiple search agents to perform the re-search.

[0050] In some embodiments, the targeted query validator can be made to determine which material to query by the intelligence of the model (i.e., the orchestration agent) itself. For example, the orchestration agent can also generate a targeted query validator based on the information fragment with the highest source confidence score in the logical contradiction relationship and the information fragment related to the information fragment with the highest source confidence score in the original information set.

[0051] In practical implementation, the aforementioned method is performed using computer equipment, thus forming a clinical decision support system (CDSS) applicable to complex medical scenarios. It should be noted that in clinical practice in rheumatology and collagen disease, physicians often face extremely difficult differential diagnoses. For example, when a patient with systemic lupus erythematosus (SLE) presents with central nervous system symptoms (epilepsy, impaired consciousness, etc.), it is often difficult to determine whether this is due to a relapse of the underlying disease (neuropsychiatric lupus) or a secondary central nervous system infection. However, the treatments for these two situations are diametrically opposed. The former requires high-dose steroid pulse therapy, while the latter is absolutely contraindicated for steroid use.

[0052] To address this challenge, this embodiment constructs a diagnostic support system (i.e., the clinical decision support system) based on a "collaborative multi-agent verifiable RAG". This system does not rely on "black-box" reasoning using a single model, but provides physicians with interpretable clinical advice through rigorous process orchestration, evidence retrieval from multiple sources, and dynamic conflict verification.

[0053] This system is designed using a cognitive architecture and consists primarily of six functionally heterogeneous but closely cohesive agents (namely, an orchestration agent, a query analysis agent, a search agent, an information verification / critique agent, a ground generation agent, and an audit agent). Unlike conventional linear systems, in this system, the orchestration agent functions as a central control unit, and a clinical path reasoning engine is built into it, allowing the "diagnostic tree" state machine to be dynamically maintained according to the task state.

[0054] At the input end of this system, the Query Analysis Agent is responsible for converting natural language into a machine-understandable medical ontology language. This agent integrates a BioBERT model to perform entity recognition and utilizes the Unified Medical Language System (UMLS) to map clinical terms to international standard codes such as SNOMED CT and RxNorm.

[0055] The search layer of this system consists of a heterogeneous group of search agents (i.e., multiple heterogeneous search agents), including a PubMedBERT-based dense search agent for processing unstructured medical histories, a BM25-based sparse search agent for processing accurate guidelines, and a Neo4j-based knowledge graph search agent. The aforementioned knowledge graph search agent is specifically designed to execute Cypher queries, thereby capturing graph pathway relationships between "drugs-contraindications" or "symptoms-diseases."

[0056] The security kernel of this system resides in the Verification & Critique Agent. This agent is equipped with a natural language reasoning (NLI) model finely tuned specifically for medical logic, which identifies inconsistencies in retrieved evidence (e.g., conflicts in medication) and dynamically scores sources according to the pyramid principle of evidence-based medicine. The final output is generated by the Ground Generation Agent, but data interactions throughout the entire process are recorded as HIPAA-compliant JSON logs by the Audit Agent.

[0057] In the following, we will use a clinical case in which "an SLE patient suddenly develops epileptic seizures accompanied by a high fever" to explain in detail the complete execution logic from S10 to S60 in the embodiment of this application.

[0058] Phase 1: Understanding intent and building standardized queries (S10-S20)

[0059] When a physician entered the query, "A patient with SLE has developed a high fever and status epilepticus. What is the differential diagnosis: neuropsychiatric lupus or central nervous system infection? Is steroid pulse therapy recommended?", the system first initiated the "Acquisition and Orchestration (S10)" process. The orchestration agent determined that this was a complex task involving two subtasks, "Differential Diagnosis" and "Treatment Decision-Making," and therefore set a high-priority validation threshold due to the potential medical risk.

[0060] Subsequently, the data flowed into the query analysis (S20) stage. The query analysis agent did not simply search the raw text entered by the user, but rather performed processing such as reconstructing and expanding the raw text. For example, using entity linking technology, the description "butterfly rash" was standardized to SNOMED CT:55464009 (Systemic lupus erythematosus), where "Systemic lupus erythematosus" refers to systemic lupus erythematosus. Also, "epilepsy" was expanded to SNOMED CT:246545002 (Generalized seizure), where "Generalized seizure" refers to generalized seizures. Furthermore, complex graph query statements were constructed for the knowledge graph search agent with the aim of identifying overlapping areas of symptoms between SLE and central nervous system infections, as well as drug contraindication relationships.

[0061] Second stage: Parallel search of multi-source heterogeneous evidence (S30)

[0062] At this stage, the orchestration agent scheduled multiple search agents in parallel. For example, a dense search agent identified key literature in the vector database that stated, "High fever (>39°C) and decreased cerebrospinal fluid glucose more strongly suggest infection." Simultaneously, a knowledge graph search agent returned the crucial pathway (Methylprednisolone) → (CNS Infection), indicating that methylprednisolone is contraindicated in cases of central nervous system infection risk. A sparse search agent extracted a clear textual definition from the EULAR guidelines: "NPSLE is a diagnosis of exclusion." Information from these different dimensions was aggregated as a pool of primary evidence (i.e., the aforementioned set of primary information).

[0063] Third stage: Dynamic verification and collision resolution (S40)

[0064] This was a key innovation step in this embodiment. The information verification and criticism agent scanned the above pool of evidence and sharply detected the following series of logical inconsistencies: Source A (early literature) recommended steroid pulse therapy for neuropsychiatric lupus, while Source B (infectious disease guidelines) warned that steroids could lead to the spread of infection and potentially fatal consequences.

[0065] The information verification and criticism agent, using an NLI model, determined that this was a "conditional dependency collision." Given the patient's "high fever," a strong indicator of infection, the information verification and criticism agent, based on evidence-based medicine rules, reduced the weight of source A (referring to the source confidence score) and activated a re-retrieval loop. This system automatically generated a new query, "cerebrospinal fluid glucose critical value for bacterial meningitis and cerebrospinal fluid glucose critical value for lupus," and further obtained the quantitative criterion that "cerebrospinal fluid glucose <40 mg / dL strongly suggests bacterial meningitis." This completely established the contextual environment of "high infection risk."

[0066] Stage 4: Grounding Generation and Audit Tracking (S50-S60)

[0067] Based on the context after "cleansing" and "verification," the ground-generating agent constructed the final response (i.e., the aforementioned final response). This final response was not ambiguous, but clearly warned: "The clinical features, given the patient's high fever and decreased cerebrospinal fluid glucose, more strongly support a central nervous system infection. To avoid exacerbating a potentially fatal infection, it is strongly recommended to discontinue methylprednisolone pulse therapy until the infection is ruled out."

[0068] Simultaneously, the audit agent completed end-to-end process recording in the background. This resulted in a JSON-formatted audit log containing timestamps, hash values ​​of the decision-making logic, and metadata for all cited sources. This log not only recorded the advice issued by the system but also documented why the system "did not implement the recommendation for steroid pulse therapy"—specifically, because it detected inconsistencies between specific guidelines. This mechanism fully complies with regulatory requirements for interpretability and traceability of medical software (SaMD).

[0069] This embodiment demonstrates the unique advantages of this method in handling high-risk medical decision-making. By introducing a hybrid retrieval that combines knowledge graphs and vector search, the system captures logical taboos at a physical level that are difficult to discover with pure text search. Furthermore, it effectively blocks "hallucinations"—errors that could lead to medical errors—through a "brake mechanism" of a dynamic verification layer (i.e., information verification and critical agents). Finally, tamper-proof audit tracking provides legal guarantees for achieving compliance in medical AI.

[0070] In practical applications, medical data exhibits extremely strong multimodal heterogeneity. A complete medical history typically includes both text (e.g., chief complaint, medical history, radiology report, etc.) and images (e.g., X-ray images, CT images, MRI images, pathology section images, etc.). In clinical practice, a very specific and dangerous problem exists: "image-report inconsistency." For example, a radiologist might use an old template to create a text report stating "lung fields are clear," but a recent chest X-ray (image) actually shows "infiltration in the right lower lung field." Therefore, in such cases, if the search agent searches only the text report, the field generation agent (LLM) may generate the conclusion that "the patient's lungs are normal" based on the incorrect text, potentially leading to a serious medical error. This is a "specific medical problem" that cannot be solved with general-purpose RAGs.

[0071] To solve the above technical problems, embodiments of the present invention refer to an image search agent and a cross-modal alignment module incorporated into the information verification and criticism agent.

[0072] In some embodiments, the query analysis agent generates not only text queries but also visual query vectors. For example, in response to the query "check for signs of pneumonia," the query analysis agent not only generates the keyword "pneumonia" but also generates visual embedding vectors representing "characteristics of pneumonia" using a pre-trained medical visual-language model.

[0073] In specific implementations, dense and sparse search agents function as text search agents, capable of processing Electronic Health Records (EHR) text and playing a role in retrieving physicians' diagnostic reports from the EHR text library. On the other hand, the search agents in this embodiment further include an image search agent. Unlike the text search agent, this image search agent retrieves relevant DICOM slice series from a Picture Archiving and Communication System (PACS). This image search agent has a content-based image retrieval (CBIR) function and can identify lesion areas based on visual similarity.

[0074] In some embodiments, the specific processing flow of the cross-modal mutual verification logic within the information verification and criticism agent is as follows: 1. Feature Projection: The information verification and criticism agent receives a text fragment T (e.g., "Reports: Normal chest") and an image fragment I (corresponding CXR image). The agent uses a common multimodal encoder (e.g., Contrastive Language—Image Pre-training, a medical-specific refinement of CLIP) to project the text fragment T and image fragment I into the same high-dimensional semantic space, and generates a text embedding vector E. Tand image embedding vector E I Obtain it. 2. Consistency Scoring: The cosine similarity between the text embedding vector ET and the image embedding vector EI is calculated as follows.

number

[0075] Embodiments of the present invention utilize the heterogeneity in the physical representation of medical images and text to achieve "visual grounding" through computer vision technology. This not only solves the problem of data heterogeneity but also directly prevents hallucinations in RAG caused by errors in recording medical history.

[0076] This application further provides a verifiable search extension generation device based on cooperative multi-agent technology, as shown in Figure 5, the device is A data collection module 100 for receiving user queries via an orchestration agent to retrieve user queries and perform task decomposition and resource scheduling, A query analysis module 200 is provided to pass user queries to a query analysis agent, which then performs intent analysis and optimization of the user queries and generates optimized query validators. A search module 300 for distributing the optimized query validants to multiple search agents to perform parallel searches and outputting a set of raw information, A verification module 400 is provided for inputting the aforementioned raw information set into an information verification and criticism agent, and for performing dynamic verification processing on the raw information set, including contradiction detection, confidence scoring, and noise filtering, to generate a verified context. An output module 500 for passing the verified context to a ground generation agent, and for the ground generation agent to generate a final response based on the verified context, An audit module 600 for generating audit tracking data based on the data flow of the orchestration agent, query analysis agent, search agent, information verification / criticism agent, and ground generation agent captured and recorded in real time by the audit agent, Includes.

[0077] In another embodiment, the verifiable search extension generator based on the cooperative multi-agent described above includes a processor, which is used to execute the program modules located in memory, the program modules including a data acquisition module 100, a query analysis module 200, a search module 300, a verification module 400, an output module 500, and an audit module 600.

[0078] The functional implementation of each module of the verifiable search extension generation device based on the cooperative multi-agent described above corresponds to each step in the embodiment of the verifiable search extension generation method based on the cooperative multi-agent described above, and a detailed explanation of its functions and implementation process will not be repeated here.

[0079] This application further provides a computer-readable recording medium on which a program is stored, the computer-readable recording medium being a medium for storing data, including but not limited to floppy disks, optical disks, hard disks, flash memory, USB memory, and / or memory sticks, and the computer being a general-purpose computer, a dedicated computer, a computer network, or other programmable device. For the operating process, operating details, and technical effects of the computer-readable recording medium according to this embodiment, refer to the above embodiment of the verifiable search extension generation method based on cooperative multi-agent, and a detailed description is omitted here.

[0080] This application further provides a computer program product including computer instructions, wherein when the computer instructions are executed by a processor, steps of a verifiable search extension generation method based on a cooperative multi-agent system as described in any of the above embodiments are implemented.

[0081] Those skilled in the art will understand that all or some of the processes in the methods of the embodiments described above can be instructed to be executed by a computer program on the relevant hardware. The computer program may be stored on a non-volatile computer-readable recording medium and may include the steps of each embodiment of the methods described above when the computer program is executed. Here, any reference to memory, storage, database, or other medium used in each embodiment of this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. For example, RAM can be obtained in several forms, including, but is not limited to, static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), extended SDRAM (ESDRAM), synchlink DRAM (SLDRAM), rambus direct RAM (RDRAM), direct rambus dynamic RAM (DRDRAM), and rambus dynamic RAM (RDRAM).

[0082] Each of the technical features of the embodiments described above can be combined in any way, and for the sake of brevity, not all possible combinations of each of the technical features of the embodiments described above have been described. However, as long as there is no inconsistency in these combinations of technical features, they should be considered to fall within the scope described herein. The above describes only some embodiments of this application, and although the description is specific and detailed, it should not be interpreted as limiting the scope of the patent of the invention. A person skilled in the art can make several modifications and improvements without departing from the spirit of this application, and all of these fall within the scope of protection of this application. Therefore, the scope of patent protection of this application should be determined by the appended claims.

Claims

1. A computer device including memory, a processor, and a computer program stored in memory and executable on the processor, wherein when the processor executes the computer program, a verifiable search extension generation method based on cooperative multi-agent is realized, and the method is This involves receiving user queries via an orchestration agent to retrieve them and perform task decomposition and resource scheduling, The user query is passed to a query analysis agent, which performs intent analysis and optimization of the user query and generates an optimized query validator. The optimized query validator is distributed to multiple search agents to perform parallel searches and output a set of raw information. The process involves inputting the aforementioned raw information set into an information verification and criticism agent, performing dynamic verification processing on the raw information set by the information verification and criticism agent, and generating a verified context, wherein the dynamic verification processing includes inconsistency detection, confidence scoring, and noise filtering of the raw information set. The verified context is passed to the ground generation agent, and the ground generation agent generates a final response based on the verified context. Based on the data flow of the orchestration agent, query analysis agent, search agent, information verification / criticism agent, and ground generation agent, which are captured and recorded in real time by the audit agent, audit tracking data is generated. A computer device characterized by including the following:

2. A computer device according to claim 1, Performing dynamic verification processing of the original information set by the aforementioned information verification and criticism agent is: If the information verification and criticism agent detects that the confidence score of the original information set falls below a predetermined threshold, it sends a re-search command to the orchestration agent. The orchestration agent generates a targeted query valid based on the received re-search command, and transmits the targeted query valid to the multiple heterogeneous search agents. A computer device characterized by including the following:

3. The computer device according to claim 2, Performing dynamic verification processing of the original information set by the aforementioned information verification and criticism agent is: In the aforementioned set of raw information, identify and mark logical inconsistencies between information fragments retrieved from different sources. Assigning a source confidence score to each of the aforementioned information fragments, Filtering information fragments from the aforementioned raw information set whose confidence score falls below a predetermined threshold, Integrating filtered information fragments to form the verified context, A computer device characterized by including the following:

4. The computer device according to claim 2, Distributing the optimized query validator across multiple search agents to perform parallel searches and outputting the raw information set is: The dense search agent and the sparse search agent are activated, the dense search agent performs semantic vector matching on the optimized query valid, and the sparse search agent performs keyword matching on the optimized query valid based on a keyword model to obtain first search data. The process involves launching a knowledge graph search agent, scanning nodes and relationships within a structured knowledge graph, and obtaining second-level search data. This involves calling a real-time network search agent to obtain current events data and then obtaining a third set of search data. The first search data, the second search data, and the third search data are aggregated to obtain the original information set, A computer device characterized by including the following:

5. The computer device according to claim 2, A computer device characterized in that the ground generation agent is constructed based on a large-scale language model.

6. The computer device according to claim 2, A computer device characterized in that the data flow includes at least the user's initial query, the reconstruction result of the query analysis agent, the data source of the search agent, and the specific text chunks that were searched, and the audit tracking data has a timestamp.

7. A computer device according to claim 1, The computer device is applied to the diagnosis and treatment of diseases, the user query includes disease information, the optimized query validator includes a reconstructed or extended medical ontology language corresponding to the disease information, the raw information set includes graph pathway relationships between drugs and contraindications, or between symptoms and diseases, corresponding to the optimized query validator, and the final response includes disease diagnosis results and treatment suggestions corresponding to the disease information. Inputting the aforementioned raw information set into an information verification and criticism agent, and performing dynamic verification processing of the raw information set by the information verification and criticism agent, is: Inputting the aforementioned set of raw information into the aforementioned information verification and criticism agent, The information verification and criticism agent identifies and marks logical contradictions between information fragments retrieved from different sources in the original information set, wherein the logical contradictions include conflicts of dosage. The aforementioned information verification and criticism agent assigns a source confidence score to the information fragments corresponding to the logical contradictions in accordance with the rules of evidence-based medicine, and sends a re-search command to the orchestration agent. The orchestration agent generates a targeted query valid based on the received re-search command, and sends the targeted query valid to the multiple search agents to perform the re-search. A computer device characterized by including the following:

8. A verifiable search extension generation device based on cooperative multi-agent systems, A data collection module for receiving user queries via an orchestration agent to retrieve user queries and perform task decomposition and resource scheduling, A query analysis module for passing user queries to a query analysis agent, which then performs intent analysis and optimization of the user queries and generates optimized query validators. A search module for distributing the optimized query validators across multiple search agents to perform parallel searches and outputting a set of raw information, A verification module for inputting the aforementioned raw information set into an information verification and criticism agent, performing dynamic verification processing on the raw information set including contradiction detection, confidence scoring, and noise filtering by the information verification and criticism agent, and generating a verified context. An output module for passing the verified context to a ground generation agent, and for the ground generation agent to generate a final response based on the verified context, An audit module for generating audit tracking data based on the data flow of the orchestration agent, query analysis agent, search agent, information verification / criticism agent, and ground generation agent captured and recorded in real time by the audit agent, An apparatus characterized by including

9. A computer-readable recording medium, wherein a program is stored on the computer-readable recording medium, and when the program is executed by a processor, a verifiable search extension generation method based on cooperative multi-agent is realized, and the method is This involves receiving user queries via an orchestration agent to retrieve them and perform task decomposition and resource scheduling, The user query is passed to a query analysis agent, which performs intent analysis and optimization of the user query and generates an optimized query validator. The optimized query validator is distributed to multiple search agents to perform parallel searches and output a set of raw information. The process involves inputting the aforementioned raw information set into an information verification and criticism agent, performing dynamic verification processing on the raw information set by the information verification and criticism agent, and generating a verified context, wherein the dynamic verification processing includes inconsistency detection, confidence scoring, and noise filtering of the raw information set. The verified context is passed to the ground generation agent, and the ground generation agent generates a final response based on the verified context. Based on the data flow of the orchestration agent, query analysis agent, search agent, information verification / criticism agent, and ground generation agent, which are captured and recorded in real time by the audit agent, audit tracking data is generated. A computer-readable recording medium characterized by including the following:

10. A computer program that, when executed by a processor, realizes a verifiable search extension generation method based on cooperative multi-agent, the method is This involves receiving user queries via an orchestration agent to retrieve them and perform task decomposition and resource scheduling, The user query is passed to a query analysis agent, which performs intent analysis and optimization of the user query and generates an optimized query validator. The optimized query validator is distributed to multiple search agents to perform parallel searches and output a set of raw information. The process involves inputting the aforementioned raw information set into an information verification and criticism agent, performing dynamic verification processing on the raw information set by the information verification and criticism agent, and generating a verified context, wherein the dynamic verification processing includes inconsistency detection, confidence scoring, and noise filtering of the raw information set. The verified context is passed to the ground generation agent, and the ground generation agent generates a final response based on the verified context. Based on the data flow of the orchestration agent, query analysis agent, search agent, information verification / criticism agent, and ground generation agent, which are captured and recorded in real time by the audit agent, audit tracking data is generated. A computer program characterized by including the following.

Images