Enterprise proprietary agent rapid construction method, program, platform and device based on large model fine tuning

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to an enterprise special agent rapid construction method based on large model fine tuning. The method comprises the following steps: S1, performing agent basic configuration; s2, outputting an enterprise special agent based on the agent basic configuration scheme and the general large model; s3, setting rule parameters of an agent response strategy, and generating an agent interaction strategy; s4, performing full-life-cycle management on enterprise knowledge entries to ensure the accuracy and traceability of agent answering; and S5, verifying the accuracy of the agent knowledge and continuously optimizing the agent knowledge. According to the invention, through low-code configuration, an enterprise can construct an agent without a technical team, and autonomous controllability is realized; a large model adaptation technology enables the intelligent agent to accurately understand proprietary knowledge, solves the problem of term misunderstanding, and achieves high-precision adaptation; the knowledge full-life-cycle management and rapid updating mechanism ensures the timeliness of the agent knowledge; customizable interaction rules enable the intelligent agent to adapt to multi-scene requirements, and the user experience is improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method, program, platform and device for rapidly constructing enterprise proprietary intelligent agents based on large model fine-tuning. Background Technology

[0002] As enterprises deepen their digital transformation, the demand for intelligent interaction tools is becoming increasingly urgent. In scenarios such as customer service, internal training, and business consulting, enterprises urgently need intelligent agents that can understand specialized terminology (such as "enterprise development plan" and "digital economy talent") and align with business logic to improve communication efficiency and the accuracy of knowledge transfer. However, traditional intelligent agent construction models face three core challenges: First, the technical threshold is high: enterprise-specific intelligent agents need to be adapted to specific knowledge systems. Traditional development relies on algorithm engineers to manually code, requiring professional skills such as natural language processing and model training. Most enterprises (especially small and medium-sized enterprises) find it difficult to complete this independently.

[0003] Secondly, the customization cycle is long: Although general large models (such as the GPT series) have strong language capabilities, they lack understanding of enterprise-specific terminology and business details, and direct use is prone to "answering the wrong question"; while full-scale customized development requires investment in the entire process from data annotation to model training, and the cycle usually lasts for several weeks or even months, making it difficult to respond quickly to business needs.

[0004] Third, there is a lack of knowledge security and controllability: Enterprise knowledge (such as technical parameters and internal processes) often involves confidential information. If relying on third-party intelligent agent platforms, there is a risk of data leakage. At the same time, the response style and knowledge scope of general large models are uncontrollable, making it difficult to meet the personalized needs of enterprises such as "rigor" and "compliance".

[0005] Against this backdrop, how to lower the threshold for building enterprise-specific intelligent agents and achieve "rapid customization, security and controllability, and precise adaptation" has become a pain point in the industry.

[0006] Currently, there are three main types of technical solutions for enterprises to build proprietary intelligent agents: Direct access to general-purpose models: Enterprises can directly use open-source or commercial general-purpose models (such as ChatGPT and Wenxin Yiyan) via API interfaces, inputting only enterprise knowledge through prompts. For example, in customer service scenarios, product manual content can be embedded in prompts, allowing the model to generate answers based on general capabilities. Its limitations are: general-purpose models have a vague understanding of enterprise-specific terminology (such as "flow cytometer model parameters"), easily generating incorrect associations; and the limited length of prompts cannot cover the vast amount of enterprise knowledge, leading to one-sided answers.

[0007] Traditional question-answering system development involves manually writing rules (such as keyword matching and regular expressions) or training small, specialized models, combined with an enterprise knowledge base. For example, a manufacturing company's after-sales system responds to customer inquiries using a rule base that maps questions to answers. Its disadvantages include: high rule maintenance costs (rules need to be manually updated for each new piece of knowledge); inability to handle complex semantics (such as synonyms and contextual relationships); and poor scalability, making it difficult to adapt to business changes.

[0008] Fully customized development: Enterprises hire algorithm teams to train custom models based on raw data. For example, financial institutions collect years of historical dialogue data for "compliance consulting" scenarios and train models from scratch. The problems with this approach are: high cost (the development cost of a single model can reach hundreds of thousands of yuan); long cycle (requiring 3-6 months); strong technical dependence, requiring enterprises to maintain algorithm teams for a long time, making it difficult to popularize.

[0009] Based on the above technical solutions, the core defects of the existing technology can be summarized as follows: The high technical threshold and insufficient enterprise autonomy: While general large model calls are simple, they cannot be adapted to enterprise knowledge; while customized development requires professional teams, which is difficult for small and medium-sized enterprises to afford. As a result, enterprises rely entirely on external technical parties for the functional adjustment and knowledge update of intelligent agents, resulting in a lag in response.

[0010] Poor model adaptability and insufficient understanding of proprietary knowledge: General large models have not been trained with enterprise data, and their semantic understanding of specific terms (such as "enterprise cultivation plan" and "equipment model parameters") is inaccurate, which can easily lead to misattribution (such as misinterpreting "digital economy talent" as "traditional economic field talent"); Traditional rule-based systems cannot handle the contextual relationship of terms (such as "the equipment" in "the power consumption of the equipment" refers to a specific model).

[0011] Weak knowledge security and controllability: When relying on third-party platforms, enterprises need to upload knowledge bases to external servers, which poses a risk of confidentiality leakage; at the same time, the response style of intelligent agents (such as whether they are rigorous) and knowledge boundaries (such as whether they are allowed to call external networks) are uncontrollable, which may generate content that does not conform to the enterprise's standards (such as exaggerating product functions).

[0012] Low iteration efficiency and difficulty in responding quickly to business changes: Enterprise knowledge (such as product parameters and business processes) is dynamically updated over time. In traditional solutions, each update of knowledge requires retraining the model or modifying the rules, which takes several days and cannot meet the needs of "real-time iteration" (such as quickly updating the agent's product knowledge when a new product is launched).

[0013] In summary, existing solutions fail to address the core requirements of "low barrier to entry, rapid iteration, and high adaptability," sacrificing either accuracy (for general models) or efficiency (for custom development), thus failing to balance technical feasibility and business practicality. This application's rapid build platform based on large model fine-tuning provides a technical path to solve these problems through a "low-code configuration + automated fine-tuning" model. Summary of the Invention

[0014] Explanation of terms used in this invention: 1. Enterprise-specific intelligent agents: These are artificial intelligence interactive systems built on the specific knowledge, terminology, and business logic of an enterprise. They can provide targeted question-and-answer, consultation, and process assistance, such as "research institute talent training question-and-answer intelligent agents" and "product after-sales intelligent customer service".

[0015] 2. Large Model Fine-tuning: This refers to the technical process of incrementally training a general pre-trained language model (such as DeepSeek, LLaMA, etc.) using enterprise-specific data (such as manuals, process documents, and historical dialogues) to adapt the model to specific enterprise terminology and business scenarios.

[0016] 3. Low-code configuration: This refers to a development model where non-technical personnel can complete the parameter setting, function customization, and process configuration of intelligent agents through a visual interface (without writing code), thus lowering the technical threshold.

[0017] 4. Incremental training: A model optimization method that updates parameters only for enterprise-specific data while retaining the original capabilities of the general large model, thus avoiding the model "forgetting" general knowledge.

[0018] 5. Enterprise Knowledge Base: A structured database storing internal enterprise knowledge, including manuals, specifications, cases, glossaries, etc., such as "Product User Manual" and "Product Technical Parameter Table".

[0019] 6. Customized Interaction Rules: This refers to the function of personalizing the response style (rigorous / flexible), response length (concise / detailed), and knowledge retrieval scope (enterprise library only / enterprise library + general knowledge) of the intelligent agent.

[0020] 7. Knowledge tracing: This refers to the ability of an intelligent agent to locate the original document (such as a chapter or page number in a manual) from which the answer originates when answering a question, and to support jumping to view the relevant information to ensure that the information is verifiable.

[0021] 8. RAG technology: Retrieval-augmented Generation, which refers to the process by which an agent, when generating a response, first retrieves relevant information from the enterprise knowledge base and then combines the retrieval results to generate an accurate response, thereby improving the accuracy of the answer.

[0022] The technical problem to be solved by this invention is to design a method, program, platform, and device for rapidly constructing enterprise-specific intelligent agents based on large-model fine-tuning. Specific objectives include: Lowering the technical threshold and enabling autonomous construction: Through a low-code visual interface, non-technical personnel (such as enterprise administrators) can independently complete the configuration and deployment of intelligent agents without relying on algorithm engineers, thus solving the problem of "technical dependence".

[0023] Improve model adaptability and accurately understand proprietary knowledge: By using incremental fine-tuning technology, enterprise data is integrated into the general large model, enabling the agent to accurately understand specific terminology and business logic, thus solving the problem of "answering the wrong question".

[0024] Ensuring knowledge security and controllability: Supports local deployment of enterprise knowledge bases, with intelligent agents only accessing knowledge within the authorized scope; it also allows for customization of response styles and knowledge boundaries to ensure that outputs conform to enterprise standards.

[0025] Enables rapid iteration and adapts to dynamic business changes: Supports real-time updates of the knowledge base and incremental updates of the model, enabling intelligent agents to complete knowledge synchronization within hours and solving the problem of "iteration lag".

[0026] To address the aforementioned technical problems, this invention provides a method for rapidly constructing enterprise-specific intelligent agents based on large-model fine-tuning, comprising the following steps: Step S1: Perform basic configuration of the agent, including setting basic agent information and importing and extracting knowledge points from the enterprise knowledge base.

[0027] Step S2: Output enterprise-specific intelligent agents based on the basic configuration scheme and general large model of intelligent agents.

[0028] Step S3: Set the rule parameters of the agent's response strategy and generate the agent's interaction strategy, which includes style, knowledge scope and recommendation rules.

[0029] Step S4: Manage the entire lifecycle of enterprise knowledge items to ensure the accuracy and traceability of agent responses.

[0030] Step S5: Verify and continuously optimize the accuracy of the agent's knowledge, and report the test results back to Step S1 for basic configuration, which is used to improve knowledge items or adjust rules.

[0031] Furthermore, in step S1, the basic information settings for the intelligent agent include: inputting a name (such as "Research Institute Talent Training Q&A"), role positioning (such as "Knowledge Advisor" or "Business Assistant"), and application scenarios (such as internal consultation or employee training). The system automatically generates a configuration template.

[0032] Furthermore, in step S1, the enterprise knowledge base import and knowledge point extraction includes: uploading enterprise documents (such as the "Intelligent Research Institute Manual" and "Product Technical Specifications"), automatically extracting structured knowledge points through natural language processing technology, including core terms (such as "Enterprise Cultivation Plan" and "Digital Economy Talent"), business logic (such as "Talent Application Process"), and relationships (such as "Connection Rules between Enterprise Cultivation Plan and Digital Economy Talent Training"), and forming an editable knowledge entry base.

[0033] Furthermore, step S2 specifically includes the following process: P1: Knowledge Mapping and Integration: Associating structured knowledge items with the semantic space of a general model, enabling the model to understand the specific meaning of enterprise-specific terms (such as "Enterprise Incubation Program" specifically referring to the talent development project of the research institute) and business logic (such as "Application conditions require a work experience of ≥2 years").

[0034] P2: Terminology Boundary Definition: For easily confused corporate terms (such as "digital economy talent" and "traditional industry talent"), semantic annotation is used to clarify the differences and ensure that the model is unambiguous in its answers (such as "the enterprise development program only covers talent in the digital economy field").

[0035] P3: Knowledge Adaptation Validation: Randomly select knowledge items for sampling validation (e.g., "Query 'Start time of the Enterprise Development Plan'"). If the model's answer does not match the knowledge item, it will automatically provide feedback on the knowledge association points that need to be optimized (e.g., "Strengthen the binding relationship between 'Start in 2024' and 'Enterprise Development Plan'").

[0036] Furthermore, step S3 specifically includes the following process: P1: Response Style Settings: Provides style options of "rigorous and standardized", "concise and easy to understand" and "professional depth" for different agents. For example, the "technical manual agent" should select "professional depth" to prioritize the use of industry terminology; the "newcomer training agent" should select "concise and easy to understand" to explain the terminology with examples.

[0037] P2: Knowledge access scope limitation: Set the knowledge boundaries that the agent can access, such as "only use the content of the 2024 version of the internal manual", and freely select the knowledge base that needs to be associated.

[0038] P3: Related Knowledge Recommendation: Define the rules for recommending related content after an answer. For example, after answering "Enterprise Incubation Program Application Process", related knowledge items such as "Required Materials List" and "Frequently Asked Questions" will be automatically recommended to form a knowledge chain for guidance.

[0039] Furthermore, step S4 specifically includes the following process: P1: Knowledge Entry Maintenance: Supports administrators to edit, add, and abolish structured knowledge entries (e.g., "2024 version of the Enterprise Development Plan replaces the 2023 version"). The system automatically marks the validity of knowledge (e.g., "currently valid" or "expired"). The agent only calls valid knowledge.

[0040] P2: Knowledge traceability mechanism: When the agent answers, it automatically marks the source of the knowledge entry corresponding to the answer (such as "from Section 3.1 'Core Content of Enterprise Development Plan' in the Research Institute Manual"), and supports jumping to the original knowledge entry or document page, so that users can verify the information.

[0041] P3: Dynamic Update Synchronization: When a knowledge item is updated (such as "Adjustment of the deadline for application for the Enterprise Development Program"), it is automatically synchronized to the enterprise-specific intelligent agent in step S2, triggering the intelligent agent to relearn the knowledge association and complete the intelligent agent knowledge update.

[0042] Furthermore, step S5 specifically includes the following process: P1: Knowledge Accuracy Test: The administrator selects key knowledge items (such as "core objectives of the enterprise development plan" and "definition of digital economy talents"), manually enters relevant questions to test and verify the matching degree between the agent's answers and the knowledge items, such as "complete match", "partial match", and "no match".

[0043] P2: Interaction Scenario Simulation: Simulate typical questions in actual business scenarios (such as "Do I meet the application requirements for the Enterprise Incubation Program?"), observe whether the agent follows the interaction rules (such as style and related recommendations), and record the knowledge call path (such as "calling knowledge entry ID: YQJH-005") to locate the cause of deviation.

[0044] This invention also provides a rapid construction platform for enterprise proprietary intelligent agents based on large model fine-tuning. This platform executes the aforementioned rapid construction method for enterprise proprietary intelligent agents based on large model fine-tuning, and specifically includes the following modules: Low-code configuration module: Enterprise administrators can complete the basic configuration of the agent and generate knowledge points through a visual interface, generate an agent configuration scheme, which includes basic attributes and a structured knowledge item library, and push it to the large model adaptation module.

[0045] Large Model Adaptation Module: Based on the knowledge entries generated by the low-code configuration module, it achieves accurate adaptation between general large models and enterprise-specific knowledge, generates enterprise-specific intelligent agents, and pushes them to the interaction rule customization module.

[0046] Customizable Interaction Rules Module: This module allows administrators to customize the response strategies of intelligent agents, ensuring that the output aligns with the needs of enterprise scenarios. It generates intelligent agent interaction strategies, which include style, knowledge scope, and recommendation rules, and pushes them to the knowledge management and traceability module.

[0047] Knowledge Management and Traceability Module: Used for the full lifecycle management of enterprise knowledge items, ensuring the accuracy and traceability of intelligent agent responses, and pushing them to the user interaction interface.

[0048] Testing and Optimization Module: Provides tools to verify the accuracy of agent knowledge and continuously optimize it, generates test result reports and pushes them to the low-code configuration module for knowledge entry improvement or rule adjustment.

[0049] This invention also provides a device for rapidly constructing enterprise-specific intelligent agents based on large model fine-tuning, comprising: At least one processor; and At least one memory communicatively connected to the processor; The memory stores instructions that can be executed by a processor, which in turn execute the aforementioned method for rapidly constructing enterprise-specific intelligent agents based on large model fine-tuning.

[0050] The present invention also provides a computer program product, including computer instructions that, when executed by a processor, cause a computer device to perform the aforementioned method for rapidly constructing enterprise proprietary intelligent agents based on large model fine-tuning.

[0051] This invention relates to natural language processing, large-scale model fine-tuning, low-code development, and enterprise knowledge management technologies, applicable to scenarios such as intelligent customer service, internal knowledge Q&A, and business process assistance. This invention builds a technical system around "rapid construction of enterprise-specific intelligent agents," achieving full-process automation "from general-purpose large models to enterprise-specific intelligent agents" through the collaborative efforts of five modules. This covers a closed loop of "configuration-fine-tuning-deployment-optimization," with seamless integration between modules through data flow, supporting enterprises in autonomously and efficiently building intelligent agents. Specifically, this invention includes the following beneficial effects: 1. Significantly improved build efficiency The development cycle for enterprise-specific intelligent agents has been shortened from the traditional 2-4 weeks to 1-2 days (including document upload, knowledge extraction, and model adaptation), reducing labor costs by 60% (no algorithm engineer involvement required). For example, a research institute built a "talent training question-answering intelligent agent" through the platform, completing the entire process from uploading the manual to official launch in just one day, saving 100,000 yuan compared to outsourcing development.

[0052] 2. The accuracy of proprietary knowledge comprehension has been greatly improved. Actual tests show that the intelligent agent has an accuracy rate of 92% in understanding enterprise-specific terms (such as "enterprise development plan" and "digital economy talent"), which is 53% higher than the general large model (60%). The matching degree of the answers to business logic (such as application conditions and training process) is 95%, which is 27% higher than the traditional rule system (75%).

[0053] 3. Optimize knowledge management and update efficiency After knowledge entries are updated, the agent can complete the synchronization within 24 hours, which is 168 times more efficient than the traditional solution (7 days). For example, after a company updates its "product technical parameters", the administrator only needs to upload the new document, and the agent can accurately answer related questions on the same day, increasing the accuracy of customer inquiry response from 80% to 98%.

[0054] 4. Enhanced adaptability to multiple scenarios The same intelligent agent can be configured to adapt to multiple scenarios through rules: when switching to the "internal consultation" scenario, it adopts the "rigorous and standardized" style and calls all knowledge; when switching to the "external publicity" scenario, it adopts the "simple and easy to understand" style and only calls public knowledge. The scenario switching time is ≤30 minutes, saving 50% of maintenance costs compared to the traditional multi-system deployment mode.

[0055] 5. Information credibility and compliance assurance All intelligent agent responses are labeled with knowledge sources (e.g., "from Section 3.2 of the Manual"), which users can trace and verify, increasing information credibility by 40%. At the same time, through knowledge scope control, the access rate of sensitive information (such as undisclosed technology) is 0, fully complying with enterprise compliance requirements. Attached Figure Description

[0056] The specific embodiments of the present invention will be further explained below with reference to the accompanying drawings.

[0057] Figure 1 This is a flowchart of the method for rapidly constructing enterprise-specific intelligent agents based on large model fine-tuning according to the present invention.

[0058] Figure 2 This is a system block diagram of the enterprise proprietary intelligent agent rapid construction platform based on large model fine-tuning according to the present invention.

[0059] Figure 3 This is a schematic diagram illustrating the operational process of the enterprise-specific intelligent agent rapid construction platform based on large model fine-tuning according to the present invention. Detailed Implementation Example 1

[0060] Combination Figure 1 The specific steps of the rapid construction method for enterprise proprietary intelligent agents based on large model fine-tuning in this embodiment are as follows: Step S1: Perform basic agent configuration, including setting basic agent information and importing and extracting knowledge points from the enterprise knowledge base. This allows enterprise administrators to complete basic agent configuration and knowledge point generation through a visual interface without writing code.

[0061] In this preferred embodiment, step S1 includes setting the basic information of the intelligent agent, such as: inputting the name (e.g., "Research Institute Talent Training Q&A"), role positioning (e.g., "Knowledge Consultant", "Business Assistant"), and application scenario (e.g., internal consultation, employee training), and the system automatically generates a configuration template.

[0062] In this preferred embodiment, step S1, the import and extraction of enterprise knowledge base includes: uploading enterprise documents (such as the "Intelligent Research Institute Manual" and "Product Technical Specifications"), automatically extracting structured knowledge points through natural language processing technology, including core terms (such as "Enterprise Cultivation Plan" and "Digital Economy Talent"), business logic (such as "Talent Application Process"), and relationships (such as "Connection Rules between Enterprise Cultivation Plan and Digital Economy Talent Training"), and forming an editable knowledge entry base.

[0063] Step S2: Output enterprise-specific intelligent agents based on the basic configuration scheme and general large model of intelligent agents.

[0064] In this preferred embodiment, step S2 specifically includes the following process: P1: Knowledge Mapping and Integration: Associating structured knowledge items with the semantic space of a general model, enabling the model to understand the specific meaning of enterprise-specific terms (such as "Enterprise Incubation Program" specifically referring to the talent development project of the research institute) and business logic (such as "Application conditions require a work experience of ≥2 years").

[0065] P2: Terminology Boundary Definition: For easily confused corporate terms (such as "digital economy talent" and "traditional industry talent"), semantic annotation is used to clarify the differences and ensure that the model is unambiguous in its answers (such as "the enterprise development program only covers talent in the digital economy field").

[0066] P3: Knowledge Adaptation Validation: Randomly select knowledge items for sampling validation (e.g., "Query 'Start time of the Enterprise Development Plan'"). If the model's answer does not match the knowledge item, it will automatically provide feedback on the knowledge association points that need to be optimized (e.g., "Strengthen the binding relationship between 'Start in 2024' and 'Enterprise Development Plan'").

[0067] Specifically, in this embodiment, an agent based on Retrieval Enhancement Generation (RAG) can be constructed without relying on large model adaptation. Instead, enterprise knowledge entries are stored in a retrieval database. After receiving a question, the agent first retrieves relevant knowledge from the database, and then a general-purpose model generates an answer based on the retrieval results. For example, when a user asks about the "application conditions for the Enterprise Development Program," the system retrieves the corresponding knowledge entry and the model organizes it into a natural language response. By supplementing enterprise knowledge through retrieval, no model training is required, making it suitable for scenarios with extremely frequent knowledge updates (such as daily updates), while still achieving accurate answers based on proprietary knowledge. Existing RAG frameworks (such as LangChain) can support this solution; the technology is mature, and those skilled in the art can implement it by configuring the retrieval engine. However, their ability to understand complex semantic relationships is weaker than the adaptation mechanism of this invention.

[0068] Specifically, in this embodiment, the intelligent agent can also be configured by manually writing rules. Instead of using low-code automatic extraction, the administrator manually writes knowledge rules (e.g., "If a user asks about the 'Enterprise Development Plan,' return the definition + application conditions"), which directly associate with a general model to generate an answer. This manual rule-based knowledge association is suitable for enterprises with very few knowledge items (≤100), yet still achieves the goal of intelligent agent customization. Rule writing is a conventional technology (e.g., based on Python conditional statements), but the maintenance cost is high (rules need to be written manually for each new knowledge item), making it unsuitable for scenarios with a large amount of knowledge.

[0069] Specifically, in this embodiment, a fully trained enterprise-specific model can also be used. Instead of adapting to a general-purpose model, the entire process from data annotation to model training is customized, using enterprise knowledge to train a dedicated model from scratch. For example, collecting question-and-answer data related to the "Enterprise Development Program" can train a small neural network model. Full-scale training achieves deep adaptation, suitable for enterprises with massive amounts of data (millions of question-and-answer pairs) and sufficient technical reserves. While full-scale training technology is mature, it requires a large amount of labeled data (≥100,000 records) and computational power (multi-GPU clusters), taking 2-4 weeks and costing 5-10 times more than this invention.

[0070] All of the above solutions can achieve the construction of enterprise-specific intelligent agents, but the "low-code configuration + large model adaptation" mode in this embodiment is superior in terms of the balance between "efficiency, cost, and adaptability": RAG technology relies on retrieval accuracy and performs poorly in complex scenarios; manual rules have weak extensibility; and full-scale training is costly. Those skilled in the art can choose a suitable solution based on the enterprise's knowledge and technical reserves.

[0071] Step S3: Set the rule parameters of the agent's response strategy and generate the agent's interaction strategy, which includes style, knowledge scope and recommendation rules.

[0072] In this preferred embodiment, step S3 specifically includes the following process: P1: Response Style Settings: Provides style options of "rigorous and standardized", "concise and easy to understand" and "professional depth" for different agents. For example, the "technical manual agent" should select "professional depth" to prioritize the use of industry terminology; the "newcomer training agent" should select "concise and easy to understand" to explain the terminology with examples.

[0073] P2: Knowledge access scope limitation: Set the knowledge boundaries that the agent can access, such as "only use the content of the 2024 version of the internal manual", and freely select the knowledge base that needs to be associated.

[0074] P3: Related Knowledge Recommendation: Define the rules for recommending related content after an answer. For example, after answering "Enterprise Incubation Program Application Process", related knowledge items such as "Required Materials List" and "Frequently Asked Questions" will be automatically recommended to form a knowledge chain for guidance.

[0075] Step S4: Manage the entire lifecycle of enterprise knowledge items to ensure the accuracy and traceability of agent responses.

[0076] In this preferred embodiment, step S4 specifically includes the following process: P1: Knowledge Entry Maintenance: Supports administrators to edit, add, and abolish structured knowledge entries (e.g., "2024 version of the Enterprise Development Plan replaces the 2023 version"). The system automatically marks the validity of knowledge (e.g., "currently valid" or "expired"). The agent only calls valid knowledge.

[0077] P2: Knowledge traceability mechanism: When the agent answers, it automatically marks the source of the knowledge entry corresponding to the answer (such as "from Section 3.1 'Core Content of Enterprise Development Plan' in the Research Institute Manual"), and supports jumping to the original knowledge entry or document page, so that users can verify the information.

[0078] P3: Dynamic Update Synchronization: When a knowledge item is updated (e.g., "Adjustment to the deadline for application for the Enterprise Development Program"), it is automatically synchronized to the enterprise-specific intelligent agent in step S2, triggering the intelligent agent to relearn the knowledge association and complete the intelligent agent knowledge update. Specifically, in this embodiment, the time limit for the intelligent agent knowledge update to take effect is guaranteed to be within 24 hours.

[0079] Step S5: Verify and continuously optimize the accuracy of the agent's knowledge, and report the test results back to Step S1 for basic configuration, which is used to improve knowledge items or adjust rules.

[0080] In this preferred embodiment, step S5 specifically includes the following process: P1: Knowledge Accuracy Test: The administrator selects key knowledge items (such as "core objectives of the enterprise development plan" and "definition of digital economy talents"), manually enters relevant questions to test and verify the matching degree between the agent's answers and the knowledge items, such as "complete match", "partial match", and "no match".

[0081] P2: Interaction Scenario Simulation: Simulate typical questions in actual business scenarios (such as "Do I meet the application requirements for the Enterprise Incubation Program?"), observe whether the agent follows the interaction rules (such as style and related recommendations), and record the knowledge call path (such as "calling knowledge entry ID: YQJH-005") to locate the cause of deviation.

[0082] The rapid construction method of enterprise-specific intelligent agents based on large model fine-tuning in this embodiment is applied to the question-answering intelligent agent for talent training in research institutes. The process logic and interaction sequence are as follows: Configuration phase: The administrator enters the agent name "Research Institute Talent Cultivation Q&A" through the "Low-Code Configuration Module", selects the role "Knowledge Advisor" and the application scenario "Internal Consulting" → uploads the "Artificial Intelligence Research Institute Manual" → the system automatically extracts structured knowledge items (such as "Enterprise Cultivation Plan: Launched in 2024, focusing on the cultivation of digital economy talents, the application conditions are ≥2 years of work experience" and "Digital Economy Talents: covering fields such as artificial intelligence and big data, and need to pass 3 core course assessments") → outputs the "Agent Configuration Scheme" to the large model adaptation module.

[0083] Large model adaptation phase: The "Large Model Adaptation Module" receives the configuration scheme, associates the knowledge items with the semantic space of the general large model, defines the term boundaries (e.g., clarifies that "Enterprise Incubation Program" ≠ "Other Talent Programs", the core difference is "Focusing on the Digital Economy Field"), and performs sampling verification (randomly queries "the start time of the Enterprise Incubation Program", the model answers "2024", matches with the knowledge item, and outputs "Research Institute Proprietary Intelligent Agent").

[0084] Rule customization phase: In the "Interaction Rules Customization Module", the administrator sets the response style as "rigorous and standardized" (prioritizing the use of manual terminology), the knowledge retrieval scope as "only the 2024 version of the manual content", and the association recommendation rule as "recommend 3 related knowledge items after answering" → outputs "interaction strategy" (such as recommending "application materials list", "assessment criteria", and "frequently asked questions" after answering "application conditions for the enterprise development plan").

[0085] Deployment and testing phase: After the AI ​​agent went online, the administrator entered the question "What is the core objective of the Enterprise Cultivation Program?" through the "Testing and Optimization Module". The AI ​​agent responded: "The core objective of the Enterprise Cultivation Program is to cultivate compound talents in the digital economy field, covering artificial intelligence, big data and other directions (from Section 3.1 of the Research Institute Manual)", and recommended relevant knowledge. The test showed that the answer completely matched the knowledge item and the response style was in line with the settings. The AI ​​agent was officially put into use.

[0086] Update phase: Three months later, the research institute released a "Supplementary Explanation of the Enterprise Incubation Program" (adding "the second batch of application time in 2024") → the administrator uploaded the supplementary document to the "Low-code Configuration Module", the system automatically extracted the new knowledge item → the "Knowledge Management and Traceability Module" was updated synchronously and triggered the "Large Model Adaptation Module" to relearn → within 24 hours, the agent could accurately answer "the second batch of application time" and the response was marked with the new source "Section 1.2 of the Supplementary Explanation".

[0087] The low-code configuration and automatic structured knowledge extraction method in this embodiment enables the configuration of basic agent information through a visual interface, automatically extracting structured knowledge items (including terminology, logical relationships, and association rules) from enterprise documents without manual coding. This method transforms knowledge processing from "manual organization" to "automatic extraction," representing a core innovation that lowers the technical threshold. The adaptation mechanism between the general large model and enterprise knowledge in this embodiment associates enterprise knowledge items with the semantic space of the general large model. Through terminology boundary definition and knowledge mapping fusion, the model accurately understands proprietary terminology and business logic, avoiding ambiguity. This mechanism solves the core problem of "general models not understanding enterprise knowledge." The knowledge lifecycle management and traceability technology in this embodiment supports the full-process management of knowledge item addition, editing, and obsolescence, automatically marking validity; when the agent answers, the knowledge source (such as document chapters) is simultaneously marked, and jump verification is supported. This technology ensures the timeliness and credibility of knowledge, forming a "knowledge-answer-verification" closed loop. The scenario-based interaction rule customization framework in this embodiment can configure response style (rigorous / popular), knowledge call scope (public / internal), and association recommendation rules, and rule adjustments take effect in real time without retraining the model. This framework meets the personalized needs of enterprises in multiple scenarios and enhances the flexibility of interaction. Example 2

[0088] Combination Figure 2 and Figure 3 The enterprise proprietary intelligent agent rapid construction platform based on large model fine-tuning in this embodiment implements the enterprise proprietary intelligent agent rapid construction method based on large model fine-tuning in Embodiment 1, specifically including the following modules: Low-code configuration module: Enterprise administrators can complete the basic configuration of the agent and generate knowledge points through a visual interface, generate an agent configuration scheme, which includes basic attributes and a structured knowledge item library, and push it to the large model adaptation module.

[0089] In this embodiment, the low-code configuration module serves as the platform's "entry point," enabling enterprise administrators to complete basic agent configuration and knowledge point generation through a visual interface without writing code. Its core functions include: ① Intelligent Agent Basic Information Settings: Enter the name (e.g., "Research Institute Talent Training Q&A"), role positioning (e.g., "Knowledge Advisor", "Business Assistant"), and application scenario (e.g., internal consultation, employee training), and the system will automatically generate a configuration template.

[0090] ② Import and extraction of enterprise knowledge base: Supports uploading enterprise documents (such as "Intelligent Research Institute Manual" and "Product Technical Specifications"), and automatically extracts structured knowledge points through natural language processing technology, including core terms (such as "Enterprise Cultivation Plan" and "Digital Economy Talent"), business logic (such as "Talent Application Process"), and relationships (such as "Connection Rules between Enterprise Cultivation Plan and Digital Economy Talent Training"), and forms an editable knowledge entry base.

[0091] This module's data input and output: The input consists of basic information configured by the administrator and enterprise documents; the output is an agent configuration scheme, which includes basic attributes and a structured knowledge entry library, and is pushed to the large model adaptation module.

[0092] Large Model Adaptation Module: Based on the knowledge entries generated by the low-code configuration module, it achieves accurate adaptation between general large models and enterprise-specific knowledge, generates enterprise-specific intelligent agents, and pushes them to the interaction rule customization module.

[0093] In this embodiment, specifically, the large model adaptation module, as the "core engine" of the platform, achieves accurate adaptation between the general large model and enterprise proprietary knowledge based on the knowledge entries generated by the low-code configuration module. The core process includes: ① Knowledge Mapping and Integration: Linking structured knowledge items with the semantic space of a general model, enabling the model to understand the specific meaning of enterprise-specific terms (such as "Enterprise Incubation Program" specifically referring to the talent development project of the research institute) and business logic (such as "Application conditions require a work experience of ≥2 years").

[0094] ② Terminology Boundary Definition: For easily confused corporate terms (such as "digital economy talent" and "traditional industry talent"), semantic annotation is used to clarify the differences and ensure that the model is unambiguous in its answers (such as "the enterprise development program only covers talent in the digital economy field").

[0095] ③ Knowledge Adaptation Validation: Randomly select knowledge items for sampling validation (e.g., "Query 'Start time of the Enterprise Development Program'"). If the model's answer does not match the knowledge item, it will automatically provide feedback on the knowledge association points that need to be optimized (e.g., "Strengthen the binding relationship between 'Start in 2024' and 'Enterprise Development Program'").

[0096] This module's data input and output: The input consists of an intelligent agent configuration scheme and a general large model; the output consists of an enterprise-specific intelligent agent, which is a model instance adapted to enterprise knowledge and is pushed to the interaction rule customization module.

[0097] Customizable Interaction Rules Module: This module allows administrators to customize the response strategies of intelligent agents, ensuring that the output aligns with the needs of enterprise scenarios. It generates intelligent agent interaction strategies, which include style, knowledge scope, and recommendation rules, and pushes them to the knowledge management and traceability module.

[0098] In this embodiment, the interaction rule customization module specifically supports administrators in customizing the response strategy of the intelligent agent to ensure that the output meets the needs of the enterprise scenario. Its core functions include: ① Response style settings: Provides style options such as "rigorous and standardized", "concise and easy to understand" and "professional depth". For example, the "technical manual agent" can be set to "professional depth" to prioritize the use of industry terminology; the "newcomer training agent" can be set to "concise and easy to understand" to explain terminology with examples.

[0099] ② Knowledge access scope limitation: Set the knowledge boundaries that the agent can access, such as "only use the content of the 2024 version of the internal manual", and freely select the knowledge base that needs to be associated.

[0100] ③ Related knowledge recommendation: Define the rules for recommending related content after answering. For example, after answering "Enterprise Incubation Program Application Process", related knowledge items such as "Required Materials List" and "Frequently Asked Questions" will be automatically recommended to form a knowledge chain for guidance.

[0101] This module's data input and output: The input consists of rule parameters set by the administrator and enterprise-specific intelligent agents; the output is the intelligent agent interaction strategy, which includes style, knowledge scope, and recommendation rules, and is pushed to the knowledge management and traceability module.

[0102] Knowledge Management and Traceability Module: Used for the full lifecycle management of enterprise knowledge items, ensuring the accuracy and traceability of intelligent agent responses, and pushing them to the user interaction interface.

[0103] In this embodiment, the knowledge management and traceability module is specifically used to realize the full lifecycle management of enterprise knowledge items, ensuring the accuracy and traceability of the agent's answers. Its core functions include: ① Knowledge entry maintenance: Supports administrators to edit, add, and abolish structured knowledge entries, such as "2024 version of the Enterprise Development Plan replaces the 2023 version"; the system automatically marks the validity of knowledge, such as "currently valid" or "expired", and the agent only calls valid knowledge.

[0104] ② Knowledge traceability mechanism: When the intelligent agent answers, it automatically marks the source of the knowledge entry corresponding to the answer, such as "from Section 3.1 'Core Content of Enterprise Cultivation Plan' of the Research Institute Manual", and supports jumping to the original knowledge entry or document page, so that users can verify the information.

[0105] ③ Dynamic Update Synchronization: When a knowledge item is updated, such as "Adjustment of the Application Deadline for the Enterprise Incubation Program," the system automatically synchronizes it to the large model adaptation module, triggering the model to relearn the knowledge associations and ensuring that the intelligent agent's knowledge update takes effect. Specifically, in this embodiment, the intelligent agent's knowledge update takes effect within 24 hours.

[0106] This module's data input and output: The input consists of updated knowledge entries and user permission tags; the output consists of response content with traceability information and knowledge validity markers, which are pushed to the user interaction interface.

[0107] Testing and Optimization Module: Provides tools to verify the accuracy of agent knowledge and continuously optimize it, generates test result reports and pushes them to the low-code configuration module for knowledge entry improvement or rule adjustment.

[0108] Specifically, in this embodiment, the testing and optimization module provides tools to verify the accuracy of the agent's knowledge and continuously optimize it. Its core functions include: ① Knowledge accuracy test: Administrators select key knowledge items, such as "core objectives of the enterprise development plan" and "definition of digital economy talents", and manually input relevant questions to test and verify the matching degree between the agent's answers and the knowledge items, such as "complete match", "partial match" and "no match".

[0109] ② Interaction scenario simulation: Simulate typical questions in actual business scenarios (such as "Do I meet the application requirements for the Enterprise Incubation Program?"), observe whether the intelligent agent follows the interaction rules (such as style and related recommendations), and record the knowledge call path (such as "calling knowledge entry ID: YQJH-005") to locate the cause of deviation.

[0110] This module's data input and output: The input consists of questions manually asked by the administrator and simulated scenario questions; the output is a test result report, which includes matching degree and deviation analysis, and is pushed to the low-code configuration module for knowledge entry improvement or rule adjustment. Example 3

[0111] The rapid construction device for enterprise proprietary intelligent agents based on large model fine-tuning in this embodiment includes: At least one processor; and At least one memory communicatively connected to the processor; The memory stores instructions that can be executed by a processor, which are then executed by the processor to enable the device to perform the enterprise proprietary intelligent agent rapid construction method based on large model fine-tuning in Embodiment 1. Example 4

[0112] The computer program product of this embodiment includes computer instructions that, when run by a processor, cause a computer device to execute the enterprise proprietary intelligent agent rapid construction method based on large model fine-tuning in Embodiment 1.

[0113] Many specific details have been set forth in the foregoing description to provide a thorough understanding of the present invention. However, the above description is merely a preferred embodiment of the present invention, and the present invention can be implemented in many other ways different from those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed above. Furthermore, any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention, or modify them into equivalent embodiments, using the methods and techniques disclosed above, without departing from the scope of the present invention. Any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention, without departing from the content of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for quickly building enterprise-specific intelligent agents based on large model fine-tuning, characterized in that: Comprise the following steps: Step S1: Perform agent basic configuration, including agent basic information setting and enterprise knowledge base import and knowledge point extraction; Step S2: Output enterprise-specific agents based on agent basic configuration scheme and general large model; Step S3: Set rule parameters of agent response strategy, generate agent interaction strategy, the interaction strategy includes style, knowledge range and recommendation rule; Step S4: Whole life cycle management of enterprise knowledge items, ensure the accuracy and traceability of the agent's answer; Step S5: Verify and continuously optimize the accuracy of the agent's knowledge, and feed back the test results to step S1 for basic configuration, for knowledge item improvement or rule adjustment.

2. The large model fine-tuning-based enterprise-specific intelligent agent rapid construction method according to claim 1, characterized in that: In step S1, the agent basic information setting includes: input name, role positioning, application scenario, and the system automatically generates a configuration template.

3. The large model fine-tuning-based enterprise-specific intelligent agent rapid construction method of claim 1, wherein: In step S1, the enterprise knowledge base import and knowledge point extraction includes: uploading enterprise documents, automatically extracting structured knowledge points including core terms, business logic, and associated relationships through natural language processing technology, and forming an editable knowledge item library.

4. The large model fine-tuning-based enterprise-specific intelligent agent rapid construction method of claim 1, wherein: Step S2 specifically includes the following processes: P1: Knowledge mapping and fusion: associate structured knowledge items with the semantic space of the general large model, so that the model understands the specific meaning of enterprise-specific terms; P2: Term boundary definition: for easily confused enterprise terms, clearly distinguish the differences through semantic annotation to ensure that the model has no ambiguity when answering; P3: Knowledge adaptation verification: randomly extract knowledge items for sampling verification, if the model answer does not match the knowledge item, automatically feedback the knowledge association point that needs to be optimized.

5. The large model fine-tuning-based enterprise-specific intelligent agent rapid construction method according to claim 1, characterized in that: Step S3 specifically includes the following processes: P1: Response style setting: provide "strict and normative", "concise and popular" and "professional and in-depth" style options for setting different agents; P2: Knowledge calling range limitation: set the knowledge boundary that the agent can call; P3: Associated knowledge recommendation: define the associated content recommendation rules after answering to form a knowledge chain guide.

6. The large model fine-tuning-based enterprise-specific intelligent agent rapid construction method according to claim 1, characterized in that: Step S4 specifically includes the following processes: P1: Knowledge item maintenance: support administrators to edit, add, and abolish structured knowledge items, the system automatically marks the validity of the knowledge, and the agent only calls valid knowledge; P2: Knowledge traceability mechanism: when the agent answers, automatically mark the source of the knowledge item corresponding to the answer, and support jumping to the original knowledge item or document page for easy user verification of the information; P3: Dynamic update synchronization: when the knowledge item is updated, automatically synchronize to the enterprise-specific agent in step S2, trigger the agent to relearn the knowledge association, and complete the agent knowledge update and take effect.

7. The large model fine-tuning-based enterprise-specific intelligent agent rapid construction method according to claim 1, characterized in that: Step S5 specifically includes the following processes: P1: Knowledge accuracy test: administrators select key knowledge items and manually input related questions for testing to verify the matching degree of the agent's answer and the knowledge item; P2: Interaction scenario simulation: simulate typical questions in actual business scenarios to observe whether the agent follows the interaction rules, and record the knowledge calling path to locate the deviation reason.

8. A large model fine-tuning-based enterprise-specific intelligent agent rapid construction platform, characterized in that: The platform performs the large model fine-tuning-based enterprise-specific intelligent agent rapid construction method of any one of claims 1-7, specifically comprising the following modules: Low-code configuration module: enterprise administrators complete intelligent agent basic configuration and knowledge point generation through a visual interface, generate intelligent agent configuration schemes, and push them to the large model adaptation module; Large model adaptation module: based on the knowledge entries generated by the low-code configuration module, the module realizes precise adaptation of general large models and enterprise-specific knowledge, generates enterprise-specific intelligent agents, and pushes them to the interaction rule customization module; Interaction rule customization module: supports administrators to customize the response strategies of intelligent agents, ensures that the output meets the needs of enterprise scenarios, generates intelligent agent interaction strategies, and pushes them to the knowledge management and traceability module; Knowledge management and traceability module: used for the whole life cycle management of enterprise knowledge entries, ensures the accuracy and traceability of intelligent agent answers, and pushes them to the user interaction interface; Test and optimization module: provides tools to verify the accuracy of intelligent agent knowledge and continuously optimize it, generates test result reports, and pushes them to the low-code configuration module for knowledge entry improvement or rule adjustment.

9. An enterprise-specific intelligent agent rapid construction device based on large model fine-tuning, characterized in that: It comprises: at least one processor; and at least one memory connected to the processor in communication; wherein the memory stores instructions executable by the processor, and the instructions are executed by the processor to cause the device to perform the large model fine-tuning-based enterprise-specific intelligent agent rapid construction method of any one of claims 1-7.

10. A computer program product, characterised in that: It comprises computer instructions, which, when executed by a processor, cause a computer device to perform the large model fine-tuning-based enterprise-specific intelligent agent rapid construction method of any one of claims 1-7.

Images