A method and system for multi-user knowledge interconnection and intelligent representative interaction

Abstract

The application relates to the technical field of knowledge social interaction, and discloses a method and system for multi-user knowledge interconnection and intelligent representative interaction. The method comprises the following steps: relying on a large language model and retrieval enhancement generation technology, analyzing user multi-modal content, extracting structured knowledge units and generating multi-dimensional knowledge representation, and constructing a personal knowledge portrait and storing it in a private knowledge base; constructing an intelligent representative, realizing cross-user knowledge accurate retrieval and permission control; dynamically calculating the knowledge credibility level of a user, and driving adaptive adjustment of permissions; converting social behavior into knowledge graph optimization signals based on reinforcement learning, constructing a knowledge interaction closed loop, and depositing high-value knowledge; and through a double-track recording mechanism, the knowledge value is verified and quantified, the Sharpe value algorithm is used to distribute the income, and a reliable settlement basis is generated. The application can solve the problems of loose binding of existing platform knowledge and identity, difficult guarantee of credibility, and inefficient interaction.

Description

Technical Field

[0001] This application relates to the technical field of knowledge-based social interaction, and in particular to a method and system for multi-user knowledge interconnection and intelligent representative interaction. Background Technology

[0002] In current knowledge-sharing platforms, the binding between an individual's core identity and multimodal knowledge resources is not tight enough, knowledge credibility lacks effective verification methods, and multimodal content parsing efficiency is low, with poor results in knowledge unit association and deduplication. When interconnecting knowledge across users, existing systems lack dynamic quantification of knowledge credibility, have rigid access control strategies, making it difficult to achieve differentiated sharing. At the same time, the closed-loop mechanism for knowledge dissemination, accumulation, and value distribution is imperfect, resulting in insufficient interaction efficiency.

[0003] How to achieve efficient integration of multimodal knowledge, dynamic assessment of knowledge credibility, and flexible access control and trusted knowledge interaction around an individual's core identity remains to be solved. As can be seen from the above, the problems of loose binding between knowledge and identity, difficulty in ensuring credibility, and inefficient interaction in existing platforms still need to be addressed. Summary of the Invention

[0004] To address the problems of loose binding between knowledge and identity, difficulty in ensuring credibility, and inefficient interaction in existing platforms, this application provides a method and system for multi-user knowledge interconnection and intelligent representative interaction.

[0005] Firstly, this application provides a method for multi-user knowledge interconnection and intelligent representative interaction, employing the following technical solution:

[0006] A method for multi-user knowledge interconnection and intelligent representative interaction includes:

[0007] Leveraging large language models and retrieval-enhanced generation technology, the system performs real-time and efficient parsing of user-provided documents, audio, and video content at the multimodal knowledge access end. It extracts structured knowledge units through the collaborative efforts of optical character recognition (OCR), speech-to-text (STO) models, and large language models. Combined with retrieval-enhanced generation technology, it achieves accurate association and deduplication of knowledge units, generating multi-dimensional knowledge representations including content vectors, topic vectors, and skill vectors. These multi-dimensional knowledge representations establish personal knowledge profiles based on entity recognition, topic clustering, and skill tag classification technologies, and the knowledge vectors are stored in a personal private knowledge base through a fine-grained access control mechanism.

[0008] Intelligent representatives are constructed based on large language models and retrieval enhancement generation technology. These intelligent representatives locate relevant knowledge units through deep matching of knowledge vectors and query intent, enabling accurate cross-user knowledge retrieval and generating high-confidence, accurate answers in question-and-answer scenarios. At the same time, they control knowledge access permissions based on user authorization scope.

[0009] In the process of multi-user knowledge interconnection, the quality of knowledge contribution, the timeliness of interactive feedback and other parameters are analyzed by relying on a large language model. The credibility level of user knowledge is dynamically calculated. The credibility level of user knowledge is updated in real time and drives the adaptive adjustment of multi-level permission control strategy to achieve differentiated knowledge sharing.

[0010] Based on the reinforcement learning mechanism, user social behavior is transformed into knowledge graph edge weight optimization signals, and a closed loop of knowledge interaction, social connection, knowledge dissemination and capability improvement is constructed. Among them, the intelligent representative actively initiates cross-user dialogue based on the user's knowledge credibility level and knowledge vector similarity, verifies the confidence of question and answer through retrieval enhancement generation technology, and automatically precipitates high-value question and answer into searchable knowledge units.

[0011] By establishing ownership and quantifying the value of knowledge contributions and usage through a dual-track recording mechanism, relying on a large language model to assist in evaluating knowledge value, allocating knowledge asset revenue based on the Shapley value algorithm, and dynamically generating credible evidence for knowledge payment, representative subscriptions, and enterprise service settlements.

[0012] Optionally, in the process of real-time parsing of user-provided documents, audio, and video content and generating multi-dimensional knowledge representations to construct a personal knowledge profile at the multimodal knowledge access terminal, the method also includes:

[0013] A lightweight optical character recognition model is used to perform layout analysis and character recognition on document images. For blurry documents, an adaptive denoising algorithm is first used to optimize the image quality, and then the structured knowledge units of the text are extracted from the optimized image.

[0014] Audio content is converted into text format through a speech-to-text model. Relying on a large language model and contextual semantics, transcription errors are corrected and the understanding of text content is deepened. The text understanding results are labeled using a pre-trained skill tag classification model, which includes a hierarchical tag system of basic skills, advanced skills, and professional skills. The model is combined with retrieval enhancement generation technology to assist in accurate tag matching.

[0015] In the process of generating multi-dimensional knowledge representation, content vectors are generated through a text embedding adaptive model, topic vectors are generated through a latent Dirichlet distribution topic clustering algorithm combined with retrieval enhancement generation technology, and skill vectors are quantized based on a hierarchical skill tag system of basic skills, advanced skills, and professional skills.

[0016] Based on entity recognition technology, topic clustering technology and skill tag classification technology, multi-dimensional knowledge representation is transformed into a personal knowledge profile. Through a fine-grained access control mechanism, the encoded knowledge vector is stored in the personal private knowledge base.

[0017] Optionally, in the process of intelligent representatives achieving accurate cross-user knowledge retrieval through large language models and retrieval enhancement generation techniques, the method also includes:

[0018] A global knowledge index is constructed for user knowledge vectors in a hierarchical manner, consisting of personal private vector clusters, domain-shared vector clusters, and global general vector clusters. Among them, personal private vector clusters only retain the user's core knowledge vectors, domain-shared vector clusters associate homogeneous knowledge vectors of users in the same domain, and global general vector clusters store basic general knowledge vectors.

[0019] Before initiating a cross-user knowledge query, the intelligent representative first determines the response depth based on the social relationship between users and the permission level of the knowledge file. The response depth includes the globally visible circle, the follower circle, the close friend circle, and the personal privacy circle. Only the response content within the corresponding circle's permission is returned to the querying party.

[0020] Users can issue commands to the intelligent representative using natural language, access and view content within the platform that they have access to, and perform various operations such as content retrieval, filtering, association, and saving using natural language.

[0021] When initiating a cross-user knowledge query, the system first parses the user's query intent using a large language model to generate a query vector. Then, relying on retrieval enhancement generation technology, it performs a coarse matching operation between the knowledge vector and the query intent in the global general vector cluster. Next, it performs a precise matching operation in the domain shared vector cluster. Finally, it performs a precise positioning operation in the target user's personal private vector cluster.

[0022] In this process, all matching processes combine contextual association analysis with retrieval enhancement generation technology. The large language model integrates the matching results to generate high-confidence accurate answers. The query results return the associated summary of knowledge vectors and the corresponding high-confidence response content, without returning the original knowledge units.

[0023] Optionally, in the process of dynamically calculating the user's knowledge trust level and driving permission adjustment based on a large language model, the method also includes:

[0024] Each user's local node calculates its local reputation sub-score based on the accuracy, completeness, and timeliness of its knowledge contribution, and uses a large language model to perform consistency verification and integration of the local reputation sub-scores.

[0025] The system pre-sets reputation calculation rules, in which the quality of knowledge contribution is the primary weighting factor, interactive feedback is the secondary weighting factor, and timeliness is the supplementary weighting factor. It relies on a large language model to analyze the actual value of knowledge contribution and the true intention of interactive feedback, and generates and records the total reputation score.

[0026] When the number of citations of a user's knowledge contribution increases significantly, the system triggers a positive incremental adjustment to the user's knowledge credibility level. When a user's feedback response times out, the system triggers a negative deduction adjustment to the user's knowledge credibility level. In particular, whenever the user's knowledge credibility level enters a new threshold range, the large language model evaluates the user's current knowledge needs and permission matching degree, and automatically triggers the adaptive adjustment of the corresponding multi-level permission control strategy.

[0027] Optionally, in the process of transforming user social behavior into knowledge graph edge weight optimization signals and constructing a closed loop based on reinforcement learning mechanisms, the method also includes:

[0028] The state space of reinforcement learning includes the current user's knowledge credibility level, knowledge vector similarity, and historical interaction frequency; the action space is the increase or decrease of the edge weights of the knowledge graph; and the reward function includes three dimensions: knowledge dissemination breadth, user ability improvement coefficient, and interaction feedback satisfaction rate.

[0029] When the intelligent representative detects that the user's knowledge credibility level has reached a preset threshold range and the knowledge vector similarity has reached a preset matching degree, it automatically initiates a cross-user dialogue.

[0030] The criteria for determining high-value question-answering are that the question-answering content is cited by multiple users, the accuracy of knowledge tracing reaches the preset standard, and the confidence level is verified by retrieval enhancement generation technology. Question-answering pairs that meet the criteria are automatically precipitated into new nodes and edges of the knowledge graph through entity linking technology, and the structured storage format of knowledge units is optimized by a large language model.

[0031] Optionally, in the process of establishing ownership and quantifying the value of knowledge contributions and usage through a dual-track recording mechanism, the methods also include:

[0032] The dual-track recording mechanism includes knowledge contribution records and behavior usage records. Knowledge contribution records record the creator, generation time, content hash value, and permission level of the knowledge unit, while behavior usage records the caller, call time, call scenario, and payment amount of the knowledge unit.

[0033] The dual-track record achieves synchronous evidence storage through a timestamp synchronization mechanism. When a dispute over intellectual property rights occurs, the ownership confirmation operation is completed by comparing the content hash value and timestamp of the dual-track record.

[0034] The value quantification dimensions of a knowledge unit include originality, scarcity, and practicality. Originality is determined by analyzing the differences between the knowledge unit and the existing knowledge base using a large language model. Scarcity is determined by the proportion of similar knowledge units in the existing stock. Practicality is determined by semantic analysis of user feedback after invocation.

[0035] Optionally, in the process of allocating knowledge asset revenue and generating credible settlement evidence based on the Shapley value algorithm, the method further includes:

[0036] For knowledge payment scenarios, the revenue distribution factors of the Shapley value algorithm include the contribution weight of knowledge creators, the promotion weight of knowledge disseminators, and the platform's technical service weight. Among them, the creator weight is set according to the value quantification score of the knowledge unit, the disseminator weight is set according to the forwarding and citation situation, and the platform weight is a fixed percentage.

[0037] For the subscription scenario, the accuracy of the smart representative's response will be included in the correction factor for the Shapley value calculation. When the response accuracy reaches the preset standard, the creator's revenue weight will be adjusted upward.

[0038] For enterprise service settlement scenarios, the Shapley value allocation coefficient is adjusted based on the enterprise user's call status and the degree of customization of the knowledge unit; in all scenarios, the settlement operation relies on the large language model to generate a detailed list, automatically trigger the settlement process and record it, and dynamically generate credible evidence for knowledge payment, representative subscription and enterprise service settlement.

[0039] Secondly, this application provides a system for multi-user knowledge interconnection and intelligent representative interaction, employing the following technical solution:

[0040] A system for multi-user knowledge interconnection and intelligent representative interaction, comprising:

[0041] The multimodal knowledge parsing and profile building module, relying on a large language model and retrieval-enhanced generation technology, performs real-time and efficient parsing of user-provided documents, audio, and video content at the multimodal knowledge access end. It extracts structured knowledge units through the collaborative extraction of optical character recognition models, speech-to-text models, and large language models, and combines retrieval-enhanced generation technology to achieve accurate association and deduplication of knowledge units, generating a multi-dimensional knowledge representation including content vectors, topic vectors, and skill vectors. The multi-dimensional knowledge representation establishes a personal knowledge profile based on entity recognition, topic clustering, and skill tag classification technologies, and stores the knowledge vectors in a personal private knowledge base through a fine-grained access control mechanism.

[0042] The intelligent representative and cross-user retrieval module are constructed based on a large language model and retrieval enhancement generation technology. The intelligent representative locates relevant knowledge units through deep matching of knowledge vectors and query intent, realizes accurate cross-user knowledge retrieval, generates high-confidence accurate answers in question-and-answer scenarios, and controls knowledge access permissions based on user authorization scope.

[0043] The reputation calculation and access control module, in the process of multi-user knowledge interconnection, relies on a large language model to analyze the quality of knowledge contribution, the timeliness of interactive feedback and other parameters, and dynamically calculates the user knowledge credibility level. The user knowledge credibility level is updated in real time and drives the adaptive adjustment of multi-level access control strategies to achieve differentiated knowledge sharing.

[0044] The social-linked knowledge closed-loop module transforms user social behavior into knowledge graph edge weight optimization signals based on reinforcement learning mechanisms, constructing a closed loop of knowledge interaction, social connection, knowledge dissemination, and capability enhancement. Among them, the intelligent representative actively initiates cross-user dialogues based on the user's knowledge credibility level and knowledge vector similarity, verifies the confidence of question and answer through retrieval enhancement generation technology, and automatically precipitates high-value question and answer into searchable knowledge units.

[0045] The Knowledge Ownership Confirmation and Revenue Module uses a dual-track recording mechanism to confirm the ownership and value of knowledge contributions and usage behaviors. It relies on a large language model to assist in evaluating the value of knowledge, allocates knowledge asset revenue based on the Shapley value algorithm, and dynamically generates credible evidence for knowledge payment, representative subscription, and enterprise service settlement.

[0046] Thirdly, this application provides an electronic device that adopts the following technical solution:

[0047] An electronic device includes a processor in which a program is running for the method of multi-user knowledge interconnection and intelligent representative interaction as described in any one of the preceding claims.

[0048] Fourthly, this application provides a storage medium, which adopts the following technical solution:

[0049] A storage medium storing a program for a method of multi-user knowledge interconnection and intelligent representative interaction as described in any one of the foregoing.

[0050] In summary, this application includes at least one of the following beneficial technical effects:

[0051] By using multimodal knowledge parsing and personal knowledge profile construction, various knowledge resources such as documents, audio, and video are transformed into multi-dimensional representations containing content, themes, and skill vectors. These representations are then deeply linked to user identities to form exclusive knowledge profiles. Coupled with a dual-track recording mechanism, knowledge contributions and usage behaviors are precisely weighted, completely solving the problem of loose binding between knowledge and user identity on existing platforms. At the same time, relying on a large language model, the credibility level of user knowledge is dynamically calculated. A comprehensive evaluation is conducted based on the quality of knowledge contributions, the timeliness of interactive feedback, and other multi-dimensional factors. Furthermore, retrieval enhancement generation technology is used to verify the confidence of question answers and the accuracy of knowledge traceability. From generation to use, the credibility of knowledge is guaranteed throughout the entire chain, effectively solving the pain point of difficulty in judging the reliability of knowledge.

[0052] By designing intelligent representatives and hierarchical knowledge indexes, and combining user social relationships with knowledge permission levels to define response depth, along with natural language command-based operation functions, the barriers to knowledge retrieval and interaction are significantly reduced. At the same time, the reinforcement learning mechanism transforms social behavior into knowledge graph optimization signals, constructs a closed loop of knowledge interaction, and automatically accumulates high-value knowledge units. This makes cross-user knowledge matching more accurate and permission control more flexible, significantly improving the efficiency of knowledge interaction and solving the problems of cumbersome interaction processes and insufficient accuracy in existing platforms.

[0053] Furthermore, compared to traditional knowledge-sharing platforms, the solution proposed in this application introduces an intelligent question-and-answer agent with automatic questioning and follow-up questioning capabilities. During the construction and maintenance of the user knowledge base, this agent proactively guides information completion and structural optimization. Combined with user feedback guidance and quality assessment dashboards based on the answer content, it achieves dynamic identification and optimization of low-quality, redundant, or incomplete knowledge content. This reduces the number of invalid responses in cross-user knowledge retrieval, lowers the risk of mis-exposure of knowledge due to lack of permissions or context, and significantly improves the efficiency and reusability of high-value knowledge accumulation within the platform. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a method for multi-user knowledge interconnection and intelligent representative interaction according to an exemplary embodiment.

[0055] Figure 2 This is a structural block diagram of a system for multi-user knowledge interconnection and intelligent representative interaction, according to an exemplary embodiment. Detailed Implementation

[0056] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0057] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0058] This application discloses a method for multi-user knowledge interconnection and intelligent representative interaction, referring to... Figure 1 ,include:

[0059] The S100, relying on a large language model and retrieval-enhanced generation technology, performs real-time and efficient parsing of user-provided documents, audio, and video content at the multimodal knowledge access end. It extracts structured knowledge units through the collaborative extraction of optical character recognition models, speech-to-text models, and large language models. Combined with retrieval-enhanced generation technology, it achieves accurate association and deduplication of knowledge units, generating multi-dimensional knowledge representations including content vectors, topic vectors, and skill vectors. The multi-dimensional knowledge representation establishes a personal knowledge profile based on entity recognition, topic clustering, and skill tag classification technologies, and stores the knowledge vectors in a personal private knowledge base through a fine-grained access control mechanism.

[0060] The S100 execution process specifically includes:

[0061] Step 1, Multimodal Data Access and Preprocessing:

[0062] It receives various types of data uploaded by users, including documents (such as PDFs, Word documents, and scanned copies), audio (such as recordings and voice files), and video (such as teaching videos and meeting recordings). All data is uniformly connected to the multimodal knowledge access terminal.

[0063] Preliminary preprocessing of various data types: redundant formatting (such as extra spaces and invalid watermarks) is removed from documents; noise reduction is performed on audio (filtering background noise and background noise); keyframes (static images) and audio tracks are extracted from videos, and then processed separately to prepare for subsequent analysis.

[0064] Step 2, Modal content parsing and text conversion:

[0065] Document parsing: A lightweight optical character recognition model is used to perform layout analysis (dividing the document image into areas such as title, body text, and chart labels) and character recognition operations. If the document has problems such as blurriness or illegible text, the image quality is first optimized by an adaptive denoising algorithm (dynamically adjusting denoising parameters), and then the text content is accurately extracted from the optimized image to form plain text data.

[0066] Audio parsing: The pre-processed audio content is converted into text format using a speech-to-text model; relying on a large language model and contextual semantic logic, it corrects homophony errors, punctuation errors and other problems that occur during the transcription process, while deepening text understanding (recognizing technical terms and core viewpoints).

[0067] Video parsing: Keyframe images extracted from the video are processed using the optical character recognition process similar to document parsing to extract text information from the images; audio tracks split from the video are transcribed and optimized according to the audio parsing process, ultimately transforming the video content into a structured form that associates text data with image features.

[0068] Step 3, Extraction and Optimization of Structured Knowledge Units:

[0069] Based on a large language model, semantic segmentation is performed on all text data after conversion of documents, audio, and video to extract independent knowledge units (such as a professional viewpoint, a set of operation steps, a technical parameter, etc.) to ensure that each knowledge unit has complete semantics.

[0070] By combining retrieval enhancement generation technology, the extracted knowledge units are compared with the platform's existing knowledge base. Through semantic similarity analysis, the knowledge units are accurately deduplicated (removing duplicate or highly similar knowledge units), and at the same time, the association mapping of related knowledge units is established (such as marking the association relationship of different knowledge units on the same topic).

[0071] Step 4, Generation of multi-dimensional knowledge representation:

[0072] Content Vector Generation: Each structured knowledge unit is vectorized and encoded using a text embedding adaptive model, transforming text semantics into high-dimensional numerical vectors to accurately capture the core content features of the knowledge unit.

[0073] Topic Vector Generation: The Latent Dirichlet Distribution topic clustering algorithm is used to classify all knowledge units into topics. Combined with retrieval enhancement generation technology, the topic recognition accuracy is optimized to generate topic vectors that can represent the topics to which the knowledge units belong.

[0074] Skill vector generation: Skill tags are labeled on knowledge units using a pre-trained skill tag classification model (including a hierarchical tag system of basic skills, advanced skills, and professional skills). This is combined with retrieval enhancement generation technology to assist in accurate tag matching (avoiding tag misjudgment). Then, the tags are quantized and encoded based on the hierarchical tag system to generate skill vectors.

[0075] Step 5, Personal Knowledge Profile Construction:

[0076] Based on entity recognition technology, user-related entities (such as user's areas of expertise, skills, and topics of interest) are extracted from multi-dimensional knowledge representation. The user's core knowledge domains are summarized through topic clustering technology, and the user's skill level is clarified by combining the skill tag classification results.

[0077] By integrating content vectors, topic vectors, and skill vectors, a personal knowledge profile can be constructed that comprehensively represents a user's knowledge reserves, skill level, and interest direction, thereby achieving a deep binding between knowledge and user identity.

[0078] Step 6, Fine-grained access control and knowledge storage:

[0079] Fine-grained permission tags (such as globally visible, visible to followers, privately visible, etc.) are assigned to each knowledge unit and its corresponding multi-dimensional knowledge vector. Permission tags are preset by the user or automatically recommended based on the sensitivity of the knowledge content.

[0080] By verifying the permission attributes of knowledge vectors through a fine-grained permission control mechanism, knowledge vectors with permission tags are stored in the user's personal private knowledge base in a unified format, and an index is built (related knowledge units, multi-dimensional vectors, permission tags, user identity) to support subsequent fast retrieval and permission verification.

[0081] By employing precise modal parsing, structured knowledge extraction, multi-dimensional vector representation, and personal knowledge profile construction, this approach thoroughly resolves the issues of inefficient multimodal knowledge integration and loose binding between knowledge and user identity in existing platforms. Simultaneously, it utilizes retrieval-enhanced generation technology to deduplicate and associate knowledge units. Combined with fine-grained access control and private knowledge base storage, it ensures the standardization, uniqueness, and security of knowledge, providing a precise, structured, and traceable knowledge foundation and user identity association basis for subsequent intelligent representative construction, cross-user knowledge retrieval, and trust level assessment.

[0082] S200 constructs intelligent representatives based on large language models and retrieval enhancement generation technology. The intelligent representatives locate relevant knowledge units through deep matching of knowledge vectors and query intent, realize accurate cross-user knowledge retrieval, generate high-confidence accurate answers in question-and-answer scenarios, and control knowledge access permissions based on user authorization scope.

[0083] The S200 execution process specifically includes:

[0084] Step 1, Smart Representative Construction and Function Initialization:

[0085] Based on a large language model (responsible for semantic understanding, intent parsing, and answer generation) and retrieval enhancement generation technology (responsible for knowledge retrieval and credibility verification), the core architecture of the intelligent representative is constructed, integrating four functional modules: query intent parsing module, knowledge retrieval matching module, high-confidence answer generation module, and permission verification module.

[0086] Initialize the intelligent representative's functions: import the user's personal knowledge profile and corresponding multi-dimensional knowledge vector (from the S100 private knowledge base), synchronize the user's preset authorization scope rules (such as the visibility layer of knowledge units, cross-user interaction permission thresholds, etc.), and establish retrieval interfaces with the platform's global knowledge index and other users' private knowledge bases to ensure that the intelligent representative has cross-source knowledge access capabilities.

[0087] Step 2, Query intent parsing and vector transformation:

[0088] It receives query requests initiated by users (supports natural language input, such as text questions and text queries after speech transcription). In the case of multi-turn dialogue scenarios, the query intent parsing module will combine the historical dialogue context to avoid intent deviation caused by isolated parsing.

[0089] By leveraging a large language model to semantically decompose the query text and extract core query elements (such as query topic, required knowledge type, target user scope, etc.), the query intent is transformed into a standardized query vector through a text embedding adaptive model, ensuring that the vector dimension is consistent with the knowledge vector generated by S100, thus laying the foundation for subsequent matching.

[0090] Step 3, Permission verification and response depth determination for cross-user knowledge retrieval:

[0091] The permission verification module first extracts the identity information of the querying user and their social relationship with the target user (the user who owns the knowledge) (such as ordinary users, followers, close friends), and then calls the permission tags of the target user's knowledge unit (from S100 fine-grained permission control).

[0092] The response depth is determined based on "social relationships between users + the permission level of the knowledge file": matching the corresponding level in the globally visible circle, follower circle, close friend circle, and personal privacy circle, clarifying the scope of knowledge that the querying party can only access (such as directly rejecting the search of knowledge units in the personal privacy circle), and preventing unauthorized access.

[0093] Step 4, Hierarchical knowledge retrieval and precise matching:

[0094] The intelligent representative accesses the platform's hierarchical knowledge index (personal private vector clusters, domain-shared vector clusters, and globally common vector clusters) through the knowledge retrieval and matching module, and performs the retrieval in the order of "coarse matching first, then fine matching":

[0095] First, in the global general vector cluster, the semantic similarity between the query vector and the knowledge vector is calculated to filter out the preliminarily relevant knowledge units (excluding completely irrelevant domain knowledge).

[0096] Furthermore, within the domain-shared vector cluster, we focus on homogeneous knowledge units of users in the same domain to perform more refined similarity matching and narrow down the search scope;

[0097] Finally, within the target user's private vector cluster, and combined with the permission range corresponding to the response depth, the precisely matching knowledge unit is located to ensure that the search results both match the query intent and comply with permission requirements.

[0098] Step 5, High-confidence answer generation and secondary permission verification:

[0099] The retrieval enhancement generation technology inputs precisely matched knowledge units (including multi-dimensional vectors, original knowledge fragments, and source annotations) into a large language model. The model integrates and logically organizes the knowledge based on the query intent, generating a complete and semantically fluent accurate answer. At the same time, it annotates the knowledge source to improve the credibility of the answer.

[0100] After the answer is generated, the permission verification module performs a second verification: confirming that the answer content does not exceed the permission range corresponding to the response depth and does not contain privacy knowledge that the target user has not authorized. Only after the verification is passed will the answer result be returned to the querying party.

[0101] Step 6, Natural Language Interaction Support and Operation Response:

[0102] Users can initiate additional operations to the intelligent representative via natural language commands, such as "filter related knowledge from the past 3 months", "view similar knowledge units of this user", and "save the current answer to my knowledge base". The intelligent representative will automatically execute the corresponding operation after parsing the command, without the need for a fixed process switching.

[0103] By constructing an intelligent representative integrating "intent parsing, permission verification, hierarchical retrieval, and high-confidence generation," the existing platform addresses issues such as low retrieval accuracy, rigid permission control, and high interaction barriers in cross-user knowledge interaction. It leverages a large language model and retrieval enhancement generation technology to achieve deep matching between query intent and knowledge vectors. Combined with social relationships and permission-level response depth judgment, it ensures both the accuracy and efficiency of cross-user knowledge retrieval while strengthening knowledge security through dual permission verification. Simultaneously, natural language interaction reduces user operation costs, and high-confidence responses and source annotations enhance the credibility of knowledge use.

[0104] In the process of multi-user knowledge interconnection, S300 relies on a large language model to analyze the quality of knowledge contributions, the timeliness of interactive feedback, and other parameters, dynamically calculates the trust level of user knowledge, updates the trust level of user knowledge in real time, and drives the adaptive adjustment of multi-level permission control strategies to achieve differentiated knowledge sharing.

[0105] The S300 execution process specifically includes:

[0106] Step 1, Core Parameter Acquisition and Data Preprocessing:

[0107] Raw data for the three core parameters is collected, with data sources covering both preceding processes and real-time interactive scenarios:

[0108] Knowledge contribution quality data: comes from user-uploaded and structured knowledge units in S100, including the accuracy (comparison results with authoritative knowledge bases), completeness (whether core information is missing), and professionalism (accuracy of professional terminology and logical rigor).

[0109] Interaction feedback data: from S200 cross-user knowledge interaction records, including the number of times other users have cited the user's knowledge unit, positive / critical feedback content, and Q&A acceptance rate (whether the queryer approves of the answer);

[0110] Timeliness data: Records the update frequency of user knowledge units (whether the latest information is added), the response speed of interactive feedback (whether questions / inquiries are answered in a timely manner), and the degree of matching between the creation time of knowledge units and industry trends.

[0111] The collected raw data is preprocessed: the data format is standardized (e.g., the ratings of different dimensions are standardized to a range of 0-10 points), abnormal data is removed (e.g., maliciously obtained fake references and meaningless feedback), and the subjective feedback content is semantically analyzed through a large language model (distinguishing between valid positive reviews, constructive questions, and invalid complaints) and converted into quantitative indicators.

[0112] Step 2, Parameter weight allocation and quantization modeling:

[0113] The system pre-sets weighting rules and clarifies the priority of the three core parameters: the quality of knowledge contribution is the primary weighting factor, interactive feedback is the secondary weighting factor, and timeliness is the supplementary weighting factor (the weight ratio can be dynamically adjusted according to the platform scenario, but the core priority is fixed).

[0114] Based on the large language model, the preprocessed parameters are quantitatively modeled: the sub-indicators of knowledge contribution quality (accuracy, completeness, professionalism) are weighted and summed to obtain the total score of knowledge contribution quality; similarly, the sub-indicators of interactive feedback and timeliness are quantitatively summed to obtain the total score of the corresponding dimension; finally, the total scores of the three dimensions are weighted a second time according to the preset weights to generate the initial quantitative score.

[0115] Step 3, Calculation and classification of user knowledge credibility level:

[0116] The large language model integrates the initial quantitative scores and combines them with the platform's preset level division rules (such as the level corresponding to the score range) to divide the user's knowledge credibility level into several levels (such as excellent, good, average, and low, and the number of levels can be flexibly configured).

[0117] During the ranking calculation process, a dynamic calibration mechanism is introduced: by comparing the parameter distribution of users in the same domain with a large language model, the ranking deviation caused by the special characteristics of a single user in the domain is avoided (for example, when a user in a niche domain contributes less but of higher quality, the calibration score is appropriately increased); at the same time, the historical ranking change trend is referenced to ensure that the ranking fluctuation is stable (for example, a short-term surge in citations will not result in a direct and significant upgrade, but stability needs to be observed).

[0118] Step 4, Real-time update of knowledge trust level triggering and execution:

[0119] Set the level update trigger conditions to ensure real-time performance:

[0120] Active triggering: The level is recalculated immediately when a user uploads a new knowledge unit, a knowledge unit is cited / feedback, or a user updates an existing knowledge unit;

[0121] Passive triggering: The system performs batch verification and updates of all users' levels daily. At the same time, when the cumulative changes of a user's core parameters reach a preset threshold (such as a change of more than 2 points in the total knowledge contribution quality score), the update is automatically triggered.

[0122] After the level is updated, it is synchronized to the user's personal knowledge profile (associated with S100), the platform's permission control system, and the smart representative (associated with S200) to ensure that the latest level data is used in all aspects.

[0123] Step 5, adaptive adjustment of multi-level permission control strategy:

[0124] The system has a pre-defined "level-permission" mapping rule, with different knowledge trust levels corresponding to different permission configurations. Permission adjustments are automatically triggered when the level is updated.

[0125] High-trust level users: gain a wider knowledge sharing scope (such as the ability to independently set an increased limit on the number of knowledge units that are "globally visible"), priority search rights (their knowledge units are ranked higher when searching across users), and autonomy in adjusting permissions (they can temporarily open up some private knowledge to specific users).

[0126] Medium-trust level users: Maintain basic sharing permissions, and can only set some knowledge to "visible to followers". Search ranking is based on normal weight.

[0127] Users with low trust levels: Sharing scope is restricted (visible only to close friends or private users), knowledge unit search ranking is lower, and permissions need to accumulate effective contributions and positive feedback to be upgraded.

[0128] During the permission adjustment process, the fine-grained permission tags of S100 are linked to ensure that permission control does not conflict (e.g., the private knowledge of high-level users is still protected by the initial permissions and is not automatically opened due to the level upgrade).

[0129] Step 6, Implementing Differentiated Knowledge Sharing:

[0130] In cross-user knowledge interaction scenarios, the access control strategy takes effect: when the querying party retrieves information, the system first verifies the knowledge credibility level of the target user, and then combines the access tags of its knowledge units to display differentiated results.

[0131] For example, the "globally visible" knowledge units of users with high trust levels can be retrieved by all users, and the answers are more comprehensive; the knowledge units of users with low trust levels are only open to their close friends, and the answers will be marked with "for reference only";

[0132] At the same time, feedback on the permission adjustment results to knowledge contributors (such as prompts that more sharing permissions can be opened after the level is upgraded) will incentivize users to accumulate high credibility levels.

[0133] By collecting multi-dimensional parameters, using dynamic level calculation driven by a large language model, and adjusting permissions adaptively, the existing platform completely solves the problems of difficulty in ensuring knowledge credibility, rigid permission control, and insufficient differentiated sharing. It strongly binds the quality of knowledge contributions, interactive feedback, timeliness, and user credibility level, making the "credibility" of knowledge quantifiable and traceable. At the same time, through the linkage between level and permission, it realizes a differentiated sharing mechanism of "high credibility, high permission; low credibility, strict control," which not only ensures the reliability of knowledge during cross-user interactions but also incentivizes users to output high-quality knowledge.

[0134] S400, based on reinforcement learning, transforms user social behavior into knowledge graph edge weight optimization signals, constructing a closed loop of knowledge interaction, social connection, knowledge dissemination, and capability enhancement. In this loop, the intelligent representative proactively initiates cross-user dialogues based on the user's knowledge credibility level and knowledge vector similarity. It verifies the confidence of question-and-answer questions through retrieval enhancement generation technology and automatically precipitates high-value questions and answers into searchable knowledge units.

[0135] The S400 execution process specifically includes:

[0136] Step 1, User social behavior data collection and preprocessing:

[0137] Collect core social behavior data in multi-user knowledge interconnection scenarios, covering the entire knowledge interaction chain: including behaviors such as following, unfollowing, liking, commenting (including knowledge-related supplementary, questioning, and affirmation comments), forwarding knowledge units, privately messaging for knowledge consultation, and jointly participating in knowledge discussions; at the same time, associate the knowledge unit information associated with the behavior (such as the knowledge topic and skill type being interacted with) and the knowledge credibility level of the two interacting users.

[0138] The collected social behavior data is preprocessed: invalid behaviors (such as accidental clicks and meaningless spam comments) are removed; semantic analysis of text-based social behaviors (such as comments and private messages) is performed using a large language model to extract core intentions (such as seeking knowledge supplementation, confirming the accuracy of knowledge, and sharing relevant experiences); the behavior data is classified and labeled (such as positive interactions, neutral interactions, and question-based interactions), and attributes such as the time of occurrence, frequency, and related knowledge dimensions of the behavior are recorded to form a standardized behavior dataset.

[0139] Step 2, Reinforcement learning model initialization and signal transformation:

[0140] Initialize the core framework of reinforcement learning and clarify the key elements of the model:

[0141] State space: includes the current user's knowledge credibility level, knowledge vector similarity (related to S100 multi-dimensional knowledge representation), historical social interaction frequency, and current edge weight distribution of the knowledge graph;

[0142] Action space: The range of increase or decrease in the weights of the knowledge graph edges (e.g., weight increase, maintenance, decrease), and the range of weight adjustment is set according to the strength of knowledge association and the degree of social intimacy;

[0143] Reward function: It covers three core dimensions: breadth of knowledge dissemination (number of users covered by interactive behavior), user ability improvement coefficient (quality change of knowledge uploaded by users in the future), and positive feedback rate (proportion of positive interaction), and calculates the reward value by weighting according to preset weights.

[0144] Based on reinforcement learning models, preprocessed social behavior data is transformed into edge weight optimization signals for knowledge graphs: positive interaction behaviors (such as mutual likes between users with high credibility levels and supplementary knowledge comments) generate weight increase signals; question-type interactions (such as reasonable questions that receive effective responses) generate weight fine-tuning signals; and invalid or negative interactions (such as meaningless comments and malicious questions) generate weight maintenance or reduction signals.

[0145] Step 3, Dynamic Optimization of Knowledge Graph Edge Weights:

[0146] The platform's knowledge graph (which is associated with the S100 personal knowledge profile and knowledge unit relationships) is invoked to locate the graph nodes and edges corresponding to the optimization signals: nodes include user nodes, knowledge unit nodes, topic nodes, and skill nodes, and edges include the "creation / ownership" relationship between users and knowledge units, the "social interaction" relationship between users, and the "association" relationship between knowledge units.

[0147] The weights of corresponding edges are adjusted according to the optimization signals: high-value positive signals (such as deep interactions between highly trusted users based on core knowledge) significantly increase edge weights and enhance the retrieval priority of associated nodes; regular positive signals slightly increase weights; fine-tuning signals fine-tune weights according to the direction of interaction semantics; negative signals maintain or reduce weights to reduce the impact of invalid associations.

[0148] After weight optimization, the knowledge graph index is updated synchronously to ensure that subsequent cross-user knowledge retrieval (related to S200) can prioritize matching high-weight relationships, thereby improving the accuracy of knowledge dissemination.

[0149] Step 4, initiate judgment and execution across user dialogue:

[0150] The intelligent representative (associated with S200) reads the optimized knowledge graph edge weights, user knowledge credibility level, and knowledge vector similarity in real time, and sets the conditions for initiating a dialogue:

[0151] Basic conditions: The similarity of the knowledge vectors of both parties reaches the preset matching level (such as the similarity of core topic or skill vectors meeting the standard), and the knowledge credibility level of the target user meets the interaction permission requirements of the initiator;

[0152] Optimization conditions: The edge weights of the knowledge graphs of both parties reach a preset threshold (such as the formation of high-weight social-knowledge associations due to multiple positive interactions), or the knowledge units of one party are highly matched with the knowledge gaps of the other party (identified through knowledge vector comparison).

[0153] Once the conditions are met, the intelligent representative generates a personalized dialogue invitation: combining the knowledge connections between the two parties, it is pushed to the initiating user in natural language; after the initiating user confirms, the intelligent representative pushes a dialogue request to the target user, clarifying the interaction topic and knowledge connection basis, thereby increasing the willingness to respond.

[0154] Step 5, Question-answer confidence verification and high-value judgment:

[0155] During cross-user dialogues, the intelligent representative records the question-and-answer content (questions, answers, supplementary explanations, etc.) in real time, forming complete question-and-answer pairs; after the dialogue ends, the confidence level of the question-and-answer pairs is verified using retrieval-enhanced generation technology.

[0156] Compare the answers to the platform's knowledge base (including the S100 private knowledge base and the shared knowledge base) to verify the accuracy and consistency of the answers.

[0157] By analyzing the rigor of question-and-answer logic and the completeness of knowledge coverage through large language models, questions and answers with logical contradictions or missing information are eliminated.

[0158] The confidence level of the question and answer is determined by combining interactive feedback (such as whether the questioner agrees or asks for further consultation).

[0159] The criteria for determining high-value Q&A are as follows: the confidence level of the Q&A meets the standard, the content of the answer can fill the knowledge gap of the platform (low similarity with the existing knowledge base), the Q&A is cited by both parties in the dialogue or third-party users, and the accuracy of knowledge tracing meets the standard (clearly indicating the source of the knowledge unit corresponding to the answer). Q&A pairs that meet the above conditions are marked as high-value Q&A.

[0160] Step 6, High-Value Question and Answer Accumulation and Closed-Loop Iteration:

[0161] High-value question-and-answer pairs are structured: core questions, key answer points, and supplementary explanations are extracted and transformed into standardized knowledge units (using the multi-dimensional knowledge representation format of S100 to generate content vectors, topic vectors, and skill vectors); a "question-and-answer source" tag is added to the knowledge unit (associating the user identities of both parties and the dialogue scenario), and fine-grained permission tags are assigned according to the authorization scope of both parties (associating with S100 permission control).

[0162] Structured knowledge units are stored in the platform's knowledge base and simultaneously updated to the personal knowledge profiles of both parties (associated with S100), and added to the knowledge graph (adding or strengthening relevant nodes and edge relationships). At the same time, the accumulation of high-value questions and answers triggers the reward mechanism of the reinforcement learning model, further optimizing the edge weights of the knowledge graph and promoting more accurate social connections and knowledge dissemination in the future.

[0163] The newly accumulated knowledge units serve as searchable resources, which can be retrieved and used by other users through intelligent representatives (linked to S200), forming a complete closed loop of "knowledge interaction → social connection → knowledge dissemination → capability enhancement → new knowledge accumulation → new interaction".

[0164] By leveraging reinforcement learning mechanisms to connect social behavior with knowledge graphs, a dynamic and cyclical knowledge interaction loop is constructed, completely resolving the existing platform's problems of disconnect between social connections and knowledge dissemination, difficulty in accumulating high-value knowledge, and low interaction efficiency. It transforms user social behavior into a driving force for knowledge graph optimization, making knowledge connections more aligned with user interaction needs. Simultaneously, it proactively initiates precise dialogues based on trust levels and knowledge similarity, and combines retrieval enhancement generation technology to ensure the credibility of question-and-answer exchanges, achieving the automatic accumulation and reuse of high-value knowledge. This process not only strengthens the deep integration of knowledge and social interaction but also continuously optimizes the knowledge dissemination path and interaction accuracy through closed-loop iteration.

[0165] S500 uses a dual-track recording mechanism to confirm the ownership and value of knowledge contributions and usage behaviors, relies on a large language model to assist in evaluating knowledge value, allocates knowledge asset revenue based on the Shapley value algorithm, and dynamically generates credible evidence for knowledge payment, representative subscription, and enterprise service settlement.

[0166] The S500 execution process specifically includes:

[0167] Step 1, Initialization of the dual-track recording mechanism and data acquisition:

[0168] Initialize the dual-track recording core architecture, and clarify the storage rules and related logic of the two recording dimensions:

[0169] Knowledge Contribution Records: Specifically records the core ownership and attribute information of knowledge units, including the knowledge unit creator's identity (associated with S100 user identity), generation time, content hash value (unique identifier), permission level (from S100 fine-grained permission control), and knowledge type (structured knowledge corresponding to document / audio / video).

[0170] Usage behavior records: Focus on the flow and usage behavior of knowledge units, including the caller's identity, call time, call scenario (knowledge payment / representative subscription / enterprise service), payment amount (if any), usage duration, and secondary dissemination behavior (forwarding / quoting).

[0171] Real-time collection of end-to-end data: Knowledge contribution record data is synchronously accessed from the S100 knowledge unit generation stage, and behavior usage record data is captured in real time from scenarios such as S200 cross-user retrieval and S400 knowledge interaction; all data is stored in a standardized format and time sequence is marked by timestamps to ensure data traceability.

[0172] Step 2, execution of knowledge ownership confirmation:

[0173] The dual-track recording system achieves real-time data alignment through a timestamp synchronization mechanism: knowledge contribution records and behavior usage records are bound to the content hash value of the same knowledge unit, forming a one-to-one "ownership-use" association link to avoid data fragmentation.

[0174] When a dispute arises regarding intellectual property ownership, the ownership verification process is initiated: the content hash value of the disputed knowledge unit is extracted, and the creator's identity and generation timestamp in the dual-track record are compared (the timestamp of the knowledge contribution record is given priority). Combined with the binding relationship between the knowledge unit and the individual knowledge profile in S100, a complete chain of ownership evidence is formed to quickly determine the legitimate rights holder. The ownership verification result is simultaneously stored in the dual-track record as the basis for subsequent value distribution.

[0175] Step 3, Quantitative assessment of knowledge value:

[0176] A three-dimensional value assessment system is built based on a large language model, clarifying the assessment logic for each dimension:

[0177] Originality assessment: The large language model compares the semantic similarity of the knowledge unit to be evaluated with the platform's existing knowledge base (including all user private knowledge bases and shared knowledge bases), and determines the originality in reverse by the degree of difference (the higher the degree of difference, the higher the originality score).

[0178] Scarcity assessment: The scarcity level is determined by statistically analyzing the proportion of similar knowledge units within the platform (such as the number of knowledge units for specific professional skills) and combining this with the popularity of demand in the field (through analysis of search frequency and interaction volume).

[0179] Practicality assessment: Analyze the feedback data of knowledge unit calls, including the positive review rate after user calls, the skill improvement coefficient (determined by the quality change of knowledge subsequently uploaded by the caller), and the application effect in enterprise service scenarios (such as implementation conversion rate). Transform subjective feedback into quantitative scores through semantic analysis.

[0180] The scores for originality, scarcity, and practicality are weighted and summed according to preset weights to generate a total value quantification score for the knowledge unit. The score is synchronously linked to the dual-track record and serves as the core basis for revenue distribution.

[0181] Step 4, Calculate the profit distribution based on the Shapley value algorithm:

[0182] Configure the profit allocation factor and weighting rules of the Shapley value algorithm for different application scenarios:

[0183] In the knowledge payment scenario, the allocation factors include the contribution weight of knowledge creators, the promotion weight of knowledge disseminators, and the technical service weight of the platform. Among them, the creator weight is dynamically set according to the quantitative score of the knowledge unit value (the higher the score, the higher the weight), the disseminator weight is calculated based on the actual conversion effect brought by forwarding / quoting, and the platform weight is a fixed percentage.

[0184] Subscription scenario: Add a smart representative response accuracy correction factor (from S200 response effect data) to the basic allocation factor. When the response accuracy reaches the preset standard, the creator's revenue weight will be automatically increased.

[0185] For enterprise service settlement scenarios: additional adjustment coefficients for "degree of customization" (such as the adaptability of enterprise customized knowledge) and "frequency of use" are introduced, and the allocation coefficients are dynamically adjusted according to the actual usage scale of enterprise users.

[0186] The Shapley value algorithm is based on the principle of "marginal contribution" to calculate the actual contribution ratio of each participant (creator, disseminator, platform) in the process of realizing knowledge value, and generate an accurate revenue distribution plan.

[0187] Step 5, Settlement Basis Generation and Process Triggering:

[0188] Leveraging a large language model to assist in generating scenario-based settlement details: the list includes knowledge unit information (hash value, name), ownership information, value quantification score, revenue distribution details (amount for each participant), and usage behavior summary (number of calls / scenario), ensuring that each revenue is traceable.

[0189] Automatically trigger the settlement process: Generate corresponding settlement vouchers based on the scenario type (knowledge payment order, representative subscription settlement statement, enterprise service statement), and store the settlement vouchers in a dual-track record simultaneously, ensuring immutability through timestamps and content hash values;

[0190] Settlement data is synchronized to the platform's financial system and users' personal accounts. Users can query settlement details through the smart representative (linked to S200), while corporate users generate standardized settlement reports as a reliable basis for financial reconciliation.

[0191] Step 6, Dual-track recording and settlement data archiving:

[0192] Regularly archive and organize the dual-track records and settlement data, classify and store them according to knowledge type, time period, and scenario type, and retain complete "rights confirmation-quantification-allocation-settlement" link data;

[0193] Archived data supports subsequent auditing and traceability. In case of disputes over revenue distribution, the dual-track records and settlement details can be retrieved for quick verification, ensuring the transparency and fairness of the entire value distribution process.

[0194] By combining a dual-track recording mechanism, three-dimensional value quantification, and the Shapley value algorithm, the existing platform's problems of difficulty in knowledge ownership confirmation, ambiguity in value assessment, and unfair revenue distribution are thoroughly solved. It records the entire chain of knowledge contribution and usage behavior, achieves precise ownership based on content hash value and timestamp, makes knowledge value quantifiable and comparable through a large language model, and ensures the fairness of revenue distribution with the Shapley value algorithm. Ultimately, it provides a reliable settlement basis for scenarios such as knowledge payment, subscription, and enterprise services. This process not only protects the legitimate rights and interests of knowledge creators but also incentivizes users to actively contribute high-quality knowledge, while providing standardized and transparent value distribution support for the platform's commercial operation.

[0195] Based on the solutions in the above embodiments of this application, the following examples of solutions in the context of workplace skills and knowledge social networking are provided for illustration:

[0196] Suppose user Li is a senior product manager who, through this knowledge-sharing platform centered on people and business cards, uploads multimodal knowledge resources (including instructional videos, case studies, and audio explanations) such as product prototype design, requirement document writing, and user research methods, using his personal business card as the core entry point. The platform, through the S100 process and relying on large language models and retrieval-enhanced generation technology, performs modal analysis on these resources—documents are extracted into structured knowledge units using an optical character recognition model, audio is converted into text and corrected for errors using a speech-to-text model, and videos are split to extract image text and audio text separately. Ultimately, a multi-dimensional knowledge representation containing content, themes, and skill vectors is generated. Combined with entity recognition and topic clustering technologies, Li's personal knowledge profile is constructed. All knowledge vectors are deeply bound to Li's personal business card (identity) and stored in a private knowledge base, completely solving the problem of the separation between knowledge and user identity on existing platforms, achieving a credible expression of "business card as knowledge entry point."

[0197] The core purpose of this platform is to ensure the trustworthy use of knowledge centered around personal profiles. Knowledge units uploaded by Xiao Li undergo the S300 process, where a large language model analyzes their professionalism, completeness, and interactive feedback to dynamically calculate an "excellent" level of knowledge credibility. When other users search for "product requirement document writing skills" using an intelligent representative, the S200 process first determines the response depth as "visible to followers" based on the searcher's social relationship with Xiao Li (e.g., follower circles) and knowledge access level. Then, it matches relevant knowledge units from Xiao Li through hierarchical retrieval, simultaneously labeling the knowledge source as "Product Manager Xiao Li (high credibility level)." Combined with search enhancement generation technology to verify question-and-answer confidence, this allows the searcher to clearly understand the source and level of knowledge credibility, effectively solving the pain point of existing platforms where knowledge credibility is unsubstantiated and difficult to judge, making knowledge use more reassuring.

[0198] To address the inefficiency of existing platform interactions, the solution achieves efficient interconnection through a knowledge interaction closed loop built using the S400 process: Xiao Li's personal AI agent reads the platform's knowledge graph in real time and discovers that user Xiao Zhang (also interested in product management, with high knowledge vector similarity) has repeatedly browsed knowledge related to requirement documents, and since they are followers, Xiao Li proactively initiates a conversation invitation: "Hello Xiao Zhang, regarding the requirement document writing techniques you've recently been interested in, I have highly credible practical cases. Would you like to discuss them in detail?" Xiao Zhang uses natural language commands to "filter core cases from the past six months," and the AI ​​agent quickly responds and returns the corresponding resources. The high-value Q&A after the exchange is verified by retrieval enhancement generation technology and automatically precipitated as searchable knowledge units. The entire process requires no complex operations; the AI ​​agent proactively matches based on social relationships and knowledge similarity, natural language interaction lowers the operational threshold, and high-value knowledge is automatically precipitated, significantly improving knowledge interaction efficiency. This perfectly aligns with the platform's core design of using user-uploaded knowledge bases and externally interactive AI agents for credible knowledge expression.

[0199] In this embodiment of the application, the method for constructing a personal knowledge profile by real-time parsing of user-provided documents, audio, and video content and generating multi-dimensional knowledge representations at the multimodal knowledge access terminal specifically includes:

[0200] Step 1, Document Image Parsing and Structured Knowledge Extraction:

[0201] A lightweight optical character recognition model is used to first perform layout analysis on the document image (such as scanned copies or screenshots) to divide it into functional areas such as titles, body text, and chart labels. If problems such as blurriness or illegible text are detected in the document, the denoising parameters are dynamically adjusted through an adaptive denoising algorithm to optimize the image quality. The text content is then accurately extracted from the optimized clear image and broken down according to semantic integrity to form structured knowledge units (such as a single knowledge point or a set of operation steps).

[0202] Step 2, Audio Analysis and Deeper Text Understanding + Skill Tagging:

[0203] Audio content (such as audio explanations and meeting recordings) is converted into plain text format using a speech-to-text model. Leveraging a large language model and contextual semantic logic, homophony and punctuation errors are corrected during transcription, while text comprehension is deepened (identifying technical terms, core viewpoints, and logical connections). A pre-trained skill tag classification model (containing a hierarchical system of basic, advanced, and professional skills) is invoked to tag the text after enhanced understanding. Combined with retrieval-enhanced generation technology, the text to be tagged is semantically compared with high-quality knowledge units already tagged on the platform to assist in accurate tag matching and avoid misjudgment (e.g., mislabeling "advanced data analysis" as "basic skills").

[0204] Step 3, Generation of multi-dimensional knowledge representations:

[0205] Content Vector: Each structured knowledge unit is vectorized and encoded using a text embedding adaptive model, transforming text semantics into a high-dimensional numerical vector to capture the core content features of the knowledge unit;

[0206] Topic Vectors: The Latent Dirichlet Algorithm (LDA) topic clustering algorithm is used to classify all knowledge units into topics. At the same time, retrieval enhancement generation technology is combined to optimize the topic recognition accuracy (eliminating topic confusion and boundary ambiguity issues) and generate topic vectors representing the topics to which the knowledge units belong.

[0207] Skill Vector: Based on a hierarchical skill tagging system of "basic skills - advanced skills - professional skills", the labeled skill tags are quantified and encoded (e.g., different levels correspond to different numerical ranges) to generate skill vectors.

[0208] Step 4, Personal Knowledge Profile Construction and Permission Storage:

[0209] Based on entity recognition technology, core entities such as users' areas of expertise, skills, and topics of interest are extracted from three types of vectors: content, topic, and skills.

[0210] By using topic clustering technology to summarize users' core knowledge areas (such as "Product Manager - Requirements Document Writing" and "Data Analysis - SQL Practice"), and combining the results of skill tag classification, the user's skill level level is clarified (such as "SQL Skills - Advanced").

[0211] By fusing the three types of vectors, a comprehensive personal knowledge profile is constructed that represents a user's knowledge reserves, skill level, and interest direction, thereby achieving a deep binding between knowledge and user identity;

[0212] Fine-grained permission tags (such as globally visible, visible to followers, etc.) are assigned to each knowledge unit and its corresponding vector. After verification through the permission control mechanism, the knowledge vector with permission tags is stored in the user's personal private knowledge base.

[0213] By clarifying the implementation logic of specific technologies such as lightweight OCR, adaptive noise reduction, hierarchical skill tags, and LDA+RAG, we can further improve the accuracy of multimodal knowledge parsing, the standardization of knowledge representation, and the fit of personal knowledge profiles.

[0214] In this embodiment of the application, the method specifically includes the following steps in the process of intelligent representatives achieving accurate cross-user knowledge retrieval through large language models and retrieval enhancement generation technology:

[0215] Step 1, Building a hierarchical global knowledge index:

[0216] A platform-wide knowledge index is constructed based on a three-level structure of "personal private vector clusters - domain shared vector clusters - global general vector clusters". The index is bound to the user's personal knowledge profile (associated with S100) and knowledge unit permission tags.

[0217] Personal private vector clusters: Only user core knowledge vectors (such as in-depth knowledge of the user's area of ​​expertise and privacy-level knowledge) are retained, while redundant secondary knowledge is excluded to ensure the efficiency of core knowledge retrieval;

[0218] Domain-shared vector clusters: These clusters associate homogeneous knowledge vectors from users in the same domain (such as knowledge related to requirement document writing and prototype design, which are common to users in the "product manager" domain), forming a domain knowledge aggregation.

[0219] Global general vector cluster: Stores basic general knowledge vectors (such as general office skills and basic industry concepts), covering common knowledge needs across fields.

[0220] Step 2, Determining the response depth before cross-user queries:

[0221] The intelligent representative first extracts the social relationship between the queryer and the target user (the user who owns the knowledge) (such as ordinary users, followers, close friends), and then calls the permission level tags of the target user's knowledge file (from S100 fine-grained permission control).

[0222] Based on "social relationship + permission level", the matching response is divided into deep circles: globally visible circle (accessible to all users), follower circle (accessible only to those who follow each other), close friend circle (accessible only to core social relationships), and personal privacy circle (accessible only to knowledge owners).

[0223] Clearly define the boundaries of the queryer's permissions, retain only knowledge units within the corresponding circle as the search scope, directly exclude unauthorized knowledge, and avoid invalid searches and privacy leaks.

[0224] Step 3, Natural Language Command Interaction and Operation Response:

[0225] Users can issue operation commands to the intelligent representative through natural language, without the need for fixed formats or process switching;

[0226] Supported core operations include: retrieval (e.g., "find advanced knowledge of product prototype design"), filtering (e.g., "filter domain-shared knowledge from the past year"), association (e.g., "associate and view the case documents corresponding to this knowledge"), and saving (e.g., "save the answer content to my knowledge base").

[0227] The intelligent representative uses a large language model to parse the intent of the command, automatically calls the corresponding functional module to perform the operation, and returns the result in real time, reducing the threshold for user interaction.

[0228] Step 4, Hierarchical cross-user knowledge matching execution:

[0229] The large language model parses the user's query intent and generates a standardized query vector (with dimensions consistent with the S100 knowledge vector).

[0230] Coarse matching: Relying on retrieval enhancement generation technology, semantic similarity is calculated in the global general vector cluster to filter out basic general knowledge units related to the query intent, exclude completely irrelevant domain knowledge, and narrow the search scope;

[0231] Fine matching: Focusing on domain-shared vector clusters, it performs more refined semantic comparisons on homogeneous knowledge units within the same domain, further identifying knowledge highly relevant to the query intent;

[0232] Precise positioning: Within the target user's personal private vector cluster, combined with the permission range corresponding to the response depth, the core knowledge unit that best matches the query intent is located, ensuring that the matching results are accurate and compliant.

[0233] Step 5, High-confidence response generation and result return:

[0234] The retrieval enhancement generation technology performs contextual analysis on the knowledge units obtained by the three-level matching to verify the consistency and completeness of the knowledge;

[0235] The large language model integrates and correlates knowledge units to generate high-confidence, accurate answers with complete structure and logical coherence, while also labeling the knowledge source (e.g., "from domain-shared knowledge of user XX").

[0236] Ultimately, it returns a related summary of the knowledge vector and the response content, without returning the original knowledge unit, thus protecting the rights of the knowledge owner while meeting the usage needs of the querying party.

[0237] By clearly defining the specific division of the three-level knowledge index and the four-level response depth, and refining the natural language operation types and hierarchical matching logic, the accuracy and efficiency of cross-user knowledge retrieval are improved, knowledge security is ensured through strict permission control, and the user interaction threshold is lowered.

[0238] In this embodiment of the application, the method specifically includes the following steps in dynamically calculating the user's knowledge credibility level and driving permission adjustment based on a large language model:

[0239] Step 1, Local Reputation Sub-Score Calculation and Consistency Verification:

[0240] Each user's local node focuses on the core dimensions of knowledge contribution and independently calculates the local reputation sub-score: ① Accuracy: compares the fit between the knowledge unit uploaded by the user and the authoritative knowledge base; ② Completeness: verifies whether the core information of the knowledge unit (such as operation steps and logical deduction) is missing; ③ Timeliness: evaluates whether the knowledge unit is in line with the latest industry trends and whether it is updated in a timely manner; the three sub-indicators are summed with equal weight to obtain the local reputation sub-score.

[0241] Based on the large language model, all local reputation sub-scores are consistent: the distribution of sub-scores for knowledge contributions in the same domain and of the same type is compared, and abnormally high / low extreme values ​​are removed (such as score distortion caused by local calculation bias). At the same time, the differences in sub-scores across nodes are integrated (such as the scoring calibration of the same knowledge unit in different local nodes) to ensure the objectivity and consistency of sub-scores.

[0242] Step 2, Generation and Recording of Global Reputation Score:

[0243] The system has pre-set fixed weighting rules: knowledge contribution quality (primary weighting factor), interaction feedback (secondary weighting factor), and timeliness (supplementary weighting factor). The weighting ratio can be fine-tuned according to the platform scenario, but the priority remains unchanged.

[0244] Large language models are deeply involved in score integration:

[0245] Analyze the actual value of knowledge contributions (such as professional depth and suitability for application scenarios) and correct the biases of relying solely on quantitative indicators;

[0246] Perform semantic analysis on interactive feedback (distinguishing between genuine positive reviews, constructive questions, and malicious negative reviews) and extract quantitative values ​​of effective feedback;

[0247] By combining the local reputation sub-score after consistency verification, the quantitative value of interactive feedback, and the timeliness indicator, the user's global reputation total score is generated by weighting and summing according to preset weights, and is synchronously recorded to the user's personal knowledge profile association system.

[0248] Step 3, dynamic adjustment of user knowledge trust level triggered:

[0249] Positive adjustment trigger: When the system detects a significant increase in the number of times a user's knowledge unit is cited (such as doubling the number of citations in a short period of time or an increase in cross-domain citations), it determines that the value of the knowledge has been widely recognized and automatically triggers a positive incremental adjustment to the user's knowledge credibility level (such as upgrading from "good" to "excellent").

[0250] Negative adjustment trigger: When a user receives feedback such as knowledge inquiries or questions from other users, and the response timeout (such as exceeding the preset 24-hour response window) occurs, it is determined that the quality of the interaction service is substandard, and the user's knowledge credibility level is automatically negatively reduced (such as from "average" to "lower").

[0251] After the level is adjusted, it is synchronized in real time to the platform's permission management system and smart representatives to ensure that the latest level data is used in subsequent interaction scenarios.

[0252] Step 4, adaptive adjustment of multi-level permission control strategy:

[0253] When a user's knowledge credibility level enters a new threshold range (such as moving from "good" to "excellent"), the large language model initiates "user's current knowledge needs and permission matching assessment": analyzing the user's recent search behavior, knowledge upload direction, and interaction scenarios to determine the user's current core needs (such as whether a wider sharing scope or higher search priority is required).

[0254] Automatic permission adjustments based on evaluation results:

[0255] High-level users (e.g., "Excellent"): Increase the upper limit of "globally visible" knowledge units, open cross-domain knowledge sharing permissions, and grant priority display rights for search results;

[0256] Intermediate-level users (e.g., "Good"): Maintain basic sharing permissions, only open the "visible to followers" scope;

[0257] For low-level users (e.g., "lower"): the sharing scope is restricted to "visible to close friends", and the permission to initiate cross-user conversations is disabled; the adjusted permissions are strongly bound to the user's knowledge trust level and synchronized to the fine-grained permission tag system (associated with S100) to ensure that there are no conflicts in permission control.

[0258] By employing a two-tiered logic of "local sub-score verification + global total score calculation," clear weight priorities, specific level adjustment trigger scenarios, and a large language model to evaluate the matching degree, the calculation of user knowledge credibility level becomes more accurate, the adjustment more dynamic, and the permission control more aligned with the actual user situation. This further solves the problems of vague knowledge credibility assessment and rigid permission adjustment in existing platforms.

[0259] In this embodiment of the application, the method specifically includes the following steps in the process of transforming user social behavior into knowledge graph edge weight optimization signals and constructing a closed loop based on a reinforcement learning mechanism:

[0260] Step 1, Initialize the core elements of the reinforcement learning model:

[0261] Define the composition of the state space: integrate three key parameters, including the current user's knowledge credibility level (from the dynamic calculation results of S300), knowledge vector similarity (from the semantic comparison results of S100 multi-dimensional knowledge representation), and historical interaction frequency (the cumulative number of past knowledge interactions and social interactions between users), to comprehensively represent the current knowledge interaction and social association state.

[0262] Define the scope of the action space: The action space focuses on the increase or decrease of the edge weights of the knowledge graph, and is divided into gradients according to "significant increase, slight increase, no change, slight decrease, and significant decrease". The weight adjustment range is positively correlated with the value of social behavior and the strength of knowledge association (e.g., high-value positive social behavior corresponds to a significant increase in weight).

[0263] Define the dimensions of the reward function: The reward function includes the breadth of knowledge dissemination (the number of users covered by the interactive behavior and the scope of secondary dissemination of knowledge units), the user ability improvement coefficient (the proportion of improvement in the professionalism and completeness of knowledge uploaded by users in the future), and the positive feedback rate of interaction (the proportion of positive interaction to all effective interaction). The reward value is calculated by weighting the three dimensions according to preset weights, which drives the direction of model optimization.

[0264] Step 2, Condition Determination and Execution of Cross-User Dialogue Initiation:

[0265] The intelligent representative (associated with S200) synchronizes two core data in real time: the user's latest knowledge credibility level (from S300) and the knowledge vector similarity between the user and potential interaction objects (from S100 multi-dimensional knowledge representation comparison).

[0266] Trigger condition verification: When both "user knowledge credibility level reaches the platform's preset threshold range (e.g., 'good' or above)" and "knowledge vector similarity reaches the preset matching degree (e.g., core topic semantic similarity meets the standard)" are met, it is determined that it has high-efficiency interaction value and the dialogue initiation process is started;

[0267] Personalized dialogue invitation generation: The intelligent representative combines the knowledge relevance of both parties (such as the common interest in "product prototyping") and trust level information to generate a natural language invitation (such as "Hello, I have 'good' level trust knowledge in the area of ​​prototyping skills that you are interested in, and our knowledge match is relatively high. Would you like to exchange experiences?"), and pushes it to both users to increase their willingness to respond.

[0268] Step 3, Identification and Structured Retention of High-Value Questions and Answers:

[0269] After a cross-user conversation ends, the system initiates a high-value question-and-answer assessment process, which requires three conditions to be met simultaneously:

[0270] The Q&A content has been cited by multiple users (third-party users who are not the two parties in the dialogue retrieved and cited the Q&A);

[0271] The accuracy of knowledge tracing meets the standard (clearly tracing back to the original knowledge unit and creator corresponding to the question and answer, without any ambiguous ownership issues).

[0272] The confidence level was verified by Retrieval Enhancement Generation (RAG) technology (by comparing with the platform's knowledge base, it was confirmed that the question-and-answer logic was rigorous, the content was accurate, and there were no contradictions or errors).

[0273] Question-answer pairs that meet the criteria enter the accumulation process: Through entity linking technology, the core questions and key answers in the questions and answers are transformed into new nodes in the knowledge graph, while establishing the connection edges of "question-answer", "answer-original knowledge unit" and "answer-creator" to enrich the knowledge graph structure;

[0274] Structured format optimization: The large language model standardizes the format of the accumulated knowledge units, and unifies the core points, logical levels and skill tags to ensure that the knowledge units are consistent and readable, and facilitate subsequent retrieval and reuse (such as organizing scattered conversational answers into a structured form of "core conclusions + practical steps + precautions").

[0275] Step 4, Reinforcement learning model feedback and closed-loop iteration:

[0276] After high-value questions and answers are collected, the system calculates the reward value for this interaction (based on the three dimensions of the reward function) and feeds the reward signal back to the reinforcement learning model;

[0277] The model adjusts subsequent action strategies based on the reward value: if the reward value is high enough (such as wide knowledge dissemination and high approval rating), then a stronger weight increase signal is generated for similar social behaviors; if the reward value is low, then the action gradient is optimized (such as reducing the weight adjustment magnitude corresponding to inefficient social behaviors).

[0278] The optimized knowledge graph weights are synchronized to the platform's retrieval system, providing support for the next round of cross-user dialogue initiation and accurate knowledge matching, and promoting the continuous iteration of the closed loop of "knowledge interaction - social connection - knowledge dissemination - capability enhancement - new knowledge accumulation".

[0279] By clarifying the core elements of reinforcement learning models, the conditions for initiating cross-user dialogues, the criteria for judging high-value questions and answers, and the details of knowledge accumulation and optimization, the optimization of knowledge graphs becomes more targeted, cross-user interactions become more precise and efficient, and the accumulation of high-value knowledge becomes more standardized. This further strengthens the deep integration of social behavior and knowledge dissemination, and improves the operational efficiency of the knowledge interaction loop and the knowledge reuse rate.

[0280] In this embodiment of the application, the method for confirming the ownership and quantifying the value of knowledge contributions and usage behavior through a dual-track recording mechanism specifically includes:

[0281] Step 1, Initialization of the dual-track recording mechanism and data acquisition:

[0282] The two core modules for initializing dual-track records define their respective record fields and storage rules:

[0283] Knowledge Contribution Record: Specifically records the ownership and core attributes of knowledge units. Required fields include creator (linked to user's personal identity / business card), generation time (accurate to the second), content hash value (a unique identifier for the knowledge unit, generated by an algorithm that encrypts the content), and permission level (from S100 fine-grained permission tags, such as globally visible, visible to followers, etc.), ensuring that knowledge units are "traceable from birth".

[0284] Behavior Usage Records: Focusing on the entire knowledge unit flow chain, required fields include caller (the user's identity using the knowledge), call time (accurate to the second), call scenario (knowledge payment / representative subscription / enterprise service), and payment amount (0 is recorded if no payment is made), comprehensively capturing knowledge usage behavior data;

[0285] Real-time data collection: Knowledge contribution record data is synchronously accessed from the S100 knowledge unit generation stage, and behavior usage record data is captured in real time from S200 retrieval, S400 interaction and other scenarios to ensure that no data is missed.

[0286] Step 2, Dual-track record synchronization, evidence storage, and association binding:

[0287] A timestamp synchronization mechanism is adopted to give each knowledge contribution record and the corresponding behavior usage record the same timestamp (based on the knowledge contribution generation time, the timestamps of subsequent usage behaviors are all associated with this time), so as to achieve real-time alignment between the two;

[0288] Using the "content hash value" of a knowledge unit as the core association key, a one-to-one correspondence between "knowledge contribution record - behavior usage record" is established, forming a complete link between "ownership source - usage flow", avoiding the risk of data fragmentation or tampering (the content hash value changes uniquely with the content of the knowledge unit, and the hash value changes synchronously after tampering, which can be quickly identified).

[0289] Step 3, Execution of Knowledge Confirmation Operations (Dispute Scenario):

[0290] When a dispute arises regarding intellectual property ownership (such as multiple parties claiming to be the creators of the same knowledge unit), the ownership confirmation process is initiated:

[0291] Extract the content hash value of the disputed knowledge unit, and retrieve the corresponding knowledge contribution record from the dual-track record using this unique identifier;

[0292] By comparing the "Creator" field (associated with user identity / business card) and the "Generation Timestamp" (with the earliest generation time taking precedence) in the knowledge contribution record, and combining the binding relationship between knowledge units and personal knowledge profiles in S100, a complete chain of evidence of "Creator Identity - Generation Time - Content Uniqueness" is formed.

[0293] The legitimate rights holder is determined based on the chain of evidence, and the confirmation result is simultaneously updated to the dual-track record, serving as the sole basis for subsequent value distribution and handling of rights disputes.

[0294] Step 4, Three-dimensional quantitative assessment of the value of knowledge units:

[0295] Originality assessment: Based on the large language model, the knowledge unit to be evaluated is compared with the platform's existing knowledge base (including all user private and shared knowledge bases) in full semantic similarity. The originality is determined by calculating the difference between the two - the higher the difference (i.e. the lower the overlap with existing knowledge), the higher the originality score.

[0296] Scarcity assessment: Analyze the proportion of similar knowledge units in the platform (e.g., the total number of knowledge units in the sub-field of "artificial intelligence-deep learning-model tuning"), and combine this with the search frequency and interaction volume of the field to analyze the demand. The lower the proportion of existing knowledge units and the higher the demand, the higher the scarcity level.

[0297] Practicality assessment: Collect user feedback data for this knowledge unit (such as positive / negative reviews, follow-up consultations, and case studies of successful applications), perform semantic analysis on the feedback content using a large language model, extract information such as the percentage of effective positive feedback and descriptions of application value, and convert it into a quantitative practicality score;

[0298] The scores for originality, scarcity, and practicality are weighted and summed according to preset weights to generate the final value quantification score of the knowledge unit, which is then synchronously linked to the dual-track record.

[0299] By clarifying the specific fields of the dual-track record, the synchronous evidence storage method and the logic of the rights confirmation operation, as well as the specific judgment method of the three-dimensional quantification of knowledge value, the knowledge rights confirmation is more accurate and efficient, and the value assessment is more scientific and transparent. This provides an objective and traceable core basis for the subsequent revenue distribution based on the Shapley value algorithm, and effectively solves the problems of difficult definition of knowledge ownership and ambiguous value assessment on existing platforms.

[0300] In this embodiment of the application, the method for allocating knowledge asset revenue and generating credible settlement evidence based on the Shapley value algorithm specifically includes:

[0301] Step 1, Shapley value algorithm allocation rules initialization:

[0302] The system pre-defines the core allocation rules for three major scenarios, clearly defining the allocation factors, weight logic, and adjustment basis for each scenario, and linking them to the data sources of preceding processes.

[0303] Basic data synchronization: retrieve knowledge unit value quantification scores (from knowledge value three-dimensional assessment), forwarding and citation data, intelligent representative response effect data (from S200), enterprise call records and customized information from the dual-track records as the core input for allocation calculation;

[0304] Unified rules and benchmarks: The Shapley value algorithm is based on the "marginal contribution" principle, that is, the revenue is distributed according to the actual contribution of each participant (creator, disseminator, platform) in the process of realizing the value of knowledge, so as to ensure the fairness of distribution.

[0305] Step 2, Calculate revenue distribution by scenario:

[0306] Knowledge payment scenarios:

[0307] Allocation factor weight settings:

[0308] Knowledge creator contribution weight: positively correlated with the quantitative score of knowledge unit value (the higher the score, the higher the weight percentage, such as a full score knowledge unit corresponds to 60% weight, and a low score corresponds to 30% weight).

[0309] Knowledge disseminator promotion weight: calculated based on the actual paid conversion effect brought about by the disseminator's forwarding and citation (e.g., if 10 paid orders are generated after dissemination, 15% weight will be allocated according to the preset ratio).

[0310] Platform technical service weight: a fixed percentage (e.g., 25%), covering technical support and platform operating costs;

[0311] The Shapley value algorithm calculates the revenue of each participant: total revenue × corresponding weight, resulting in the specific revenue for creators, distributors, and the platform.

[0312] Representative subscription scenario:

[0313] The basic allocation factor follows the rules of the knowledge payment scenario, and a new correction factor is added: the accuracy of the intelligent representative's response (from S200 response effect data, such as an accuracy rate of over 95% is considered satisfactory).

[0314] Weighting adjustment logic: When the accuracy of the response reaches the platform's preset standard (e.g., accuracy rate ≥ 90%), the knowledge creator's revenue weight will be automatically increased (e.g., increased by 5%-10%), while the platform weight will be simultaneously decreased (keeping the total weight at 100%), incentivizing creators to provide high-quality knowledge to improve the intelligent representative's response effect;

[0315] The algorithm calculates the final revenue: the total subscription revenue is allocated according to the adjusted weights.

[0316] Enterprise service settlement scenarios:

[0317] The allocation coefficient is adjusted based on: enterprise user usage (frequency of use, duration of use, scope of cross-departmental use); and the degree of customization of the knowledge unit (e.g., whether it is developed exclusively for the enterprise, whether it is adapted to the enterprise's specific business scenarios).

[0318] Adjustment rules: The higher the frequency of calls and the deeper the degree of customization, the higher the contribution weight of the knowledge creator (e.g., the creator of the exclusive customized knowledge unit can have a weight of up to 70%). The weight of the disseminator is flexibly set according to the internal forwarding and citation situation of the enterprise, and the platform weight is adjusted according to a fixed proportion (e.g., 20%).

[0319] The algorithm calculates the distribution amount of the total revenue of enterprise services, and clarifies the revenue share of creators and platforms (including disseminators if applicable).

[0320] Step 3, Settlement Details List Generation and Process Triggering:

[0321] All scenarios rely on a large language model to help generate a detailed settlement list: the list includes core information - basic information of knowledge unit (name, content hash value, value quantification score), revenue distribution details (identity, weight, and revenue amount of each participant), and scenario-related data (number of paid orders, subscription period, and enterprise call records), ensuring that each revenue is traceable and verifiable;

[0322] Automatically trigger the settlement process: Generate corresponding settlement vouchers based on the scenario type (knowledge payment order receipt, representative subscription monthly statement, enterprise service quarterly statement), and store the settlement vouchers in a dual-track record simultaneously, ensuring immutability through timestamps and content hash values;

[0323] Data synchronization and query: Settlement data is synchronized to the platform's financial system and users' personal accounts in real time. Users can query details through the smart representative (linked to S200), and enterprise users can generate standardized financial reconciliation reports.

[0324] Step 4, Settlement data archiving and dispute resolution:

[0325] Regularly archive settlement details, vouchers, and allocation calculation logs, and store them in categories according to scenario type and time period to retain complete data on the "value quantification - revenue distribution - settlement" chain.

[0326] In the event of a dispute over profit distribution, archived data and dual-track records are retrieved, and information such as distribution factors, weight settings, and adjustment basis is checked to quickly verify the facts and ensure efficient resolution of the dispute.

[0327] By clarifying the revenue distribution factors, weighting rules, and adjustment logic for the three major scenarios, and combining them with a large language model to generate standardized settlement details, the distribution of knowledge asset revenue becomes fairer, more accurate, and more transparent, and the settlement basis becomes more credible and traceable. This effectively solves the problems of vague revenue distribution rules, poor scenario adaptability, and lack of credible settlement basis in existing platforms.

[0328] This application discloses a system for multi-user knowledge interconnection and intelligent representative interaction, referring to... Figure 2 ,include:

[0329] The Multimodal Knowledge Parsing and Profile Building Module 001, relying on a large language model and retrieval-enhanced generation technology, performs real-time and efficient parsing of user-provided documents, audio, and video content at the multimodal knowledge access end. It extracts structured knowledge units through the collaborative extraction of optical character recognition models, speech-to-text models, and large language models. Combined with retrieval-enhanced generation technology, it achieves accurate association and deduplication of knowledge units, generating multi-dimensional knowledge representations including content vectors, topic vectors, and skill vectors. The multi-dimensional knowledge representation establishes a personal knowledge profile based on entity recognition, topic clustering, and skill tag classification technologies, and stores the knowledge vectors in a personal private knowledge base through a fine-grained access control mechanism.

[0330] The intelligent representative and cross-user retrieval module 002 are built based on a large language model and retrieval enhancement generation technology. The intelligent representative locates relevant knowledge units through deep matching of knowledge vectors and query intent, realizes accurate cross-user knowledge retrieval, generates high-confidence accurate answers in question-and-answer scenarios, and controls knowledge access permissions based on user authorization scope.

[0331] The reputation calculation and access control module 003, in the process of multi-user knowledge interconnection, relies on the large language model to analyze the quality of knowledge contribution, the timeliness of interactive feedback and other parameters, dynamically calculates the user knowledge credibility level, updates the user knowledge credibility level in real time and drives the adaptive adjustment of multi-level access control strategies to achieve differentiated knowledge sharing.

[0332] The social linkage knowledge closed loop module 004, based on the reinforcement learning mechanism, transforms users' social behavior into knowledge graph edge weight optimization signals, and constructs a closed loop of knowledge interaction - social connection - knowledge dissemination - capability enhancement. Among them, the intelligent representative actively initiates cross-user dialogue based on the user's knowledge credibility level and knowledge vector similarity, verifies the confidence of question and answer through retrieval enhancement generation technology, and automatically precipitates high-value question and answer into searchable knowledge units.

[0333] The Knowledge Ownership Confirmation, Quantification, and Revenue Module 005 uses a dual-track recording mechanism to confirm the ownership and value of knowledge contributions and usage behaviors. It relies on a large language model to assist in evaluating the value of knowledge, allocates knowledge asset revenue based on the Shapley value algorithm, and dynamically generates credible evidence for knowledge payment, representative subscription, and enterprise service settlement.

[0334] This application also discloses an electronic device, including a processor, wherein the processor runs a program for the method of multi-user knowledge interconnection and intelligent representative interaction described in any one of the above-described embodiments.

[0335] This application also discloses a storage medium storing a program for the method of multi-user knowledge interconnection and intelligent representative interaction as described in any one of the above embodiments.

[0336] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for multi-user knowledge interconnection and intelligent representative interaction, characterized in that, include: Leveraging large language models and retrieval-enhanced generation technology, the system performs real-time and efficient parsing of user-provided documents, audio, and video content at the multimodal knowledge access end. It extracts structured knowledge units through the collaborative efforts of optical character recognition (OCR), speech-to-text (STO) models, and large language models. Combined with retrieval-enhanced generation technology, it achieves accurate association and deduplication of knowledge units, generating multi-dimensional knowledge representations including content vectors, topic vectors, and skill vectors. These multi-dimensional knowledge representations establish personal knowledge profiles based on entity recognition, topic clustering, and skill tag classification technologies, and the knowledge vectors are stored in a personal private knowledge base through a fine-grained access control mechanism. Intelligent representatives are constructed based on large language models and retrieval enhancement generation technology. These intelligent representatives locate relevant knowledge units through deep matching of knowledge vectors and query intent, enabling accurate cross-user knowledge retrieval and generating high-confidence, accurate answers in question-and-answer scenarios. At the same time, they control knowledge access permissions based on user authorization scope. In the process of multi-user knowledge interconnection, the quality of knowledge contribution, the timeliness of interactive feedback and other parameters are analyzed by relying on a large language model. The credibility level of user knowledge is dynamically calculated. The credibility level of user knowledge is updated in real time and drives the adaptive adjustment of multi-level permission control strategy to achieve differentiated knowledge sharing. Based on the reinforcement learning mechanism, user social behavior is transformed into knowledge graph edge weight optimization signals, and a closed loop of knowledge interaction, social connection, knowledge dissemination and capability improvement is constructed. Among them, the intelligent representative actively initiates cross-user dialogue based on the user's knowledge credibility level and knowledge vector similarity, verifies the confidence of question and answer through retrieval enhancement generation technology, and automatically precipitates high-value question and answer into searchable knowledge units. By establishing ownership and quantifying the value of knowledge contributions and usage through a dual-track recording mechanism, relying on a large language model to assist in evaluating knowledge value, allocating knowledge asset revenue based on the Shapley value algorithm, and dynamically generating credible evidence for knowledge payment, representative subscriptions, and enterprise service settlements.

2. The method for multi-user knowledge interconnection and intelligent representative interaction according to claim 1, characterized in that, In the process of real-time parsing of user-provided documents, audio, and video content and generating multi-dimensional knowledge representations to construct a personal knowledge profile at the multimodal knowledge access end, the methods also include: A lightweight optical character recognition model is used to perform layout analysis and character recognition on document images. For blurry documents, an adaptive denoising algorithm is first used to optimize the image quality, and then the structured knowledge units of the text are extracted from the optimized image. Audio content is converted into text format through a speech-to-text model. Relying on a large language model and contextual semantics, transcription errors are corrected and the understanding of text content is deepened. The text understanding results are labeled using a pre-trained skill tag classification model, which includes a hierarchical tag system of basic skills, advanced skills, and professional skills. The model is combined with retrieval enhancement generation technology to assist in accurate tag matching. In the process of generating multi-dimensional knowledge representation, content vectors are generated through a text embedding adaptive model, topic vectors are generated through a latent Dirichlet distribution topic clustering algorithm combined with retrieval enhancement generation technology, and skill vectors are quantized based on a hierarchical skill tag system of basic skills, advanced skills, and professional skills. Based on entity recognition technology, topic clustering technology and skill tag classification technology, multi-dimensional knowledge representation is transformed into a personal knowledge profile. Through a fine-grained access control mechanism, the encoded knowledge vector is stored in the personal private knowledge base.

3. The method for multi-user knowledge interconnection and intelligent representative interaction according to claim 2, characterized in that, In the process of intelligent representatives achieving accurate cross-user knowledge retrieval through large language models and retrieval enhancement generation techniques, the methods also include: A global knowledge index is constructed for user knowledge vectors in a hierarchical manner, consisting of personal private vector clusters, domain-shared vector clusters, and global general vector clusters. Among them, personal private vector clusters only retain the user's core knowledge vectors, domain-shared vector clusters associate homogeneous knowledge vectors of users in the same domain, and global general vector clusters store basic general knowledge vectors. Before initiating a cross-user knowledge query, the intelligent representative first determines the response depth based on the social relationship between users and the permission level of the knowledge file. The response depth includes the globally visible circle, the follower circle, the close friend circle, and the personal privacy circle. Only the response content within the corresponding circle's permission is returned to the querying party. Users can issue commands to the intelligent representative using natural language, access and view content within the platform that they have access to, and perform various operations such as content retrieval, filtering, association, and saving using natural language. When initiating a cross-user knowledge query, the system first parses the user's query intent using a large language model to generate a query vector. Then, relying on retrieval enhancement generation technology, it performs a coarse matching operation between the knowledge vector and the query intent in the global general vector cluster. Next, it performs a precise matching operation in the domain shared vector cluster. Finally, it performs a precise positioning operation in the target user's personal private vector cluster. In this process, all matching processes combine contextual association analysis with retrieval enhancement generation technology. The large language model integrates the matching results to generate high-confidence accurate answers. The query results return the associated summary of knowledge vectors and the corresponding high-confidence response content, without returning the original knowledge units.

4. The method for multi-user knowledge interconnection and intelligent representative interaction according to claim 3, characterized in that, In the process of dynamically calculating the user's knowledge credibility level and driving permission adjustment based on a large language model, the method also includes: Each user's local node calculates its local reputation sub-score based on the accuracy, completeness, and timeliness of its knowledge contribution, and uses a large language model to perform consistency verification and integration of the local reputation sub-scores. The system pre-sets reputation calculation rules, in which the quality of knowledge contribution is the primary weighting factor, interactive feedback is the secondary weighting factor, and timeliness is the supplementary weighting factor. It relies on a large language model to analyze the actual value of knowledge contribution and the true intention of interactive feedback, and generates and records the total reputation score. When the number of citations of a user's knowledge contribution increases significantly, the system triggers a positive incremental adjustment to the user's knowledge credibility level. When a user's feedback response times out, the system triggers a negative deduction adjustment to the user's knowledge credibility level. In particular, whenever the user's knowledge credibility level enters a new threshold range, the large language model evaluates the user's current knowledge needs and permission matching degree, and automatically triggers the adaptive adjustment of the corresponding multi-level permission control strategy.

5. The method for multi-user knowledge interconnection and intelligent representative interaction according to claim 4, characterized in that, In the process of transforming user social behavior into knowledge graph edge weight optimization signals and constructing a closed loop based on reinforcement learning mechanisms, the method also includes: The state space of reinforcement learning includes the current user's knowledge credibility level, knowledge vector similarity, and historical interaction frequency; the action space is the increase or decrease of the edge weights of the knowledge graph; and the reward function includes three dimensions: knowledge dissemination breadth, user ability improvement coefficient, and interaction feedback satisfaction rate. When the intelligent representative detects that the user's knowledge credibility level has reached a preset threshold range and the knowledge vector similarity has reached a preset matching degree, it automatically initiates a cross-user dialogue. The criteria for determining high-value question-answering are that the question-answering content is cited by multiple users, the accuracy of knowledge tracing reaches the preset standard, and the confidence level is verified by retrieval enhancement generation technology. Question-answering pairs that meet the criteria are automatically precipitated into new nodes and edges of the knowledge graph through entity linking technology, and the structured storage format of knowledge units is optimized by a large language model.

6. The method for multi-user knowledge interconnection and intelligent representative interaction according to claim 5, characterized in that, In the process of establishing ownership and quantifying the value of knowledge contributions and usage through a dual-track recording mechanism, the methods also include: The dual-track recording mechanism includes knowledge contribution records and behavior usage records. Knowledge contribution records record the creator, generation time, content hash value, and permission level of the knowledge unit, while behavior usage records the caller, call time, call scenario, and payment amount of the knowledge unit. The dual-track record achieves synchronous evidence storage through a timestamp synchronization mechanism. When a dispute over intellectual property rights occurs, the ownership confirmation operation is completed by comparing the content hash value and timestamp of the dual-track record. The value quantification dimensions of a knowledge unit include originality, scarcity, and practicality. Originality is determined by analyzing the differences between the knowledge unit and the existing knowledge base using a large language model. Scarcity is determined by the proportion of similar knowledge units in the existing stock. Practicality is determined by semantic analysis of user feedback after invocation.

7. The method for multi-user knowledge interconnection and intelligent representative interaction according to claim 6, characterized in that, In the process of allocating knowledge asset revenue and generating credible settlement evidence based on the Shapley value algorithm, the method also includes: For knowledge payment scenarios, the revenue distribution factors of the Shapley value algorithm include the contribution weight of knowledge creators, the promotion weight of knowledge disseminators, and the platform's technical service weight. Among them, the creator weight is set according to the value quantification score of the knowledge unit, the disseminator weight is set according to the forwarding and citation situation, and the platform weight is a fixed percentage. For the subscription scenario, the accuracy of the smart representative's response will be included in the correction factor for the Shapley value calculation. When the response accuracy reaches the preset standard, the creator's revenue weight will be adjusted upward. For enterprise service settlement scenarios, the Shapley value allocation coefficient is adjusted based on the enterprise user's call status and the degree of customization of the knowledge unit; in all scenarios, the settlement operation relies on the large language model to generate a detailed list, automatically trigger the settlement process and record it, and dynamically generate credible evidence for knowledge payment, representative subscription and enterprise service settlement.

8. A system for multi-user knowledge interconnection and intelligent representative interaction, characterized in that, include: The multimodal knowledge parsing and profile building module, relying on a large language model and retrieval-enhanced generation technology, performs real-time and efficient parsing of user-provided documents, audio, and video content at the multimodal knowledge access end. It extracts structured knowledge units through the collaborative extraction of optical character recognition models, speech-to-text models, and large language models, and combines retrieval-enhanced generation technology to achieve accurate association and deduplication of knowledge units, generating a multi-dimensional knowledge representation including content vectors, topic vectors, and skill vectors. The multi-dimensional knowledge representation establishes a personal knowledge profile based on entity recognition, topic clustering, and skill tag classification technologies, and stores the knowledge vectors in a personal private knowledge base through a fine-grained access control mechanism. The intelligent representative and cross-user retrieval module are constructed based on a large language model and retrieval enhancement generation technology. The intelligent representative locates relevant knowledge units through deep matching of knowledge vectors and query intent, realizes accurate cross-user knowledge retrieval, generates high-confidence accurate answers in question-and-answer scenarios, and controls knowledge access permissions based on user authorization scope. The reputation calculation and access control module, in the process of multi-user knowledge interconnection, relies on a large language model to analyze the quality of knowledge contribution, the timeliness of interactive feedback and other parameters, and dynamically calculates the user knowledge credibility level. The user knowledge credibility level is updated in real time and drives the adaptive adjustment of multi-level access control strategies to achieve differentiated knowledge sharing. The social-linked knowledge closed-loop module transforms user social behavior into knowledge graph edge weight optimization signals based on reinforcement learning mechanisms, constructing a closed loop of knowledge interaction, social connection, knowledge dissemination, and capability enhancement. Among them, the intelligent representative actively initiates cross-user dialogues based on the user's knowledge credibility level and knowledge vector similarity, verifies the confidence of question and answer through retrieval enhancement generation technology, and automatically precipitates high-value question and answer into searchable knowledge units. The Knowledge Ownership Confirmation and Revenue Module uses a dual-track recording mechanism to confirm the ownership and value of knowledge contributions and usage behaviors. It relies on a large language model to assist in evaluating the value of knowledge, allocates knowledge asset revenue based on the Shapley value algorithm, and dynamically generates credible evidence for knowledge payment, representative subscription, and enterprise service settlement.

9. An electronic device, characterized in that, Includes a processor, wherein the processor runs a program for a method for multi-user knowledge interconnection and intelligent representative interaction as described in any one of claims 1-7.

10. A storage medium, characterized in that, The program stores a method for multi-user knowledge interconnection and intelligent representative interaction as described in any one of claims 1-7.

Images