Intelligent agent supervision method based on enterprise-level intelligent agent platform

Abstract

The invention provides an agent supervision method based on an enterprise-level agent platform, and the method comprises the steps: constructing a supervision agent for each AI agent, the supervision agent comprising a knowledge model, a supervision module and an evaluation module; and the supervision agent carries out generation content supervision on the AI agent in an active instruction, negotiation or passive receiving mode. According to the method, high-degree-of-freedom illusion can be reserved for creative AI, and the inspiration excitation efficiency is improved; dynamic balance artistry and cost are provided for the innovative AI; and the creation type AI illusion is strictly inhibited, and the self strategy and behavior mode are dynamically adjusted according to different tasks and environments, so that the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent agent supervision technology, and specifically to an intelligent agent supervision method based on an enterprise-level intelligent agent platform. Background Art

[0002] The enterprise-level AI agent platform features a two-tier architecture consisting of an "AI technology base + agent applications." AI serves as the technical foundation for building agents. The AI technology base, relying on generative large models, provides agents with underlying capabilities such as generation, analysis, and decision-making. Agents, acting as application vehicles, integrate these AI capabilities, focusing on creative industries like advertising, film, and digital media, providing them with content creation, process automation, and user experience upgrades. For example, agents in the advertising sector can automatically generate brand copy and optimize delivery strategies. In film and television production scenarios, agents can assist with character design and storyboard generation, significantly improving industry efficiency.

[0003] However, the current platform faces a key bottleneck in intelligent agent supervision: a unified static reflection mechanism is difficult to adapt to the phased needs of the creative process. Existing technologies generally use fixed means such as a posteriori verification and prompting engineering constraints to suppress model hallucinations. Although this can improve content accuracy, it ignores the differentiated needs for hallucinations at different stages of the creative industry. In the creative stage (such as advertising metaphor association or animation concept design), high-freedom hallucinations are a source of inspiration, and excessive constraints will lead to limited innovation; in the innovation stage (such as the artistic and cost balance of storyboard design), the hallucination threshold needs to be dynamically adjusted to take into account divergence and feasibility; and in the creative stage (such as brand copy compliance verification or film and television refinement), hallucinations need to be strictly suppressed to ensure content accuracy and implementation.

[0004] In summary, the existing technology has at least the following problems:

[0005] The existing regulatory intelligent body adopts a unified static reflection mechanism, which leads to excessive restrictions on model illusions in the creative stage and hinders the generation of creative inspiration; the innovation stage is unable to dynamically balance creativity and feasibility requirements, and lacks flexibility, resulting in a poor user experience. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the existing technology and provide a regulatory intelligent agent and its regulatory method based on an enterprise-level intelligent agent platform, which can solve the problems of poor regulatory flexibility and over-restricted model illusions in existing regulatory intelligent agents, resulting in poor user experience in using the intelligent agent platform.

[0007] In order to achieve the above objectives and other objectives, the present invention is implemented by including the following technical solutions: As a first aspect, the present invention proposes an intelligent agent supervision method based on an enterprise-level intelligent agent platform, including the steps of: constructing a supervision intelligent agent for each AI intelligent agent, the supervision intelligent agent including a knowledge model, a supervision module and an evaluation module; the supervision intelligent agent generates content supervision for the AI intelligent agent through active instructions, negotiation or passive reception.

[0008] Furthermore, the AI agents include creative, innovative and creative types.

[0009] Furthermore, the knowledge model forms a knowledge graph by converting knowledge data into structured data.

[0010] Furthermore, the knowledge data includes a rule base, a professional domain knowledge base, a real-time dynamic base and a meta-knowledge base.

[0011] Furthermore, the supervision module presets hallucination threshold intervals for different AI agents, and adjusts the hallucination thresholds according to the generated content, external variables, and the correlation of knowledge graphs under different AI agents, resource rationality, and user behavior perception through a dynamic reflection mechanism; the hallucination threshold is used to control the balance between the factual accuracy and creative freedom of the generated content, and the larger the hallucination threshold, the greater the creative freedom.

[0012] Furthermore, the evaluation module performs a multi-dimensional evaluation of the generated content in terms of legality, creativity and rationality, and ranks the generated content according to the reward mechanism.

[0013] Furthermore, the evaluation process of the evaluation module includes predicting conflicts before generating content, real-time monitoring and intercepting cross-border content during content generation, and reflecting and iterating the knowledge model after generating content.

[0014] Furthermore, the evaluation factors of the reward mechanism include legitimacy, creativity, dissemination and commercial value, and the generated content is ranked according to the sum of the weights of all evaluation factors.

[0015] Furthermore, the agent supervision method also includes the steps of: the supervision agent generates personalized guidance questions through real-time interaction data and pushes them to the associated agent cluster; and knowledge sharing is achieved in the agent cluster through a cross-domain knowledge graph.

[0016] As a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent agent supervision method described in the first aspect.

[0017] One or more technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages:

[0018] 1. This invention creates a corresponding supervisory agent for each AI agent, which can supervise the knowledge models and generated data at each stage, retaining the illusion of high freedom for creative AI and improving the efficiency of inspiration; providing a dynamic balance between artistry and cost for innovative AI; and strictly suppressing the illusion of creative AI, with a highly flexible overall supervisory strategy;

[0019] 2. When AI-generated content triggers rigid rules or detects significant risks, the supervisory agent can perform one-way mandatory intervention on the AI agent. In areas where the illusion threshold can be flexibly adjusted, the supervisory agent can conduct multiple rounds of interactive negotiations with the AI agent to achieve a balance between factual accuracy and creative freedom, optimize generated content, and enhance user experience. In low-risk or real-time monitoring stages, the supervisory agent only collects data, which can improve supervision efficiency.

[0020] 3. This invention uses the knowledge graph as the knowledge boundary of the knowledge model, which can reduce model illusions and ensure the accuracy and implementation of content in the creation stage;

[0021] 4. The present invention adopts a dynamic feedback mechanism, which can dynamically adjust the hallucination threshold according to different tasks and environments, and dynamically balance the requirements of creativity and feasibility;

[0022] 5. The present invention introduces a reward mechanism in the evaluation module, which can optimize the generated content and encourage the generation of creative content;

[0023] 6. The present invention is applicable to the dynamic supervision of intelligent platforms in creative industries such as advertising, film and television, and digital media. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Shown is a schematic diagram of the construction process of the knowledge model in the present invention.

[0025] Figure 2 Shown is a schematic diagram of the dynamic adjustment of the supervision module in the present invention.

[0026] Figure 3 Shown is a schematic diagram of the evaluation process of the evaluation module in the present invention.

[0027] Figure 4 Shown is a schematic diagram of the interaction between the supervisory agent and the AI agent in the present invention. DETAILED DESCRIPTION

[0028] The embodiments of the present invention provide a regulatory agent and a regulatory method based on an enterprise-level intelligent agent platform, thereby solving the problems of poor regulatory flexibility and over-restricted model illusions in existing regulatory agents, which lead to poor user experience in using the intelligent agent platform.

[0029] Below, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments described herein. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention, rather than all, are shown in the accompanying drawings.

[0030] The overall idea of the technical solution provided by the present invention is as follows: a method for intelligent agent supervision based on an enterprise-level intelligent agent platform, comprising the steps of: constructing a supervision intelligent agent for each AI intelligent agent, the supervision intelligent agent including a knowledge model, a supervision module and an evaluation module; the supervision intelligent agent generates content supervision for the AI intelligent agent through active instructions, negotiation or passive reception.

[0031] Its main concept is to build a corresponding supervisory agent for each AI agent based on three major components: knowledge model, supervision module, and evaluation module. Specifically, a knowledge model with a knowledge graph as the knowledge boundary is constructed, a supervision module that uses a dynamic feedback mechanism to dynamically adjust the illusion threshold is constructed, and an evaluation module that includes an evaluation mechanism and a reward mechanism is constructed. This allows supervision of knowledge models at each stage and the generation of data, retaining a high degree of freedom illusion for creative AI and improving the efficiency of inspiration; providing innovative AI with a dynamic balance between artistry and cost, and between creativity and feasibility; and strictly suppressing the illusion of creative AI to ensure the accuracy and implementation of content during the creative stage. The overall supervisory strategy is highly flexible, dynamically adjusting its own strategies and behavior patterns according to different tasks and environments to improve the user experience.

[0032] After introducing the basic principles of the present invention, various non-limiting embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0033] Example 1:

[0034] Please see the attached Figure 1-4 This embodiment provides an intelligent agent supervision method based on an enterprise-level intelligent agent platform, including the following steps:

[0035] S1. Build a corresponding supervisory agent for each AI agent to supervise the AI agent; the supervisory agent includes a knowledge model, a supervision module, and an evaluation module;

[0036] S2. The supervisory agent supervises the generated content of the AI agent through active instructions, negotiation, or passive reception;

[0037] S3. The supervisory agent generates personalized guidance questions through real-time interactive data and pushes them to the associated agent cluster; knowledge sharing is achieved in the agent cluster through cross-domain knowledge graphs to achieve the purpose of improving user experience.

[0038] In step S1, AI agents in the creative industry are divided into creative categories (such as AI writing assistants, AI painting tools, AI advertising creativity, etc.), innovative categories (such as AI design optimization, AI content enhancement, etc.), and creative categories (such as AI video production, AI script writing, etc.). Each supervisory agent consists of three major components: a knowledge model, a supervisory module, and an evaluation module.

[0039] like Figure 1 As shown, the knowledge model is formed by converting and modeling knowledge data into structured data. Knowledge data includes a rule base, a professional domain knowledge base, a real-time dynamic database, and a meta-knowledge base. The rule base stores industry standards, legal provisions, and ethical guidelines, imposing rigid constraints on content to ensure the legality of content production. The professional domain knowledge base focuses on the creative industry and stores general industry knowledge and vertical industry domain knowledge. The real-time dynamic database stores minute-by-minute network dynamics and industry data. The meta-knowledge base records the original source of each fact. Knowledge data can be converted into structured data based on Neo4j's data storage and retrieval, forming a structured knowledge network for the creative industry, or knowledge graph. This knowledge model then enables automatic knowledge generation, reasoning, and continuous iteration.

[0040] like Figure 2 As shown, the supervision module can flexibly adjust the hallucination threshold through a dynamic reflection mechanism. The hallucination threshold is a dynamic fault-tolerance parameter used to control the balance between factual accuracy and creative freedom in generated content. The supervision module adjusts the hallucination threshold as follows: First, the knowledge graph relevance, resource rationality, and user behavior perception of creative, innovative, and creative AI agents are analyzed, and a basic hallucination threshold range is preset. For example, the hallucination threshold range for creative AI agents can be 8-9, for innovative AI agents can be 5-6, and for creative AI agents can be 2-3. Then, through a dynamic feedback mechanism, the hallucination threshold is adaptively adjusted based on the generated content or external variables, combined with factors such as knowledge graph relevance, resource rationality, and user behavior perception.

[0041] The knowledge graph relevance refers to determining the relevance threshold through the common nodes and shortest paths between entities in the knowledge graph, and controlling the knowledge boundary based on the relevance threshold. For example, if the relevance is greater than 0.9, it is a high threshold, retaining strongly related knowledge; if the relevance is greater than 0.5, it is a low threshold, allowing weakly related knowledge. Resource rationality refers to the rationality of the computing resources used to predict content generation, such as the rationality of resource parameters such as concurrency, resource utilization, and intelligent scheduling. For example, under high load, the hallucination threshold can be increased to reduce the generation of low-hallucination content. User behavior perception refers to the feedback on hallucinated content during the generation of interactions, such as the number of corrections, follow-up questions, and adoptions.

[0042] like Figure 3 As shown in Figure 1, the evaluation module includes an evaluation mechanism and a reward mechanism. The evaluation mechanism evaluates the three dimensions of legality, creativity, and rationality of generated content and user behavior. Legality refers to ensuring the compliance of user behavior and generated content based on the rule base; creativity refers to ensuring that generated content does not exceed the set illusion threshold range; rationality refers to the rationality of resource parameter indicators such as computing resource utilization, concurrency, and intelligent scheduling of the generated content. Specifically, the evaluation process involves assessments in three stages: before, during, and after generation, to achieve risk control and quality optimization. Testing is performed before content generation to predict conflicts; real-time monitoring is performed during content generation to intercept content that exceeds the set illusion threshold range; and after content generation, reflection and identification of error causes are carried out, and the knowledge model is iterated.

[0043] The reward mechanism is introduced to optimize the output order of generated content and encourage the production of creative content. The reward mechanism weights the generated content and user behavior based on factors such as legality, creativity, virality, and commercial value, and sorts the generated content by weighted sum from highest to lowest. Legality refers to ensuring the compliance of user behavior and generated content based on a rule base. Illegal content is weighted 0, while legal content receives a 10% weight. Creativity compares the difference between generated content and the knowledge graph based on a set illusion threshold. Generated content within the illusion threshold receives a 50% weight. Content exceeding the illusion threshold receives a 10% weight reduction for every unit of difference from the 50% weight. Virality refers to the sharing rate and the amount of derivative works generated by user behavior. Weights are dynamically adjusted based on a weekly fluctuation curve. Increases in both the sharing rate and the amount of derivative works receive a 20% weight, while increases in either the sharing rate or the amount of derivative works receive a 10% weight. Commercial value refers to the cost per conversion, with each conversion receiving a 1% weight increase.

[0044] When adding up the weights of legality, creativity, dissemination and commercial value, legality is given the highest priority. Once illegal content is detected, the weight sum will be reduced to zero.

[0045] like Figure 4 As shown in step S2, proactive instruction issuance refers to the unilateral and mandatory intervention of the supervisory agent in the AI agent when the generated content of the AI agent triggers a rigid rule (such as a legal red line) or the supervisory agent detects a major risk. Negotiation refers to the process in which the supervisory agent and the AI agent reach a consensus through multiple rounds of interaction during a stage where the illusion threshold can be flexibly adjusted (such as the innovation stage). Passive reception refers to the process in which the supervisory agent only collects data for subsequent rule optimization during the low-risk or real-time monitoring stage and does not intervene in the real-time generation of the AI agent.

[0046] Example 2:

[0047] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method of the first embodiment is implemented.

[0048] In summary, the intelligent agent supervision method provided by the present invention can supervise the knowledge models and generated data at each stage, retain a high degree of freedom illusion for creative AI, and improve the efficiency of inspiration stimulation; provide a dynamic balance between artistry and cost for innovative AI; and strictly suppress the illusion of creative AI, and the overall supervision strategy is highly flexible; using the knowledge graph as the knowledge boundary of the knowledge model can reduce the model illusion and ensure the accuracy and implementation of the content in the creation stage; the supervision module adopts a dynamic feedback mechanism, which can dynamically adjust the illusion threshold according to different tasks and environments, and dynamically balance the creativity and feasibility requirements; the evaluation module introduces a reward mechanism, which can optimize the generated content and encourage the generation of creative content.

[0049] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the same technology of the present invention, the present invention is also intended to include these modifications and variations.

Claims

1. An intelligent agent supervision method based on an enterprise-level intelligent agent platform, characterized in that: Including steps: Construct a supervisory agent for each AI agent, wherein the supervisory agent includes a knowledge model, a supervisory module, and an evaluation module; The supervisory agent supervises the generated content of the AI agent through active instructions, negotiation or passive reception.

2. The intelligent agent supervision method based on the enterprise-level intelligent agent platform according to claim 1 is characterized in that: The AI agents include creative, innovative and creative types.

3. The intelligent agent supervision method based on the enterprise-level intelligent agent platform according to claim 1 is characterized in that: The knowledge model forms a knowledge graph by converting knowledge data into structured data.

4. The intelligent agent supervision method based on the enterprise-level intelligent agent platform according to claim 3 is characterized in that: The knowledge data includes a rule base, a professional domain knowledge base, a real-time dynamic base and a meta-knowledge base.

5. The intelligent agent supervision method based on the enterprise-level intelligent agent platform according to claim 3 is characterized in that: The supervision module presets hallucination threshold intervals for different AI agents, and adjusts the hallucination threshold through a dynamic reflection mechanism based on generated content, external variables, and the correlation of knowledge graphs under different AI agents, resource rationality, and user behavior perception; the hallucination threshold is used to control the balance between the factual accuracy and creative freedom of the generated content, and the larger the hallucination threshold, the greater the creative freedom.

6. The intelligent agent supervision method based on the enterprise-level intelligent agent platform according to claim 5 is characterized in that: The evaluation module performs a multi-dimensional evaluation of the generated content in terms of legality, creativity and rationality, and ranks the generated content according to the reward mechanism.

7. The method for monitoring an intelligent agent based on an enterprise-level intelligent agent platform according to claim 6, characterized in that: The evaluation process of the evaluation module includes predicting conflicts before generating content, real-time monitoring and intercepting cross-border content during content generation, and reflecting and iterating the knowledge model after generating content.

8. The method for monitoring an intelligent agent based on an enterprise-level intelligent agent platform according to claim 6, characterized in that: The evaluation factors of the reward mechanism include legality, creativity, dissemination and commercial value, and the generated content is ranked according to the sum of the weights of all evaluation factors.

9. The method for monitoring an intelligent agent based on an enterprise-level intelligent agent platform according to claim 1, characterized in that: It also includes the steps of: the supervisory intelligent agent generates personalized guidance questions through real-time interactive data and pushes them to the associated intelligent agent cluster; and the intelligent agent cluster realizes knowledge sharing through cross-domain knowledge graph.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the intelligent agent supervision method according to any one of claims 1 to 9 is implemented.

Images