An AI agent system and a collaborative knowledge construction method

Abstract

The application discloses an AI intelligent agent system and a collaborative knowledge construction method, which comprises the following steps: collecting communication data in real time and preprocessing; constructing an intelligent agent role library; analyzing a dialogue context and switching an intelligent agent role; designing an intelligent reasoning driven decision method, making an interactive mode decision and execution; based on field knowledge, enhancing the knowledge extraction and graph construction field knowledge base of RAG; knowledge synthesis and visual output; the AI intelligent agent system comprises a data acquisition module, a context analysis engine, a core processing unit, an interactive output module and a user feedback interface. The application can realize real-time analysis of communication, knowledge refinement, conflict identification and process assistance in the team collaborative knowledge construction process, and significantly improves the efficiency and quality of team collaborative knowledge construction.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and relates to natural language processing, large language model applications and agent technology. In particular, it relates to an AI agent system and a collaborative knowledge construction method based on AI agents, which can enable active participation in team collaboration, dynamic construction of knowledge graphs and assistance in decision-making. Background Technology

[0002] In team collaboration scenarios such as enterprise R&D, project management, and academic research, members engage in frequent and information-intensive exchanges through instant messaging tools and online meetings. These exchanges generate massive amounts of unstructured data, such as chat logs and transcripts of meeting audio. How to efficiently extract, organize, and solidify structured knowledge from these fragmented and dynamically evolving discussions is a major pain point in current collaborative workflows.

[0003] Existing collaborative office software and knowledge management platforms primarily offer information recording and sharing functions, but lack the ability to intelligently analyze and process the content of communication processes. The summarization, organization, and structuring of knowledge heavily rely on manual operation, which has the following limitations:

[0004] Information overload and knowledge loss: Key viewpoints, decision-making processes, and background information are scattered across a large number of conversations, making them difficult to trace and review, resulting in a significant loss of tacit knowledge.

[0005] Manual summarization is inefficient: Team members need to spend a lot of time organizing meeting minutes, summarizing discussions, and writing knowledge base entries. The workload is heavy and prone to omissions or subjective biases.

[0006] Lack of dynamic and in-depth insights: Existing technologies cannot analyze the focus of discussions in real time, identify potential conflicts of opinion or consensus, and cannot proactively guide and promote the team's knowledge building process.

[0007] Insufficient integration of domain knowledge makes it difficult to automatically link and verify new concepts and conclusions generated in discussions with existing knowledge bases within the organization (such as technical documents and project materials), thus creating knowledge silos.

[0008] Therefore, there is an urgent need for an intelligent solution that can be deeply integrated into collaborative processes, automatically and in real time analyze, summarize and structure team communication content, and proactively assist the team in knowledge construction. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, this invention provides an AI intelligent agent system and a collaborative knowledge construction method for collaborative knowledge construction, aiming to solve the technical problems of information overload, low efficiency of knowledge accumulation, lack of deep insight, and difficulty in domain knowledge integration in collaborative communication.

[0010] This invention constructs AI agents with multiple roles and modes, builds "contextual representations" based on multi-dimensional quantitative indicators and semantic features, and switches roles by combining user-configured parameters and large-scale model reasoning; it allows multiple agents to make independent decisions based on the same context; and it designs a three-level intelligent reasoning-driven "active / passive intervention decision-making mechanism" to achieve real-time analysis, knowledge extraction, conflict identification, and process assistance for team communication, significantly improving the efficiency and quality of team collaborative knowledge construction.

[0011] This invention provides a collaborative knowledge construction method based on AI agents, comprising the following steps:

[0012] 1) Real-time acquisition and preprocessing steps of communication data: Real-time acquisition of multimodal communication data such as text and speech transcription; by calling the large language model application programming interface (API), preprocessing operations such as speaker recognition, noise filtering, and format standardization are performed on the acquired data to form dialogue stream data in a unified format.

[0013] 2) Construct an agent role library; perform dialogue context analysis and agent role switching;

[0014] The constructed agent role library includes at least three roles: coordinator, creator, and improver; and defines the boundaries of role responsibilities. Dialogue context analysis involves agents with different roles continuously analyzing the context of the dialogue flow, the frequency of participant interactions, and the content topics. Based on preset thresholds, parameters, and / or the semantic understanding capabilities of a large language model (LLM), the current stage of the discussion (e.g., brainstorming, solution review, problem debate) is determined. Then, based on the context analysis data from the agent role library, the large model infers and selects which agent role to switch from based on the role responsibility boundaries.

[0015] In the process of collaborative knowledge construction, the core technologies of the dialogue context analysis and agent role switching method of this invention include:

[0016] 21) Conduct multi-dimensional in-depth analysis of the dialogue flow, and statistically analyze quantitative indicators such as discussion duration, user speaking activity, participation rate, and message length in the dialogue between the agent and the user; use a large language model to identify semantic features such as discussion stage, discussion depth, and focus issues, and then structure these quantitative indicators and semantic features according to a preset data pattern to form contextual representation data.

[0017] The pre-defined data schema is a data structure specification defined during the system initialization phase to describe contextual representation data. This data schema includes field sets, field types, hierarchical relationships, and encoding methods. For example, the system organizes quantitative indicators into a set of numerical fields, semantic features into enumerated or labeled field sets, and predefines their hierarchical and mapping relationships within the structured object, enabling subsequent inference modules to read contextual representation data in a unified format.

[0018] 22) Allow users to configure agent parameters, including thresholds for participation, speaking frequency, and speaking length. These parameter thresholds, contextual representation data, agent roles, and responsibility boundaries are then input into a large language model for reasoning and judgment. This determines the specific behavior the agent role should take in the current context, yielding the contextual analysis results for that agent role.

[0019] 23) A single group can be configured with multiple agents. Each agent makes independent behavioral decisions based on the context analysis results of the current dialogue (dialogue context). Functional isolation and collaborative work are achieved through predefined role and responsibility boundaries.

[0020] 3) Design intelligent reasoning-driven decision-making methods for interactive decision-making and execution:

[0021] Based on the current agent role and dialogue context, decisions are made using either an active or reactive agent decision-making interaction mode to achieve collaborative interaction.

[0022] In the proactive agent decision-making interaction mode, the agent proactively publishes content in communication channels (i.e., group chats between the agent and team members) based on preset trigger conditions, without being directly instructed, such as generating discussion summaries or highlighting differing viewpoints. In the passive mode, the agent responds to explicit instructions from users (such as "@agent, summarize the morning's discussion") and performs specific information processing or question-and-answer tasks.

[0023] Design a decision-making mechanism based on autonomous reasoning using a large language model, i.e., an intelligent reasoning-driven decision-making method; including:

[0024] 31) Construct a priority-based hierarchical decision-making framework, which includes the judgment of multiple execution decision priorities;

[0025] 32) Transform the situational analysis results, role positioning, dynamic threshold, recent dialogue content and other information into structured decision prompts, send them to the large language model for reasoning, and the model outputs the judgment result of whether or not to speak;

[0026] 4) Construct a domain knowledge base, and enhance RAG's knowledge extraction and graph construction based on domain knowledge;

[0027] The agent processes the dialogue flow and performs the following operations periodically or triggered by events: extracting key information; building a domain knowledge base and performing domain knowledge enhancement (RAG); and dynamically constructing a knowledge graph: integrating the extracted information entities and semantic relationships with the enhanced knowledge retrieved by RAG to build or update a dynamic knowledge graph.

[0028] 5) Knowledge synthesis and visualization output: The agent generates diverse outputs based on a dynamically updated knowledge graph and presents them in the collaborative platform.

[0029] Output formats include: structured summaries, knowledge graph visualizations, and speech quality analysis reports.

[0030] Accordingly, the present invention also provides an AI intelligent agent system for collaborative knowledge construction, the system comprising the following modules:

[0031] 1) Data acquisition module: used to connect to the collaborative platform API to acquire and preprocess communication data in real time.

[0032] 2) Context Analysis Engine: Built-in large language model for analyzing dialogue context and determining the role and interaction mode of the agent based on the analysis results.

[0033] 3) Core processing units: including: a large language model reasoning module, responsible for performing natural language understanding and generation tasks based on prompting engineering; a RAG knowledge enhancement module, which manages the vectorized indexing and retrieval of the domain knowledge base and provides external knowledge support for the language model; and a knowledge graph management module, responsible for the creation, updating, storage and querying of the knowledge graph.

[0034] 4) Interactive output module: Used to format the text summaries, analysis reports and visualization maps generated by the core processing unit and publish them to the user interface of the collaborative platform.

[0035] 5) User feedback interface: Used to receive user instructions to correct or confirm the content output by the agent, and to use the feedback to optimize the knowledge graph and model behavior.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1) Improve knowledge transformation efficiency: This invention extracts and organizes knowledge from unstructured communication automatically and in real time to construct a dynamic knowledge graph, which greatly shortens the cycle from discussion to knowledge accumulation.

[0038] 2) Enhance decision-making insight: By proactively identifying and presenting key points, consensus and conflict in discussions, this invention provides teams with deeper insights, helping them to focus on problems more efficiently, resolve differences and make informed decisions.

[0039] 3) Achieving dynamic and interconnected knowledge: By utilizing RAG technology, this invention dynamically connects new knowledge generated in real-time discussions with existing domain knowledge bases, promoting knowledge reuse and innovation.

[0040] 4) Optimize collaborative interaction experience: The multi-role and multi-mode design of the intelligent agent enables it to flexibly adapt to different collaborative scenarios like a real team member, intervene and provide support in an appropriate manner, and improve the smoothness and intelligence level of human-machine collaboration. Attached Figure Description

[0041] Figure 1 This is a functional module structure diagram of the AI ​​intelligent agent system proposed in this invention;

[0042] Figure 2 This is a flowchart of the AI ​​agent method for collaborative knowledge construction proposed in this invention.

[0043] Figure 3 A screenshot of the system interface for defining the role responsibilities and behavioral patterns in an AI agent system provided in a specific embodiment of the present invention. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the present invention.

[0045] This invention provides an AI agent system and method for collaborative knowledge construction. By constructing an AI agent with multiple roles and modes, it enables real-time analysis of team communication, knowledge extraction, conflict identification, and process assistance, significantly improving the efficiency and quality of collaborative knowledge construction. It addresses the technical problems of information overload, low efficiency of knowledge accumulation, lack of deep insight, and difficulty in integrating domain knowledge in collaborative communication.

[0046] This invention provides a collaborative knowledge construction method and system based on AI agents. The method includes the following steps:

[0047] 1) Real-time acquisition and preprocessing steps of communication data: Real-time acquisition of multimodal communication data such as text and speech transcription; by calling the large language model application programming interface (API), preprocessing operations such as speaker recognition, noise filtering, and format standardization are performed on the acquired data to form dialogue stream data in a unified format.

[0048] 2) Dialogue Context Analysis and Agent Role Switching: Construct an agent role library and define role responsibility boundaries. The agent continuously analyzes the context of the dialogue flow, the frequency of participant interactions, and the content topic. Based on preset thresholds, parameters, and / or the semantic understanding capabilities of the Large Language Model (LLM), and from the preset role library, according to the above context analysis data, the LLM infers and judges based on the role responsibility boundaries to select and switch agent roles.

[0049] Existing technologies in the field of collaborative knowledge construction generally use a single instruction or rule to switch roles, or do not switch roles at all; however, this invention designs a new control framework by using quantitative indicators, semantic indicators, configurable thresholds, and large model reasoning as decision-making criteria.

[0050] 20) Allows users to configure agent parameters, including group size, discussion duration, suggested speaking time per person, discussion phase, and language style. These parameters can be used to define the responsibility boundaries and behavioral patterns of agent roles. Example parameter configuration: [Phase Name] Initial Phase; [Maximum Phase Duration] 5 minutes; [Phase Description] Discussion begins, members familiarize themselves with the topic.

[0051] 21) Construct a library of agent roles and define the boundaries of role responsibilities (including behavioral patterns). Roles should include at least: Coordinator, Creator, and Perfector. The definitions of role responsibilities and behavioral patterns are as follows: Figure 3 As shown in the screenshot, the "Available Variable Reference" on the left side of the image represents the parameters configured in 20), and the system prompts on the right side define the "Coordinator" role's responsibility boundaries and behavioral patterns.

[0052] 22) Conduct multi-dimensional in-depth analysis of the dialogue flow, and statistically analyze quantitative indicators such as discussion duration (statistics on the cumulative duration of the current discussion stage and the time interval between messages to determine the discussion rhythm and whether there is a "stagnation" phenomenon), user speaking activity (calculated based on the number of times a message is spoken per unit time and the density of recent messages), participation rate (statistics on the number of members participating in the discussion and the distribution of the number of times a member speaks to identify whether there is an imbalance in participation), and message length (analysis on the average number of words in a single message and the trend of word count fluctuation to identify whether it is a shallow exchange of opinions or a deep discussion stage); use a large language model to identify semantic features such as discussion stage, discussion depth, and focus issues, and structure these quantitative indicators and semantic features to form contextual representation data.

[0053] 23) Input these situational representation data, parameter thresholds, agent roles and responsibility boundaries into the large language model for reasoning and judgment, and determine which agent role should be switched to and the specific behavior that agent role should take in the current situation.

[0054] 24) A single group can be configured with multiple agents. Each agent makes independent behavioral decisions based on the context analysis results of the current dialogue, and achieves functional isolation and collaborative work through predefined role and responsibility boundaries.

[0055] The constructed intelligent agent role library includes at least the following roles: coordinator, creator, and improver; and the boundaries of role responsibilities are defined.

[0056] In the process of collaborative knowledge construction, the AI ​​assistants in existing collaborative systems adopt a single, fixed role and cannot adjust their own positioning according to the dynamic development of the discussion. This results in inadequate support at different collaborative stages (such as the idea generation stage, solution convergence stage, and decision confirmation stage), and may even cause interference. The core key technological innovation of the dialogue context analysis and agent role switching method of this invention lies in: ① Performing multi-dimensional in-depth analysis of the dialogue flow, statistically analyzing quantitative indicators such as discussion duration, activity, participation rate, and message length in the dialogue between agents and users, and using a large language model to identify semantic features such as discussion stage, discussion depth, and focus issues. These quantitative indicators and semantic features are then structured to form contextual representation data. ② Allowing users to set group size and agent configuration parameters, including thresholds for agent participation, speaking frequency, and speaking length. These parameters and thresholds, along with contextual representation and role responsibility positioning, are comprehensively input into a large language model to infer whether the current situation of the role triggers the threshold and to decide the specific behavior to be taken at that moment. ③ The system supports configuring multiple agents in a single group. Each agent makes independent behavioral decisions based on the analysis results of the current common situation, and achieves functional isolation and collaborative work through predefined role and responsibility boundaries.

[0057] The role pool includes at least three roles: coordinator, creator, and improver, and the boundaries of their responsibilities are defined as follows:

[0058] ① Coordinator: Responsible for guiding the agenda and summarizing interim consensus; in the event of conflict, responsible for neutrally presenting the viewpoints of all parties. Specific actions include inviting members to speak in a gentle manner when participation is low; raising stimulating questions when discussions stall; summarizing key points to guide focus when discussions are disorganized; and managing processes at key time points.

[0059] ② Creator: Responsible for proactively raising thought-provoking questions or generating new ideas based on the current discussion and domain knowledge. Specific actions include: proposing innovative viewpoints when the discussion stalls in the early stages, raising in-depth questions when the discussion becomes superficial, and introducing challenging perspectives when opinions converge. This role is designed with a key constraint mechanism: avoiding introducing entirely new directions in the latter half of the discussion to prevent disrupting the established consensus.

[0060] ③ Improver: Responsible for fact-checking and logically organizing existing viewpoints, and supplementing details or finding supporting evidence. Specific actions include providing an integrated framework when multiple viewpoints coexist, supplementing background knowledge when discussing key concepts, politely correcting factual errors, and supplementing important aspects that have been overlooked when asked professional questions.

[0061] 3) Design an intelligent reasoning-driven interaction mode decision-making method: Set the trigger conditions for interaction mode decision-making; perform interaction mode decision-making and execution to obtain a judgment on whether the intelligent agent role intervenes (speaks):

[0062] Depending on the current role and the dialogue context, the agent decides whether to adopt an active or reactive interaction mode.

[0063] In proactive interaction mode, the agent, without direct user instruction, proactively publishes content in communication channels (i.e., group chats between the agent and team members) based on preset intervention criteria, such as generating discussion summaries and highlighting differing viewpoints. In passive mode, the agent responds to explicit user instructions (e.g., "@Agent, summarize the morning's discussion"), performing corresponding information processing or question-and-answer tasks.

[0064] Traditional collaborative AI uses fixed rules for triggering (such as scheduled summaries or keyword triggers), which leads to inappropriate timing of intervention, missing key discussion points, or frequently interrupting normal communication, seriously affecting the user experience.

[0065] This invention abandons traditional hard-coded triggering rules (such as fixed time intervals or simple counting thresholds) and designs and adopts a decision-making mechanism based on autonomous reasoning using a large language model. The system constructs a hierarchical decision-making framework, including three levels of priority judgment: the first priority identifies the user's direct request (triggered immediately when the user explicitly requests the agent to answer); the second priority identifies explicit rejection signals (not triggered when the user explicitly indicates that the agent does not need to participate); the third priority executes the intervention judgment of the active interaction mode, determining whether the large model should intervene in the discussion, which is the core decision-making logic.

[0066] In the third priority judgment, the system judges whether there is an opportunity to intervene and add value to the discussion through the following four criteria: discussion rhythm (judged based on quantitative indicators, such as whether the activity level is below the threshold, whether the silence time exceeds the threshold, whether the participation rate is low, etc.), content value (judged based on semantic understanding to see if there is an opportunity to supplement, extend, summarize, inspire, or connect viewpoints), role value (judged based on the role responsibilities of the agent to see if the current scenario is the one in which the role should play a role), and timing (judged based on the dialogue flow to see if a natural intervention point has been reached, such as when the viewpoint has been expressed, when the topic has changed, or when there is an opportunity to deepen the discussion, etc.).

[0067] Information such as contextual analysis results, role positioning, dynamic thresholds, and recent dialogue content are used as prompts for the large language model, which then performs inference. The model outputs a judgment on whether or not to speak based on the four criteria for judging opportunities to add value to the discussion. This mechanism transforms the agent's intervention decision from rule-driven to intelligent reasoning-driven, enabling it to adapt to complex and ever-changing collaborative scenarios.

[0068] Depending on the interaction mode (active vs. passive), the system employs a differentiated strategy for message generation. When passive mode is triggered, the system focuses on the specific questions raised by the user, provides targeted answers, and adjusts the generation strategy according to the configured answer style (direct answer or heuristic guidance). When active mode is triggered, the system emphasizes the agent's role, provides complete discussion context analysis, and requires that the generated content blend naturally into the discussion flow, avoiding any abruptness.

[0069] 4) Construct a domain knowledge base and enhance the knowledge extraction and graph construction steps of RAG based on domain knowledge.

[0070] The agent processes the dialogue stream, performing the following operations periodically or event-triggered:

[0071] Key information extraction: Utilizing large language models and task-specific prompt engineering, identify and extract core entities (such as technical terms and project codes), key viewpoints, to-do items, decision conclusions, and semantic relationships (such as causality, support, and opposition) from dialogues.

[0072] Domain Knowledge Enhancement (RAG): For extracted core entities or ambiguous viewpoints, a Retrieval-Augmented Generation (RAG) mechanism is initiated. The agent uses the entities mentioned in the dialogue as core query objects and searches through the vector index of a pre-built domain knowledge base (including internal technical documents, project databases, external authoritative literature, etc.) to obtain relevant background knowledge, definitions, or data, i.e., enhanced knowledge.

[0073] Dynamic knowledge graph construction: Extracted information entities are used as nodes, semantic relationships as edges, and enhanced knowledge retrieved from RAG is integrated to construct or update a dynamic knowledge graph. This graph structurally represents the knowledge system formed during team discussions in a machine-readable manner.

[0074] 5) Knowledge Synthesis and Visualization Output Steps: Based on a dynamically updated knowledge graph, the agent generates diverse outputs and presents them on the collaborative platform. Output formats include:

[0075] Structured Summary: The system periodically or upon user request automatically generates summary reports containing key discussion points, critical decisions, and unresolved issues. The generation mechanism involves the system calling a large language model, taking recent dialogue content (excluding agent-specific messages) as input, and requesting a 200-300 word overall summary and several key discussion points from the model. The system persistently stores the generated summary data in the group database for later querying and display.

[0076] Knowledge Graph Visualization: The system generates a hierarchical concept graph structure, visually presenting the concept network and knowledge structure formed during discussions. The generation mechanism is as follows: the system calls a large language model to analyze the dialogue content, requiring the model to output core topic nodes, main concept nodes, and their sub-concept nodes in a structured format. The system converts the hierarchical concept data returned by the model into a front-end renderable graph structure format, with each node containing a unique identifier, concept name, and a list of child nodes. The generated concept graph data is also persistently stored in the group database, supporting interactive browsing and exploration.

[0077] Speaking Quality Analysis: The system generates personalized speaking quality assessment reports for group members. The analysis mechanism is as follows: the system groups members by conversation message and evaluates each member's speaking content from multiple dimensions, including content depth, logic, constructiveness, participation, and expression ability. The system uses a large language model as a professional educational assessment assistant to analyze the first few speaking messages from each member, requiring the model to provide a quality score of one to five, a quality analysis text, a list of main strengths, and a list of improvement suggestions. For members who have spoken less than three times, the system provides a basic score and encouraging feedback. The quality analysis data of all members is aggregated and stored in the group database, providing teachers or administrators with objective teaching feedback.

[0078] Existing systems rely on template filling or simple rule extraction for knowledge output, resulting in mechanical content lacking semantic coherence and unable to handle model output anomalies, leading to poor system stability. This invention's knowledge synthesis and output mechanism replaces traditional template filling or rule extraction with a large language model-driven generation method, making the output content more semantically coherent and readable. Furthermore, it employs a persistent storage mechanism to preserve the generated knowledge assets as part of the group's data long-term, supporting continuous knowledge accumulation and reuse.

[0079] Accordingly, the present invention also provides an AI intelligent agent system for collaborative knowledge construction, the system comprising the following modules:

[0080] 1) Data acquisition module: used to connect to the collaborative platform API to acquire and preprocess communication data in real time.

[0081] 2) Context Analysis Engine: Built-in large language model for analyzing dialogue context and determining the role and interaction mode of the agent based on the analysis results.

[0082] 3) Core processing units: including: a large language model reasoning module, responsible for performing natural language understanding and generation tasks based on prompting engineering; a RAG knowledge enhancement module, which manages the vectorized indexing and retrieval of the domain knowledge base and provides external knowledge support for the language model; and a knowledge graph management module, responsible for the creation, updating, storage and querying of the knowledge graph.

[0083] 4) Interactive output module: Used to format the text summaries, analysis reports and visualization maps generated by the core processing unit and publish them to the user interface of the collaborative platform.

[0084] 5) User feedback interface: Used to receive user instructions to correct or confirm the content output by the agent, and to use the feedback to optimize the knowledge graph and model behavior.

[0085] The following example illustrates a software development team's need to discuss the design of a new feature, "intelligent recommendation based on user profiles," through an instant messaging tool, which involves collaborative knowledge construction.

[0086] Figure 1 The diagram shows the functional module structure of the AI ​​intelligent agent system proposed in this invention; Figure 2 This is the flowchart of the AI ​​agent method for collaborative knowledge construction proposed in this invention. The AI ​​agent system and method of this invention operate in the following manner:

[0087] Step S201: Data Acquisition and Preprocessing

[0088] The system's data acquisition module 10 connects to the team's discussion API to collect and format all conversation messages in real time.

[0089] Step S202: Context Analysis and Role Switching

[0090] The context analysis engine 20 determined that the current conversation was an open-ended discussion about technology selection, belonging to the "brainstorming" stage. Therefore, the engine 20 decision-making agent acted as a creator to stimulate more ideas.

[0091] Step S203: Interactive Mode Decision-Making and Execution

[0092] The agent then selects active mode. After monitoring that the discussion about the algorithm has lasted for 10 minutes, in order to broaden its thinking, the interactive output module 40 actively speaks in the channel: "In addition to collaborative filtering, can we consider deep learning-based sequence models, such as GRU or Transformer? They may be more advantageous in processing time-series data of user behavior."

[0093] Step S204: Knowledge Extraction and Graph Construction

[0094] The core processing unit 30 continues to work in the background.

[0095] The large language model reasoning module 31 extracts the entity "collaborative filtering", the attribute "simple to implement", and the problem "severe cold start problem" from member B's statement "collaborative filtering is simple to implement, but the cold start problem is serious".

[0096] When member C mentioned "we could refer to the practices in our 'Smart Retail' project last year," RAG Knowledge Enhancement Module 32 was triggered. It used "Smart Retail" as a keyword to search the internal project document library and found a technical report describing how the project used a hybrid recommendation algorithm to mitigate the cold start.

[0097] The knowledge graph management module 33 updates the knowledge graph immediately.

[0098] Step S205: Knowledge Synthesis and Output

[0099] After an hour of discussion, the context analysis engine 20 determined that the discussion was losing momentum, and the agent switched to the coordinator role and performed a proactive summary.

[0100] If the team leader @s the agent and asks a question at this time, the agent will enter passive mode, query the RAG knowledge enhancement module 32 again, and quickly provide an answer.

[0101] Through the above embodiments, this invention demonstrates a complete, closed-loop collaborative knowledge construction process. The intelligent agent not only records and summarizes, but also significantly improves the efficiency and output quality of collaborative work through its dynamic role-playing, proactive interaction, and deep integration with domain knowledge.

[0102] In summary, this invention designs a dialogue scenario analysis and agent role switching technology solution integrating a dynamic role switching mechanism (coordinator, creator, and improver); provides an interaction mode decision-making technology solution based on LLM reasoning that includes dual interaction modes (active and passive); and combines Retrieval Augmented Generation (RAG) technology with a large language model, applying it to real-time team collaborative communication scenarios to achieve seamless integration of external domain knowledge and dynamic dialogue information. It also provides a method for converting unstructured team communication into a dynamic knowledge graph in real time and automatically, realizing the structuring, visualization, and continuous evolution of knowledge.

[0103] It should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the scope of the claims.

Claims

1. A collaborative knowledge construction method based on AI intelligent agents, characterized in that, Includes the following steps: 1) Real-time acquisition and preprocessing of communication data: Real-time acquisition of multimodal communication data; preprocessing of the collected data by calling the large language model application programming interface (API) to form dialogue stream data in a unified format; 2) Construct an agent role library and define the boundaries of role responsibilities; The constructed agent role library includes at least the following roles: coordinator, creator, and improver; each agent role has a corresponding role responsibility boundary. During the collaborative knowledge construction process, the current stage of the discussion is determined, and the dialogue context is analyzed. Based on the context analysis data, the large language model infers and judges the role of the intelligent agent according to the boundaries of role responsibilities, and selects to switch roles. 21) Construct an agent role library and define role responsibility boundaries and behavior patterns; users define the responsibility boundaries and behavior patterns of agent roles by configuring agent parameters; Perform multi-dimensional in-depth analysis of the dialogue flow, statistically analyze the quantitative indicators in the dialogue between the agent and the user, identify semantic features using a large language model, and structure the quantitative indicators and semantic features to form contextual representation data. 22) Input the agent parameter threshold, situational representation data, agent role and role responsibility boundary into the large language model for reasoning and judgment, determine the agent role and the specific behavior that the agent role should take in the current situation, and obtain the situational analysis results of the agent role; 23) Configure multiple intelligent agents, each of which makes independent behavioral decisions based on the context analysis results of the current dialogue, and achieves functional isolation and collaborative work through their respective role and responsibility boundaries; 3) Design intelligent reasoning-driven decision-making methods for interactive decision-making and execution: Based on the current agent role and dialogue context, adopt different agent decision-making interaction modes to make decisions and achieve collaborative interaction; Intelligent reasoning-driven decision-making methods include: 31) Construct a priority-based hierarchical decision-making framework to determine the priority of multiple execution decisions; 32) The context analysis results, role positioning, dynamic threshold, and recent dialogue content information are used as structured decision prompts and sent to the large language model for reasoning; the model outputs a judgment result on whether or not to speak; 4) Construct a domain knowledge base, and perform knowledge extraction and graph construction based on domain knowledge enhancement RAG; the agent processes the dialogue flow and performs operations periodically or event-triggered, including: key information extraction; constructing a domain knowledge base and performing domain knowledge enhancement; constructing a dynamically updated knowledge graph; 5) Knowledge synthesis and visualization output: Based on dynamically updated knowledge graphs, the intelligent agent generates diverse outputs and presents them in the collaborative platform; the output formats include: structured summaries, knowledge graph visualizations, and speech quality analysis reports.

2. The collaborative knowledge construction method based on AI agents as described in claim 1, characterized in that, In step 1), the real-time acquired multimodal communication data includes text and / or speech transcription information; the preprocessing operations include speaker recognition, noise filtering, and format standardization.

3. The collaborative knowledge construction method based on AI agents as described in claim 1, characterized in that, Step 2) defines the boundaries of role responsibilities, specifically including: The facilitator's role is to guide the agenda and summarize the consensus reached at each stage; to neutrally present the viewpoints of all parties when conflicts arise; and to use gentle language to invite members to speak when participation is low; to raise stimulating questions when discussions stall; and to summarize key points and guide focus when discussions are disorganized, and to manage processes at key time points. The creator's role is responsible for proactively raising thought-provoking questions or generating new ideas based on the current discussion and domain knowledge. Creator behaviors include proposing innovative viewpoints when the discussion stalls in the early stages, raising in-depth questions when the discussion becomes superficial, and introducing challenging perspectives when opinions converge. The creator role has a key constraint mechanism, namely, avoiding introducing completely new directions in the latter half of the discussion to avoid disrupting the consensus that has been formed. The role of the improver is to verify the facts and organize the logic of existing viewpoints, and to supplement details or find supporting materials. The role of the improver includes providing an integrated framework when multiple viewpoints coexist, supplementing background knowledge when discussing key concepts, politely correcting factual errors, and supplementing important aspects that have been missed when asked professional questions.

4. The collaborative knowledge construction method based on AI agents as described in claim 3, characterized in that, In step 2), the user configures the agent parameters, including: group size, discussion duration, suggested speaking time per person, discussion stage, and language style parameters; agents with different roles continuously analyze the context of the dialogue flow, the frequency of interaction among participants, and the topic of the content; based on preset rules and / or the semantic understanding capabilities of the Large Language Model (LLM), the current discussion stage is determined, including the brainstorming stage, the solution review stage, and the problem debate stage.

5. The collaborative knowledge construction method based on AI agents as described in claim 4, characterized in that, Quantitative metrics in dialogues between agents and users include: discussion duration, user activity level, participation rate, and message length; semantic features are identified using large language models, including: discussion stage, discussion depth, and focus issues.

6. The collaborative knowledge construction method based on AI agents as described in claim 3, characterized in that, In step 3), the agent decision-making interaction mode includes the active agent decision-making interaction mode and the passive agent decision-making interaction mode; The proactive intelligent agent decision-making and interaction mode is a mode in which an intelligent agent proactively publishes content based on preset triggering conditions without being directly instructed. The passive agent decision-making interaction mode is when an agent responds to user commands and performs information processing or question-and-answer tasks.

7. The collaborative knowledge construction method based on AI agents as described in claim 6, characterized in that, Step 31) Construct a priority-based hierarchical decision framework to determine the priority of multiple execution decisions; the execution decision priority includes three levels; The first priority is to identify the user's direct request, and trigger it immediately when the user explicitly asks the agent to answer. The second priority identifies explicit rejection signals and does not trigger when the user explicitly states that they do not need the agent's participation. The third priority is to actively participate in the judgment; in the judgment of the third priority, an aggressive strategy is adopted, that is, to actively seek opportunities to intervene and add value to the discussion.

8. The collaborative knowledge construction method based on AI agents as described in claim 7, characterized in that, Opportunities to add value to discussions fall into four main categories: discussion rhythm, content value, role value, and opportune timing. The discussion of rhythm is based on quantitative indicators, including whether the activity level is below the threshold, whether the silence duration exceeds the threshold, and whether the participation rate is too low. Content value-added categories are based on semantic understanding to determine whether there are opportunities to supplement, expand, summarize, inspire, or connect viewpoints. The role value class is used to determine whether the current scenario is when the role should play a role, based on the role responsibilities of the intelligent agent. The "right time" category is based on judging whether a natural intervention point has been reached according to the flow of the dialogue, including when the viewpoint has been expressed, when the topic has been changed, and when there is an opportunity to deepen the discussion.

9. The collaborative knowledge construction method based on AI agents as described in claim 8, characterized in that, Step 5) The automatically generated output format includes: The structured summary generation mechanism is as follows: call the large language model, take the recent dialogue content as input, and ask the model to provide an overall summary and key discussion points of 200 to 300 words; extract structured information from the model response, including summary text and list of key points, and persistently store the generated summary data in the group database; Knowledge graph visualization generates a hierarchical concept graph structure, presenting the concept network and knowledge context formed during discussions in a visual manner. The knowledge graph generation mechanism is as follows: a large language model is called to analyze the dialogue content, requiring the model to output core topic nodes, main concept nodes, and their sub-concept nodes in a structured format; the hierarchical concept data returned by the model is converted into a front-end renderable graph structure format, with each node containing a unique identifier, concept name, and a list of child nodes; the generated concept graph data is persistently stored in the group database for interactive browsing and exploration. The speaking quality analysis report is a personalized speaking quality assessment report generated by group members. The mechanism for generating the analysis report is as follows: the group members communicate in groups, and each member's speaking content is evaluated from multiple dimensions, including content depth, logic, constructiveness, participation, and expression ability; a large language model is called to analyze the members' speaking, and the model is required to provide a quality score, quality analysis text, a list of main strengths, and a list of improvement suggestions; the quality analysis data of all members is summarized and stored in the group database.

10. An AI agent system implementing the collaborative knowledge construction method of claim 1, characterized in that, The system includes: a data acquisition module, a context analysis engine, a core processing unit, an interactive output module, and a user feedback interface; among which: The data acquisition module is used to connect to the collaborative platform API to acquire and preprocess communication data in real time; The context analysis engine is used to analyze the dialogue context through a built-in large language model and decide the role and interaction mode of the intelligent agent based on the analysis results; The core processing units include: a large language model reasoning module, a RAG knowledge enhancement module, and a knowledge graph management module. The large language model reasoning module is used to perform natural language understanding and generation tasks based on prompting engineering; the RAG knowledge enhancement module is used to manage the vectorized indexing and retrieval of the domain knowledge base, providing external knowledge support for the language model; and the knowledge graph management module is used for the creation, updating, storage, and querying of the knowledge graph. The interactive output module is used to format the text summaries, analysis reports, and visualization maps generated by the core processing unit and publish them to the user interface of the collaborative platform; The user feedback interface is used to receive user instructions to correct or confirm the content output by the agent, and to use this feedback to optimize the knowledge graph and model behavior.

Images