A trustee artificial intelligence research and development system

Abstract

This invention provides a trust-based artificial intelligence R&D system. To enable a large-scale AI model to identify user intents and invoke relevant actions through chat content, the system needs to transform actions into tools recognizable by the large-scale model. Based on the ModelContextProtocol framework, it defines the core communication protocol specifications for the McpServer application service. McpServer facilitates the transformation of application interaction actions into standardized tools for the AI ​​large-scale model, adapting to mainstream AI large-scale model tool invocation protocols. This solves the gap problem where AI large-scale models cannot directly invoke application actions, enabling collaborative interaction between the AI ​​large-scale model and the application. At the application layer, it provides multiple Agent intelligent assistants to identify different business scenarios based on prompt words. It also provides McpServer management and tool retrieval, using the SemanticKernel framework and large-scale model chat intent recognition and tool matching. Action-defined prompt words are used for large-scale model recognition. For complex and multi-intent scenarios, it provides multiple intent selection, records historical dialogue content, and ultimately identifies the user intent through multiple rounds of dialogue.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a trust-based artificial intelligence research and development system. Background Technology

[0002] Currently, most trust companies rely on manual operations on the webpage for interaction with their systems. For some common business operations, which are very frequent, and for some complex business scenarios, switching back and forth between various system menus is time-consuming and laborious. Furthermore, some of the company's relevant knowledge documents are passed around offline, lacking a unified management and maintenance model.

[0003] With the rapid development of artificial intelligence technology, various applications are widely used in people's daily lives and work. At the same time, the popularization of AI big data models has promoted the upgrading of intelligent interaction. AI big data models can directly call business functions and manage knowledge base documents, thereby improving work efficiency.

[0004] While large AI models possess natural language understanding capabilities, they struggle to directly recognize and invoke native application interactions. The core issue lies in the lack of standardized tool definitions for application interactions and the absence of dedicated intermediate services to convert actions into AI large model tool formats.

[0005] Therefore, it is necessary to provide a new AI research and development system for trust to solve the above problems. Summary of the Invention

[0006] The technical problem solved by this invention is to provide a trust-based artificial intelligence R&D system that addresses the gap problem of large AI models being unable to directly invoke application actions, and enables collaborative interaction between large AI models and applications.

[0007] To address the aforementioned technical problems, the trust artificial intelligence R&D system provided by this invention comprises the following four main modules:

[0008] McpServer service module: A custom McpServer service that converts Action actions into standardized tools for large AI models and retrieves a list of tools;

[0009] The SemanticKernel framework and multi-Agent module combine agents to form multiple agents based on business scenarios. The agents are routed to the corresponding agents through the A2A model and prompt words. These include: intelligent assistant, knowledge base and intelligent questioning.

[0010] The large-scale model intent recognition and vector matching module pre-writes the content that needs intent enhancement into the vector library through page configuration and matches it with the document vector.

[0011] The normalization module includes the following steps: S1, cleaning special characters: ① Remove punctuation and special characters at the end of the text; ② Clean up redundant spaces in the text: merge consecutive spaces into a single space and remove leading and trailing spaces;

[0012] S2, Standardized Case: ① Convert to lowercase;

[0013] S3. Replace proper noun information: ① Normalize and replace unique proper nouns with noun category names; ② Specific product or project;

[0014] S4. Replace time information: ① Replace exact time point with "time"; ② Replace exact time range with "time range";

[0015] S5. Replace synonyms: ① Preset synonym list;

[0016] S6. Remove modal particles: ① Preset modal particle list;

[0017] S7, Supplementing Regular Expression Replacement:

[0018] ① Preset regular expression replacement table

[0019] Preferably, the intelligent assistant adopts the Tongyi Qianwen big model, allowing users to chat and interact with Liangbao. The intelligent assistant obtains user needs and matches them based on chat content and prompts, hitting key intentions, loading parameters, and finally calling the corresponding business action.

[0020] Preferably, the knowledge base used is the Tencent Lexiang Knowledge Base, which encapsulates the Tencent Lexiang interface and identifies documents to output them in a streaming manner.

[0021] Preferably, the intelligent query retrieves data from the data platform tool via McpServer, the Agent calls the large model to identify the intent, and the results are output using an integrated third-party McpServer.

[0022] Preferably, the document vector matching step includes: the intelligent assistant identifies the business document through the uploaded document and executes the corresponding business action. First, the document template is imported into the vector library. When uploading the document, the content in the document is extracted and matched with the vector library. The result with a similarity of more than 80% is returned. The user can choose the document type and execute the corresponding business action.

[0023] For documents with a similarity of less than 80%, a process orchestration is used to process them step by step. First, the documents are processed and parsed into a standard data dictionary format. Then, the enumeration values ​​in the documents are replaced. Finally, the parameters are loaded and the corresponding business action is called.

[0024] Preferably, the intent recognition enhancement includes: maintaining an intent recognition vector table through the interface, associating content templates with corresponding Actions, normalizing user dialogue content, matching it with the vector library, obtaining the specific Tool, loading parameters into the large model, and finally calling the Action to return data results to the front end.

[0025] Compared with related technologies, the trust artificial intelligence R&D system provided by this invention has the following beneficial effects:

[0026] This invention provides a trust-based artificial intelligence R&D system. To enable a large model to understand user intent and invoke relevant actions through chat content, the actions need to be transformed into tools that the large model can recognize. Based on the ModelContextProtocol framework, the system defines the core communication protocol specifications of the McpServer application service. Through McpServer, the system realizes the transformation of application interaction actions into standardized tools of the AI ​​large model, adapts to the tool invocation protocols of mainstream AI large models, solves the gap problem that makes it difficult for AI large models to directly invoke application actions, and realizes collaborative interaction between AI large models and applications.

[0027] At the application layer, multiple Agent intelligent assistants are provided to identify different business scenarios based on prompt words. An McpServer management and retrieval tool is also provided, based on the SemanticKernel framework and large-scale model chat intent recognition and matching. Action-defined prompt words are used for large-scale model recognition. For complex and multi-intent scenarios, multiple intent selection is provided, historical dialogue content is recorded, and the user's intent is ultimately matched through multiple rounds of dialogue.

[0028] In terms of the interaction layer, the front end provides standardized output for different CUQR Actions, and also provides some personalized content display. The back end identifies the Tool through the large model, loads the parameters, calls it through Action, and encapsulates the results into the data structure required by the front end and returns it.

[0029] Intent Enhancement: Vector matching is performed before using a large model for intent recognition. The content requiring intent enhancement is pre-written into the vector library through page configuration, binding the content to the called Action. The similarity of the vector content is used to obtain the corresponding Action, which is then loaded with parameters and invoked. This pre-determined access to the desired Tool eliminates the intermediate large model recognition step, significantly increasing the intent recognition hit rate and response speed. The vector library supports multiple modes including InMemory, Elasticsearch, and QdrantVector, and is extensible.

[0030] Proper nouns: For some proprietary names related to trust business, such as name abbreviations of departments and employees, proper nouns are configured. Before the large model recognition, proper nouns will be replaced and then the vector library matching for intent enhancement will be performed to improve the intent hit rate.

[0031] Semantic normalization: The same business scenario can be described by multiple texts, which increases the efficiency of intent hit. Therefore, semantic normalization is performed for this scenario to improve the hit rate. Attached Figure Description

[0032] Figure 1 This is the core code of McpServer;

[0033] Figure 2 This is a list of McpServer services that have been used so far;

[0034] Figure 3 It is an overall functional division;

[0035] Figure 4 It showcases Grain Treasure and Knowledge Base;

[0036] Figure 5 It is the arrangement of document workflows;

[0037] Figure 6 It is text normalization processing;

[0038] Figure 7 It is an intent recognition enhancement processing flow;

[0039] Figure 8 It is the overall process of front-end and back-end interaction. Detailed Implementation

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0041] Please see Figures 1 to 8 A trust-based artificial intelligence R&D system is proposed to address the disconnect between large AI models and application actions, enabling collaborative interaction between large AI models and applications. The system comprises the following four modules:

[0042] McpServer service module:

[0043] 1. A custom McpServer service is implemented to convert Action actions into standardized tools for large AI models and retrieve the tool list. Currently, two McpServer services have been implemented: one for existing Action actions and the other for the data platform for intelligent data exchange, retrieving information about internal trust business. See the core code for details. Figure 1 As shown;

[0044] 1. In addition, it integrates other third-party McpServers, making our functionality more comprehensive (such as...). Figure 2 (as shown)

[0045] The SemanticKernel framework and multi-Agent integration module: Agents are divided into multiple groups based on business scenarios, and routing to the corresponding agent is done through an A2A model and prompt words. See details... Figure 3 As shown;

[0046] It includes:

[0047] 1. Smart Assistant: Liangbao uses the Tongyi Qianwen big model. Users can chat and interact with Liangbao, express their needs, and Liangbao will match the chat content and prompts to hit the key intent, load parameters, and finally call the corresponding business action.

[0048] 2. Knowledge Base: It uses the Tencent Lexiang Knowledge Base, encapsulating the Tencent Lexiang interface to recognize documents and output them in a streaming manner (e.g., ...). Figure 4 (as shown)

[0049] 3. Intelligent Data Query: The Agent pulls data from the data platform tool via McpServer, calls the large model to identify intent, and outputs results by integrating with the third-party McpServer.

[0050] The large-scale model's intent recognition and vector matching module pre-writes the content requiring intent enhancement into the vector library via page configuration and matches it with document vectors. Liangbao can identify the type of business document uploaded and execute corresponding business actions. First, the document template is imported into the vector library. When a document is uploaded, its content is retrieved and matched against the vector library. Results with a similarity of 80% or higher are returned. Users can then select the document type and execute the corresponding business action. For complex documents, a process orchestration approach is used for step-by-step processing. The large-scale model (qwen-doc-turbo) first processes and parses the document data into a standard data dictionary format, then replaces the enumerated values ​​in the document, and finally loads the parameters to call the corresponding business action. See details... Figure 5 .

[0051] The normalization module includes the following steps: 1. Cleaning special characters ① Remove punctuation and special characters at the end of the text; ② Clean up redundant spaces in the text: merge consecutive spaces into a single space and remove leading and trailing spaces;

[0052] 2. Standardize case sensitivity ① Convert to lowercase (handling cases that may contain English characters);

[0053] 3. Replace proprietary noun information (replace through the content list identified by the proprietary noun plugin)

[0054] ① Make normalized replacement for the unique proprietary noun and replace it with the noun classification name (excluding the "other" category); ② For the exact product or project (when the problem text clearly contains a project or product, transform it into a unique proprietary noun and process it according to ①);

[0055] 4. Replace time information (replace through the content list identified by the time plugin) ① Replace the exact time point with "time"; ② Replace the exact time range with "time range";

[0056] 5. Replace synonyms ① Preset the synonym table (e.g., replace "have a look" with "check");

[0057] 6. Remove modal particles ① Preset the modal particle table (e.g., remove "ah", "ya", "ma");

[0058] 7. Supplement regular replacement ① Preset the regular expression replacement table (e.g., replace "\d+ individuals" with "{quantity} individuals").

[0059] Maintain the intent recognition vector table through the interface, associate the content template with the corresponding Action. After the user's conversation content is normalized, perform vector library matching, obtain the specific Tool, call the large model for parameter loading, and finally call Action to return the data result to the front end. See the details in Figure 7 as shown;

[0060] Currently, the front end has adapted the intelligent assistants for the PC side and the mobile phone side. The back end identifies the intent through the large model and calls it to return the data format to the front end, and the front end outputs it personalized to the interface; see the details of the front-end and back-end interaction process in Figure 8 as shown.

[0061] To enable a large AI model to match user intent and invoke relevant business actions based on chat content, the actions need to be transformed into tools that the large model can recognize. Based on the ModelContextProtocol framework, the core communication protocol specification of the McpServer application service is defined. McpServer facilitates the transformation of application interaction actions into standardized tools for the AI ​​large model, adapting to mainstream AI large model tool invocation protocols. This addresses the disconnect between the AI ​​large model and the application, enabling collaborative interaction between the AI ​​large model and the application. At the application layer, multiple Agent intelligent assistants are provided to identify different business scenarios based on prompt words. McpServer also manages and retrieves tools, matching chat intent recognition and tools based on the SemanticKernel framework and the large model. Action prompt words are defined for large model recognition. For complex and multi-intent scenarios, multiple intent selection is provided, historical dialogue content is recorded, and the user intent is ultimately matched through multiple rounds of dialogue. At the interaction layer, the front-end provides standardized output formats for different CUQR actions and also offers some personalized content display. The back-end loads parameters after recognizing the tool through the large model, invokes it via Action, and encapsulates the results into the data structure required by the front-end before returning it.

[0062] Intent Enhancement: Vector matching is performed before using a large model for intent recognition. The content requiring intent enhancement is pre-written into the vector library through page configuration, binding the content to the called Action. The similarity of the vector content is used to obtain the corresponding Action, which is then loaded with parameters and invoked. This pre-determined access to the desired Tool eliminates the intermediate large model recognition step, significantly increasing the intent recognition hit rate and response speed. The vector library supports multiple modes including InMemory, Elasticsearch, and QdrantVector, and is extensible.

[0063] Proper nouns: For some proprietary names related to trust business, such as name abbreviations of departments and employees, proper nouns are configured. Before the large model recognition, proper nouns will be replaced and then the vector library matching for intent enhancement will be performed to improve the intent hit rate.

[0064] Semantic normalization: The same business scenario can be described by multiple texts, which increases the efficiency of intent hit. Therefore, semantic normalization is performed for this scenario to improve the hit rate.

[0065] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A trust-based artificial intelligence research and development system, the system comprising the following four main modules: McpServer service module: A custom McpServer service that converts Action actions into standardized tools for large AI models and retrieves a list of tools; The SemanticKernel framework and multi-Agent module combine agents to form multiple agents based on business scenarios. The agents are routed to the corresponding agents through the A2A model and prompt words. These include: intelligent assistant, knowledge base and intelligent questioning. The large-scale model intent recognition and vector matching module pre-writes the content that needs intent enhancement into the vector library through page configuration and matches it with the document vector. The normalization module includes the following steps: S1, cleaning special characters: ① Remove punctuation and special characters at the end of the text; ② Clean up redundant spaces in the text: merge consecutive spaces into a single space and remove leading and trailing spaces; S2, Standardized Case: ① Convert to lowercase; S3. Replace proper noun information: ① Normalize and replace unique proper nouns with noun category names; ② Specific product or project; S4. Replace time information: ① Replace exact time point with "time"; ② Replace exact time range with "time range"; S5. Replace synonyms: ① Preset synonym list; S6. Remove modal particles: ① Preset modal particle list; S7, Supplementing Regular Expression Replacement: ① Preset regular expression replacement table.

2. The trustee artificial intelligence research and development system according to claim 1, wherein, The intelligent assistant adopts the Tongyi Qianwen big model, allowing users to chat and interact with Liangbao. The intelligent assistant obtains user needs and matches them based on chat content and prompts, hitting key intentions, loading parameters, and finally calling the corresponding business action.

3. The trust artificial intelligence R&D system according to claim 1, characterized in that, The knowledge base used is Tencent Lexiang Knowledge Base, which encapsulates the Tencent Lexiang interface and identifies documents to output them in a streaming manner.

4. The trust artificial intelligence R&D system according to claim 1, characterized in that, The intelligent query system pulls data from the data platform tool via McpServer, and the Agent calls the large model to identify the intent, outputting the results by integrating the third-party McpServer.

5. The trust artificial intelligence R&D system according to claim 2, characterized in that, The document vector matching steps include: the intelligent assistant identifies the business document through the uploaded document and executes the corresponding business action. First, the document template is imported into the vector library. When uploading the document, the content in the document is extracted and matched with the vector library. The result with a similarity of more than 80% is returned. The user can choose the document type and execute the corresponding business action. For documents with a similarity of less than 80%, a process orchestration is used to process them step by step. First, the documents are processed and parsed into a standard data dictionary format. Then, the enumeration values ​​in the documents are replaced. Finally, the parameters are loaded and the corresponding business action is called.

6. The trust artificial intelligence R&D system according to claim 1, characterized in that, The enhanced intent recognition includes: maintaining an intent recognition vector table through the interface, associating content templates with corresponding Actions, normalizing user dialogue content, matching it with the vector library, obtaining the specific Tool, loading parameters into the large model, and finally calling the Action to return data results to the front end.

Images