Multi-agent separation intelligent dialogue methods, systems, devices, storage media, and program products
By employing a multi-agent separation intelligent dialogue method, and utilizing the independent permissions and discrete state routing of the generating and verifying agents, the problems of inconsistent content generation and inaccurate behavior control in large language models are solved, achieving higher precision dialogue control and content accuracy.
Patent Information
- Application Number
- CN202611466431.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-20
- Publication Date
- 2026-11-17
AI Technical Summary
Large language models suffer from issues such as inconsistency between generated content and authorized knowledge sources (illusion) and the inability to precisely control dialogue proxy behavior according to preset rules when generating dialogue content.
The intelligent dialogue method adopts a multi-agent separation approach. The generation agent generates questions and receives user input. The discrete state routing engine is used to determine the discrete state. When the input error state count reaches the threshold, the system switches to the verification agent, which retrieves the original text of the knowledge source and presents it. The generation agent and the verification agent are independent of each other, with separate permissions and no sharing of dialogue context.
It effectively avoids cyclical verification errors between AI models, improves the accuracy of behavior control, reduces generative error rate and behavior control bias, and ensures that the generated content is consistent with the authorized knowledge source.
Smart Images

Figure CN123115351A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a multi-agent separation intelligent dialogue method, system, device, storage medium, and program product. Background Technology
[0002] Large-scale language models suffer from two types of technical defects when generating dialogue content: Problem 1, the generated content is inconsistent with the authorized knowledge source ("illusion"); Problem 2, the behavior of the dialogue agent cannot be precisely controlled according to preset rules (uncontrolled probing depth and inaccurate timing of role switching).
[0003] Regarding issue one (illusion control): Existing solutions employ post-generation multi-step verification (such as the multimodal fact verification and dynamic constraints in CN121980006A) or a closed-loop governance solution covering the entire process (data layer, model layer, inference layer, and application layer). These solutions fall under the category of "post-event detection," where verification steps are executed after generation, and the verification agent and generation agent are of the same type, resulting in a circular verification problem of "using LLM to verify LLM."
[0004] Regarding question two (dialogue behavior control): existing solutions use single agent continuous scoring parameter tuning (adjusted through continuous scoring parameters such as behavior abnormality, inquiry window, and inquiry acceptance rate) or calculate the continuity of topics based on context. The shortcomings of these solutions are: (1) a single agent cannot achieve agent-level isolation of different dialogue strategies - different strategies such as follow-up questioning, error correction, and demonstration share the same agent configuration, which interferes with each other; (2) continuous scoring is not effective in discrete decision-making scenarios (such as "the nth error triggers role switching").
[0005] Therefore, there is an urgent need to invent a new intelligent dialogue method based on large language models to overcome the problems of inconsistency between generated content and authorized knowledge sources and the inability to accurately control dialogue agent behavior in existing technologies. Summary of the Invention
[0006] In view of this, embodiments of the present invention provide a multi-agent separation intelligent dialogue method, system, device, storage medium, and program product, which at least partially solves the problems existing in the prior art.
[0007] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.
[0008] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:
[0009] According to one aspect of the present invention, a multi-agent separation intelligent dialogue method is provided, the method comprising:
[0010] The generation agent generates questions in each round of dialogue and receives user input;
[0011] The discrete state routing engine performs discretization state determination on the user input to obtain the user input state, which is either correct input, incorrect input, or ambiguous input.
[0012] When the input error count reaches the preset threshold, switch to the verification agent;
[0013] The verification agent retrieves the original text of the knowledge source in read-only mode and presents the original text of the knowledge source to the user. The verification agent does not generate new content.
[0014] In some embodiments, the method further includes:
[0015] Before outputting the question content generated by the agent to the user, the output paragraphs of the question content are marked with structured source pointers;
[0016] Determine if the source pointer of the output paragraph is in the preset authorized source list;
[0017] If the source pointer of the output paragraph is in the default list of authorized sources, the output paragraph will be output to the user.
[0018] If the source pointer of the output paragraph is not in the default list of authorized sources, the output paragraph will be intercepted and output to the user will be prohibited.
[0019] In some embodiments, the generation agent and the verification agent are independent agent processes;
[0020] The permissions for generating the proxy are content generation and source labeling;
[0021] The verification agent is authorized to retrieve knowledge sources and present the original text, but is prohibited from generating new content.
[0022] In some embodiments, the discrete state routing engine performs discretization state determination on the user input to obtain the user input state, including:
[0023] The discrete state routing engine determines whether the semantics of the user input match the semantics of the preset standard answer.
[0024] If the semantics of the user input match the semantics of the preset standard answer, then the user input status for this round is "correct input".
[0025] If the semantics of the user input do not match the semantics of the preset standard answer, then the user input status for this round is marked as an input error, and the count of input error statuses is accumulated.
[0026] If the user input contains uncertain expressions, the user input state for this round is fuzzy, and the generator will output a prompt message requiring definite input.
[0027] In some embodiments, when the input error state count reaches a preset counting threshold, the system switches to a verification agent, which retrieves the original knowledge source text and presents it to the user, including:
[0028] The system continuously monitors user input status in the dialogue flow. When the count of input error status reaches a preset threshold, it generates statements to be verified and retrieval instructions based on user input.
[0029] Send the statement to be verified and the search instruction to the verification agent;
[0030] The verification agent performs a knowledge source retrieval in read-only mode based on the statement to be verified and the retrieval instruction to obtain the original text of the knowledge source corresponding to the statement to be verified.
[0031] The original text of the knowledge source will be presented to the user.
[0032] In some embodiments, the verification agent runs in parallel with the generation agent, and the verification agent is triggered after the input error state count reaches a preset count threshold.
[0033] According to one aspect of the present invention, a multi-agent separation intelligent dialogue system is provided, the system comprising:
[0034] Generate a proxy to generate questions in each round of dialogue and receive user input;
[0035] The discrete state routing engine is used to discretize the state of user input and obtain the user input state. When the count of input error states reaches a preset count threshold, it switches to the verification agent.
[0036] The verification proxy is used to retrieve the original text of the knowledge source and present the original text of the knowledge source to the user.
[0037] According to one aspect of the present invention, a multi-agent separation intelligent dialogue device is provided, comprising:
[0038] Memory; and
[0039] A processor coupled to the memory is configured to execute a multi-agent separation intelligent dialogue method as described above based on instructions stored in the memory.
[0040] According to one aspect of the present invention, a computer-readable storage medium is provided having computer instructions stored thereon that, when executed by a processor, implement a multi-agent separation intelligent dialogue method as described in any of the preceding claims.
[0041] According to one aspect of the present invention, a computer program product is provided, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions that, when executed by a computer, cause the computer to implement the steps of the multi-agent separation intelligent dialogue method as described in any of the preceding claims.
[0042] This invention provides a multi-agent separation intelligent dialogue method, system, device, storage medium, and program product. The method includes: a generation agent generating questions in each round of dialogue and receiving user input; a discrete state routing engine discretizing the user input to determine the user input state, which can be either correct, incorrect, or ambiguous; when the count of incorrect input states reaches a preset threshold, switching to a verification agent; and the verification agent retrieving the original knowledge source text and presenting it to the user. This invention effectively avoids cyclic verification errors between AI models and improves behavior control accuracy through discrete state routing, thereby reducing the generative error rate and behavior control bias in AI dialogue. Attached Figure Description
[0043] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.
[0044] Figure 1 The flowchart illustrates some embodiments of the intelligent dialogue method with multi-agent separation according to the present invention. Detailed Implementation
[0045] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0046] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0047] Figure 1 The flowchart illustrates some embodiments of a multi-agent separation intelligent dialogue method according to the present invention.
[0048] like Figure 1 As shown, the multi-agent separation intelligent dialogue method according to an embodiment of the present invention may include steps S100, S200, S300 and S400.
[0049] In step S100, the generation agent generates questions in each round of dialogue and receives user input.
[0050] In some embodiments, before outputting the question content generated by the generation agent to the user, each output paragraph of the question content is marked with a structured source pointer; it is determined whether the source pointer of the output paragraph is in the preset authorized source list; if the source pointer of the output paragraph is in the preset authorized source list, the output paragraph is output to the user; if the source pointer of the output paragraph is empty or not in the preset authorized source list, the output paragraph is intercepted and output to the user is prohibited.
[0051] In some embodiments, the above source pointer includes a source type and an identifier, wherein the source type includes textbooks, DOIs, and guides.
[0052] In step S200, the discrete state routing engine discretizes the user input to determine the user input state, which is either correct, incorrect, or ambiguous. For ease of description, correct input is denoted as state S1, incorrect input as state S2, and ambiguous input as state S3.
[0053] In some embodiments, a discrete state routing engine determines whether the semantics of the user input match the semantics of a preset standard answer.
[0054] If the semantics of the user input match the semantics of the preset standard answer, then the user input status for this round is "correct input," and the generated agent continues to respond, which may speed up the process.
[0055] If the semantics of the user input do not match the semantics of the preset standard answer, then the user input status for this round is considered an input error, and the count of input error statuses is accumulated; if it is the first input error, then a proxy questioning reasoning chain is generated.
[0056] If the user input contains uncertain expressions ("maybe", "perhaps", etc.), then the user input state for this round is ambiguous, and the generation agent will output a prompt message requiring definite input.
[0057] In some embodiments, each user input state is hard-bound to a specific agent role and dialogue behavior, and role switching is triggered by the number of state transitions (rather than continuous scoring). The Nth time entering an error state directly triggers the intervention of the verification agent.
[0058] In step S300, when the input error status count reaches a preset count threshold, the system switches to the verification agent.
[0059] In some embodiments, the preset counting threshold is configured to be 2 by default.
[0060] In step S400, the verification agent retrieves the original text of the knowledge source in read-only mode and presents the original text of the knowledge source to the user. The verification agent does not generate new content.
[0061] In some embodiments, the user input status in the dialogue flow is continuously counted. When the input error status count reaches a preset counting threshold, a statement to be verified and a retrieval instruction are generated based on the user input. The statement to be verified and the retrieval instruction are sent to the verification agent. The verification agent performs a knowledge source retrieval in read-only mode based on the statement to be verified and the retrieval instruction to obtain the original text of the knowledge source corresponding to the statement to be verified. The original text of the knowledge source is then presented to the user.
[0062] In some embodiments, the generation agent and the verification agent are independent agent processes running on the same or different computing devices, exchanging data through inter-process communication or network communication. The generation agent's permissions are content generation and source annotation; the verification agent's permissions are authorized knowledge source retrieval and original text presentation, and the verification agent is prohibited from generating new content. The generation agent is configured with teaching system prompts, and the verification agent is configured with independent retrieval system prompts.
[0063] The validation agent does not participate in judging right or wrong; it only performs original text retrieval and presentation, thus avoiding the circular validation problem of using a second LLM to judge the first LLM.
[0064] In some embodiments, the verification agent and the generation agent run in parallel, the generation agent runs continuously, and the verification agent is triggered after the input error state count reaches a preset count threshold.
[0065] In other embodiments, a simplified implementation of multi-role separation within a single process is provided. As a simplified implementation in resource-constrained or early-stage deployment, the separation of roles between the generation agent and the verification agent can be achieved through system prompt word engineering within a single model call.
[0066] In the system prompts during the same call, different behavioral permissions and constraints were issued to the two agent roles: the content generation role was granted the permission to generate teaching content and mark the source of each item; the retrieval and verification role was limited to only performing authorized knowledge source retrieval and presenting the original text, without actively generating answers or outputting subjective judgments.
[0067] When a conversation continuously enters an error state a preset number of times, a hard rule in the system prompts a role switch, changing the questioning from the content generation role to the retrieval and verification role, presenting the original knowledge source text or search results, and ceasing further questioning. In this simplified implementation, the permission boundaries of each role are defined through rules at the prompt level, making it suitable for single-process deployment scenarios.
[0068] In other embodiments, a process-level agent separation is provided within a dialogue. The generation agent and verification agent are deployed as independent processes, running on the same or different computing devices. They exchange data via message queues or inter-process communication interfaces, but do not share the dialogue context or memory space. The generation agent process loads the system configuration required for content generation, including the hard rule that "each piece of generated content must include a source pointer." The verification agent process runs as an independent process, loading a retrieval-specific system configuration and mounting the authorized knowledge source index only with read-only permissions. The verification agent process does not hold access to the content generation interface, and its system configuration does not contain generation instructions; the generation agent process also does not hold write permissions to the knowledge source retrieval interface. This permission isolation can be enforced at the operating system process level or container level (e.g., independent container instances, different user permissions, or network namespaces). Source Pointer Filter: Deployed as an output pipeline interception layer independent of the two proxies mentioned above (e.g., deployed on an API gateway or output middleware), it performs source pointer verification on each output segment before the output of the generation agent is sent to the user terminal via the network interface. It extracts the source_pointer, compares it with the authorized source list, and intercepts outputs that fail verification at the filter layer, preventing them from entering subsequent transmission. This filter runs on the data path between the generation agent and the user terminal, and its verification logic does not rely on the generation capabilities of any large language model. Runtime Sequence: The generation agent runs continuously and responds to the dialogue; the verification agent runs in parallel and is always active or started on demand, performing retrieval and presentation only when it receives statements to be verified and retrieval instructions; the two agents do not share the dialogue context, the generation agent holds the dialogue history, and the verification agent only receives statements and retrieval instructions related to the current verification task.
[0069] It should be noted that the simplified implementation of multi-role separation within a single process and the process-level separation implementation share the same security principles. Role verification only retrieves and presents information, without generating judgments, and role switching is based on discrete state transition counts. The only difference lies in the implementation level of permission isolation: the single-process implementation achieves this through prompt word rule conventions, while the process-level implementation achieves it through operating system / container-level process isolation, the latter providing stronger and enforceable security guarantees. Those skilled in the art will understand that the separation from prompt word-level role separation to process-level proxy separation represents a progressively increasing level of isolation strength.
[0070] As a non-restrictive example, let's take a dialogue as an example to illustrate the process of state determination and route triggering. The determination logic is independent of the specific domain content of the dialogue, and only depends on the matching state between the input and the standard answer and the state transition count:
[0071] [Dialogue Round 1] Generated Agent: "Please analyze the hemodynamic status of this patient." User input: "Insufficient blood volume, fluid resuscitation is needed." Status determination: S2 (Error - Preset standard answer not found).
[0072] [Dialogue Round 2] Generating Agent: "Why do you think it's low blood volume? Explain step by step." User Input: "It might be... low blood pressure..." State Judgment: S3 (fuzzy - "maybe"), Routing: Generating agent outputs standardized prompts requiring deterministic input.
[0073] [Dialogue Round 3] User inputs: "I'm not sure." State determination: S2 (Error) — This is the second time entering S2. Routing: State transition count reaches N=2 → Force switch to verification agent. Verification agent: "The original text of the textbook is as follows: [Search results]. Please refer to it."
[0074] In another embodiment, a multi-agent separated intelligent dialogue system includes:
[0075] Generate a proxy to generate questions in each round of dialogue and receive user input;
[0076] The discrete state routing engine is used to discretize the state of user input and obtain the user input state. When the count of input error states reaches a preset count threshold, it switches to the verification agent.
[0077] The verification proxy is used to retrieve the original text of the knowledge source and present it to the user.
[0078] In another embodiment, a multi-agent separation intelligent dialogue device includes: a memory; and a processor coupled to the memory, the processor being configured to execute a multi-agent separation intelligent dialogue method as described above based on instructions stored in the memory.
[0079] In another embodiment, a computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the steps of the method corresponding to the method in the embodiment for a multi-agent separation intelligent dialogue method. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable non-transitory storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0080] In another embodiment, a computer program product includes computer program instructions that, when executed by a processor, implement the steps of a multi-agent separation intelligent dialogue method as described above.
[0081] Compared with the prior art, the present invention has the following beneficial technical effects:
[0082] This invention effectively reduces loop verification errors by separating multiple agents. The verification agent does not generate content or output judgments, thus avoiding the trust loop of "using LLM to audit LLM".
[0083] This invention effectively improves the accuracy of behavior control through discrete state routing. Compared with continuous scoring, the behavior consistency of discrete state transitions is more verifiable.
[0084] This invention effectively reduces the content error rate by forcibly verifying the source pointer, cuts off the output path of content without a source at the system level, and does not rely on LLM self-regulation.
[0085] The above technical effects can be verified through comparative experiments: on the same dialogue task set, compare the proportion of no source output when the source pointer filter is enabled and disabled, and the consistency rate of the agent intervention timing when using discrete state routing and continuous scoring mechanisms.
[0086] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.
[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, or alterations made by those skilled in the art using the disclosed technical content shall fall within the protection scope of the present invention.
Claims
1. A multi-agent separation intelligent dialogue method, characterized in that, The method includes: The generation agent generates questions in each round of dialogue and receives user input; The discrete state routing engine performs discretization state determination on the user input to obtain the user input state, which is either correct input, incorrect input, or ambiguous input. When the input error count reaches the preset threshold, switch to the verification agent; The verification agent retrieves the original text of the knowledge source in read-only mode and presents the original text of the knowledge source to the user. The verification agent does not generate new content.
2. The intelligent dialogue method with multi-agent separation according to claim 1, characterized in that, The method further includes: Before outputting the question content generated by the agent to the user, the output paragraphs of the question content are marked with structured source pointers; Determine if the source pointer of the output paragraph is in the preset authorized source list; If the source pointer of the output paragraph is in the default list of authorized sources, the output paragraph will be output to the user. If the source pointer of the output paragraph is not in the default list of authorized sources, the output paragraph will be intercepted and output to the user will be prohibited.
3. The intelligent dialogue method with multi-agent separation according to claim 1, characterized in that, The generation agent and the verification agent are independent agent processes; The permissions for generating the proxy are content generation and source labeling; The verification agent is authorized to retrieve knowledge sources and present the original text, but is prohibited from generating new content.
4. The intelligent dialogue method with multi-agent separation according to claim 1, characterized in that, The discrete state routing engine discretizes the user input to determine its state, resulting in the user input state, including: The discrete state routing engine determines whether the semantics of the user input match the semantics of the preset standard answer. If the semantics of the user input match the semantics of the preset standard answer, then the user input status for this round is "correct input". If the semantics of the user input do not match the semantics of the preset standard answer, then the user input status for this round is marked as an input error, and the count of input error statuses is accumulated. If the user input contains uncertain expressions, the user input state for this round is fuzzy, and the generator will output a prompt message requiring definite input.
5. The intelligent dialogue method with multi-agent separation according to claim 1, characterized in that, When the input error count reaches a preset threshold, the system switches to a verification agent, which retrieves the original knowledge source text and presents it to the user, including: The system continuously monitors user input status in the dialogue flow. When the count of input error status reaches a preset threshold, it generates statements to be verified and retrieval instructions based on user input. Send the statement to be verified and the search instruction to the verification agent; The verification agent performs a knowledge source retrieval in read-only mode based on the statement to be verified and the retrieval instruction to obtain the original text of the knowledge source corresponding to the statement to be verified. The original text of the knowledge source will be presented to the user.
6. The intelligent dialogue method with multi-agent separation according to claim 1, characterized in that, The verification agent runs in parallel with the generation agent, and the verification agent is triggered after the input error state count reaches a preset count threshold.
7. A multi-agent separation intelligent dialogue system, characterized in that, The system includes: Generate a proxy to generate questions in each round of dialogue and receive user input; The discrete state routing engine is used to discretize the state of user input and obtain the user input state. When the count of input error states reaches a preset count threshold, it switches to the verification agent. The verification proxy is used to retrieve the original text of the knowledge source and present the original text of the knowledge source to the user.
8. A multi-agent separation intelligent dialogue device, characterized in that, The device includes: Memory; and A processor coupled to the memory, the processor being configured to execute a multi-agent separation intelligent dialogue method as described in any one of claims 1 to 6 based on instructions stored in the memory.
9. A computer-readable storage medium, characterized in that, It stores computer instructions that, when executed by a processor, implement a multi-agent separation intelligent dialogue method as described in any one of claims 1 to 6.
10. A computer program product, characterized in that, The computer program product includes computer program instructions that, when executed by a processor, implement the steps of the multi-agent separation intelligent dialogue method as described in any one of claims 1 to 6.
Citation Information
Patent Citations
Large model anti-illusion method and system based on multi-modal fact verification and dynamic constraint
CN121980006A
