Electronic component AI technical support system and method based on enterprise private knowledge base

Abstract

The invention relates to the technical field of intelligent question answering, in particular to an electronic component AI technical support system and method based on an enterprise private knowledge base, and the method comprises the following steps: receiving enterprise private technical data or information from at least one data source, the data source is any one or more of local upload, FTP / SFTP, enterprise mail, ERP / PLM system or third-party storage interface; performing text extraction and optical character recognition, document structure analysis, table unit extraction and metadata annotation on the received data, segmenting the data into searchable semantic units and standardizing terms; generating a vector representation for the semantic unit, storing the vector representation in a vector database, extracting entities and relationships, and updating an enterprise private knowledge graph; when a user query is received. According to the method, repeated labor investment is remarkably reduced, the fault positioning and response time is shortened, and the long-term utilization value of knowledge assets is improved.

Description

Technical Field

[0001] This invention relates to an AI technology support system and method for electronic components based on an enterprise's private knowledge base, belonging to the field of intelligent question answering technology. Background Technology

[0002] Electronic component and terminal smart manufacturing companies generally face a large amount of scattered, heterogeneous and frequently iterated proprietary data—including specifications, application notes, SOPs, test reports and maintenance work orders, etc., and these data contain both tabular parameters and a large amount of unstructured text and scanned images. Traditional search relies mainly on keyword or full-text matching, which makes it difficult to understand the semantics of table cells and synonyms of terms. Furthermore, generative question answering is prone to creating "illusions" in vertical professional fields, and context breaks often occur in multi-channel conversations. At the same time, the security and compliance requirements of enterprises for private data restrict the direct use of public models and external platforms, resulting in significant deficiencies in the accuracy, traceability and compliance of existing solutions. This leads to several challenges: how to efficiently parse multi-source, multi-format private data into searchable semantic units while maintaining parameter and unit consistency; how to establish a reliable connection between semantic retrieval and generative models to ensure that answers are supported by evidence and are traceable; how to achieve secure strategic decision-making between automation and human intervention through confidence levels and high-risk rules; how to construct a closed-loop active learning system that standardizes manual confirmation and work order results and writes them back as training samples; and how to quickly complete cold start and domain model adaptation while ensuring compliance constraints such as private deployment, encryption, and access control. Therefore, there is an urgent need to improve an AI technology support system and method for electronic components based on enterprise private knowledge bases to solve the above-mentioned problems. Summary of the Invention

[0003] The purpose of this invention is to provide an AI technology support system and method for electronic components based on an enterprise's private knowledge base. This system addresses several key issues: how to efficiently parse multi-source, multi-format private data into searchable semantic units while maintaining parameter and unit consistency; how to establish a reliable connection between semantic retrieval and generation models to ensure evidence support and traceability of responses; how to achieve secure strategic decision-making between automation and human intervention through confidence levels and high-risk rules; how to construct a closed-loop active learning system that standardizes manual confirmation and work order results as training samples; and how to quickly complete cold start and domain-specific model adaptation while ensuring compliance with private deployment, encryption, and access control requirements.

[0004] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide an AI technology support system for electronic components based on an enterprise's private knowledge base, comprising the following steps: S1. Data Acquisition: Receive enterprise private technical data or information from at least one data source, wherein the data source is any one or more of local upload, FTP / SFTP, enterprise email, ERP / PLM system or third-party storage interface; S2. Preprocessing and structuring: Perform text extraction and optical character recognition (OCR), document structure parsing, table cell extraction and metadata annotation on the data received in step S1, segment the data into searchable semantic units and standardize the terms; S3. Knowledge Representation: Generate vector representations for the semantic units and store them in a vector database. Simultaneously, extract entities and relations and update the enterprise's private knowledge graph. The knowledge graph construction includes: entity recognition and synonym normalization based on the domain terminology, extraction of entity types such as device model, pin definition, performance parameters, and fault codes, and identification of causal, substitution, and subordinate relationships between entities, supporting graph-based reasoning and retrieval; S4. Query processing: When a user query is received, the query is vectorized and a set of candidate evidence fragments is retrieved based on a hybrid retrieval strategy of vector retrieval and inverted index. The hybrid retrieval strategy includes: firstly retrieving the top K semantic units in the vector database using KNN, then performing fine ranking on the above K results based on BM25 or inverted index, and finally producing the final top-N evidence fragments by integrating and ranking them according to semantic similarity, BM25 score, document credibility, time weight and historical click rate through a learning-to-rank model. S5. Evidence-based generation and credibility: Using the candidate evidence fragments obtained in step S4 as the retrieval context, a response is generated through the Retrieval Enhancement Generation (RAG) strategy, and the response is accompanied by the corresponding evidence fragment identifier and location information; Further measures include explicitly limiting the response to "only using the given evidence" in the generated model prompts, and performing unit and range verification when the generated results involve numerical values ​​or parameters; if the verification fails, the response will be marked as low confidence and trigger manual intervention or return a "manual confirmation required" response to the user. S6. Confidence Assessment and Decision: A confidence assessment is performed on the answer. The confidence assessment is based on a weighted fusion of retrieval similarity, generation model confidence score, rule assertion verification, and source credibility. When the confidence score is lower than a preset threshold or a high-risk rule is triggered, the session is upgraded to manual takeover according to a preset strategy. The manual takeover process is as follows: the conversation context, the matched evidence fragments, the confidence score and the recommended operation set are pushed to the manual collaboration panel. The panel allows manual FAEs to edit the answers and write the standardized solutions after manual confirmation back to the knowledge base as new knowledge entries and as labeled samples for subsequent model fine-tuning. S7. Feedback and Closed-Loop Learning: Collect feedback information such as user confirmation, rating, work order results and manual annotation, and use the annotated samples to update the knowledge base index or as fine-tuning data to enter the periodic model / retrieval strategy retraining process to form a closed-loop iteration. Further, the collected feedback samples are stratified and screened according to error type and confidence level. Samples with low confidence and labeled as errors by users or humans are prioritized and imported into the active learning pool. After manual labeling, the retrieval strategy or model is fine-tuned periodically in a micro-batch manner. The system performs data security and compliance processing on incoming or queried data, including sensitive information identification and on-demand desensitization, static / transmit encryption (e.g., AES and TLS), role-based access control (RBAC), and audit logs. The system is deployed in a hybrid cloud manner, either privately or via dedicated line access, to meet enterprise compliance requirements.

[0005] Secondly, embodiments of the present invention provide an AI technology support method for electronic components based on an enterprise's private knowledge base, including a data acquisition module for acquiring enterprise private technical data from local uploads, FTP / SFTP, enterprise emails, ERP / PLM, or external storage interfaces; The preprocessing module, connected to the data acquisition module, is used to perform OCR, document structure parsing, table cell extraction and terminology standardization on the acquired data and output semantic units and metadata. Vectorization service and vector database are used to generate vector representations of the semantic units and store them for semantic retrieval. A knowledge graph engine is used to extract entities and relationships from the semantic units and maintain the enterprise's private knowledge graph. The retrieval and reordering module, connected to the vector database and inverted index, is used to perform vector retrieval, BM25 / inverted index retrieval, and reordering based on fusion features to output a set of candidate evidence. The generation and credibility module is used to generate answers based on the candidate evidence obtained from the retrieval through Retrieval Enhancement Generation (RAG), and to inject evidence citations and confidence scores into the answers; The strategy engine is used to determine whether to respond automatically or escalate the session to manual takeover based on the confidence score and preset rules. The multi-channel access and session management module supports multiple access channels such as WeChat for Enterprise, WeChat Official Accounts, web customer service, ticketing system and email, and maintains multi-turn session contexts; The human collaboration panel displays the conversation context, candidate evidence, and confidence level, allowing human intervention, editing of responses, and writing human confirmation results back to the knowledge base; The feedback and learning module is used to collect user evaluations, work order results and manual annotations, and is responsible for importing the annotated samples into the active learning or fine-tuning process to update the model or retrieval strategy. The security and operations module is used to implement static and transmission encryption of data, identification and desensitization of sensitive information, role-based access control and audit logs, and supports private or hybrid cloud deployment; The system is configured to return the answer to a user's query and simultaneously provide the corresponding evidence fragment identifier and location information.

[0006] Implementing the embodiments of this application has the following beneficial effects: This system organically integrates data collection and preprocessing, semantic unitization, vectorized retrieval and BM25 ranking, knowledge graph construction, RAG evidence injection and consistency verification, a confidence-weighted and high-risk trigger strategy engine, and a human collaboration panel and feedback learning module. This not only improves the evidence-level retrieval recall and question-answering credibility, but also ensures security and compliance in low-confidence or high-risk scenarios through human-machine collaboration. Furthermore, it achieves continuous optimization of the model and retrieval strategy through closed-loop write-back and active learning. Finally, it meets enterprise compliance requirements with private deployment, encryption, and RBAC-based operational safeguards, thereby significantly reducing repetitive manual input, shortening fault location and response time, and enhancing the long-term utilization value of knowledge assets. Attached Figure Description

[0007] Figure 1 This is a flowchart of an AI technology support method for electronic components based on an enterprise's private knowledge base, according to the present invention. Figure 2 This is a system block diagram of an AI technology support system for electronic components based on a private enterprise knowledge base, according to the present invention. Detailed Implementation

[0008] The following will describe in detail the implementation of this application with reference to the accompanying drawings and embodiments, so that the implementation process of how this application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0009] like Figure 2As shown in the figure, this embodiment provides an AI technology support system for electronic components based on an enterprise's private knowledge base, including a data acquisition module for acquiring enterprise private technical data from local uploads, FTP / SFTP, enterprise emails, ERP / PLM, or external storage interfaces. The preprocessing module, connected to the data acquisition module, is used to perform OCR, document structure parsing, table cell extraction and terminology standardization on the acquired data, and output semantic units and metadata. Vectorization services and vector databases are used to generate and store vector representations of semantic units for semantic retrieval. A knowledge graph engine is used to extract entities and relationships from semantic units and maintain a private knowledge graph for an enterprise. The retrieval and reordering module, connected to the vector database and inverted index, is used to perform vector retrieval, BM25 / inverted index retrieval, and reordering based on fused features to output a set of candidate evidence. The generation and credibility module is used to generate answers based on the candidate evidence obtained from the retrieval through Retrieval Enhancement Generation (RAG), and to inject evidence citations and confidence scores into the answers; The strategy engine is used to determine whether to respond automatically or escalate the conversation to manual intervention based on confidence scores and preset rules. The multi-channel access and session management module supports multiple access channels such as WeChat for Enterprise, WeChat Official Accounts, web customer service, ticketing system and email, and maintains multi-turn session contexts; The human collaboration panel displays the conversation context, candidate evidence, and confidence level, allowing human intervention, editing of responses, and writing human confirmation results back to the knowledge base; The feedback and learning module is used to collect user evaluations, work order results and manual annotations, and is responsible for importing the annotated samples into the active learning or fine-tuning process to update the model or retrieval strategy. The security and operations module is used to implement static and transmission encryption of data, identification and desensitization of sensitive information, role-based access control and audit logs, and supports private or hybrid cloud deployment; The system is configured to return an answer along with the corresponding evidence fragment identifier and location information when generating an answer to a user's query. The system organically integrates functional modules such as data collection, preprocessing, vectorized retrieval, knowledge graph, RAG generation, confidence assessment and strategy engine, multi-channel conversation management, human collaboration and closed-loop learning, which significantly improves the retrieval of enterprise proprietary technical knowledge and the reliability and traceability of question and answer results. Vectorization services and retrieval / re-ranking mechanisms can quickly locate highly relevant evidence at the semantic level. The generation and credibility modules enforce evidence injection and consistency verification when outputting answers, thereby effectively suppressing the illusion of generative models and reducing misleading suggestions. The strategy engine and human collaboration panel realize human-machine collaboration, which can promptly upgrade to human handling in low confidence or high risk situations, ensuring operational safety and business compliance. Multi-channel access and session management ensure consistent response and contextual continuity across touchpoints. The feedback and learning modules systematically write back manual confirmations and work order results as training samples, forming a closed-loop proactive learning process. This allows the model and retrieval strategies to continuously optimize over time and adapt to the enterprise's specific domain semantics. The security and operations modules meet private deployment and compliance requirements through static / transmission encryption, data masking, and access control, reducing information leakage and compliance risks. In summary, this system can reduce repetitive manual input, shorten fault location and response time, improve customer satisfaction, and enhance the utilization and value of the enterprise's knowledge assets in the long term.

[0010] like Figure 1-2 As shown in this embodiment, an AI technology support method for electronic components based on an enterprise's private knowledge base is applied in the vertical industries of electronic components and intelligent manufacturing of terminals: Example 1 (Vertical Industry of Electronic Components) 1. Data sources: Component datasheets, application notes, production / test reports, FAE work orders, FAQs.

[0011] Document types: Includes PDF (including scanned documents), DOCX, PPT, XLS spreadsheets, images, etc.

[0012] Initial size: Approximately 5,000 documents and a glossary (approximately 8,000 entries).

[0013] 2. Standardized knowledge processing workflow: 21. Data Acquisition: Batch pull documents from enterprise FTP / PLM / internal shared drives; register a unique identifier, version number and source information for each file; perform security scanning and hash storage on uploaded files.

[0014] Files with high sensitivity levels are marked as "restricted" to restrict access during subsequent searches and exports.

[0015] 22. Preprocessing and structuring: Perform OCR on scanned documents and images, and record the OCR confidence level for subsequent quality control.

[0016] The specification table is parsed into "table cells" and an entry format (including rows, columns, field names and values, and units) is generated.

[0017] Domain NER (rules + models) is used to extract entities (model, parameters, pins, failure codes, etc.) and synonyms are normalized (e.g., "ESR" and "equivalent series resistance" are normalized to the same entity).

[0018] 23. Knowledge Representation: Generate a vector representation for each semantic unit (paragraph or table unit) and write it into the vector database along with the document positioning information (page number, character offset).

[0019] Construct knowledge graph nodes and establish relationships, such as "device-parameter", "fault-root cause", and "device-alternative". Graph nodes retain source document identifiers and timestamps.

[0020] 3. Query and Retrieval 31. Query processing flow: The user raises a question (natural language), the system extracts slots (such as model, parameter conditions, usage scenarios) and generates a query vector.

[0021] The retrieval employs a hybrid strategy: first, vector retrieval is performed on the Top-K (e.g., K=50), then the candidates are supplemented and re-ranked using BM25 inverted index retrieval. The re-ranking uses fused features (semantic similarity, BM25 score, document credibility, timeliness, click-through rate) to generate the final Top-N evidence fragments (e.g., N=5).

[0022] Evidence format requirements: Each piece of evidence should include: evidence number, document identifier, page / paragraph location information, summary fragment, and search score. This format facilitates precise referencing of the original source in the response.

[0023] 4. Evidence generation and credibility: 41. Generation Principle: The Retrieval Enhancement Generation (RAG) strategy is adopted: the Top-N evidence fragments are used as context, and the model is required to "answer based solely on evidence" according to the predefined prompt template.

[0024] The response format must include: conclusion, step-by-step investigation or recommendations, cited evidence numbers, and confidence score.

[0025] 42. Credibility Constraints: If the answer involves quantifiable parameters (voltage, current, temperature, resistance, etc.), unit consistency and numerical range verification should be performed. If the verification fails, it should be marked as low credibility or manual intervention should be triggered.

[0026] For recommendations that may lead to hardware modifications or security risks (such as "replacing the main power module"), the system should require manual review before implementation.

[0027] 5. Confidence assessment and manual takeover strategy: Example of confidence score calculation: Confidence score = w1 × average semantic similarity + w2 × model-generated confidence score + w3 × rule validation (1 for passing, 0 for failing).

[0028] Where w1, w2, and w3 are configurable weights (example: w1+w2+w3=1).

[0029] Decision threshold: conf≥0.65: Direct automatic response and provision of evidence.

[0030] 0.45≤conf<0.65: Automatic reply with the message "Manual confirmation recommended".

[0031] If conf < 0.45 or a high-risk rule is triggered: Automatically escalate to manual takeover and push the context and evidence to the manual collaboration panel.

[0032] Human Collaboration Panel Function: Human FAEs can view the conversation context, highlight evidence, edit answers, and write the confirmed content back to the knowledge base as a new knowledge entry or training sample with one click in the panel.

[0033] 6. Feedback Collection and Closed-Loop Learning Feedback types and collection process: Explicit user feedback (satisfied / unsatisfied, star rating).

[0034] Work order closed-loop results (resolved / unresolved, maintenance measures and results).

[0035] Manual FAE labeling (accept / reject / modify).

[0036] Active learning pool strategy: Prioritize the inclusion of samples with low confidence and those manually labeled as incorrect.

[0037] The samples are stratified by priority (high priority: safety-related and common failure types).

[0038] Fine-tuning cycle and process: Fine-tuning trigger conditions: sample accumulation reaches a threshold or at a fixed cycle (e.g., monthly).

[0039] Fine-tuning objectives: Improve the accuracy and evidence coverage of parameter confirmation and troubleshooting issues.

[0040] Version control and canary releases: Each model update is first validated in a gray-scale environment (e.g., 5% of users), and then gradually scaled up after comparing key KPIs (Top-1 accuracy, manual takeover rate, user satisfaction).

[0041] 7. Key points for model fine-tuning: Data preparation: The training set was constructed using manually labeled question-answer pairs, low-confidence error-correction samples, and synthetic variants.

[0042] Priority tasks: Parallel training of the quality of generative question answering (whether it returns the correct conclusion and evidence) and the binary classification task of "whether human intervention is required".

[0043] Evaluation metrics: Top-1 accuracy, evidence coverage (the percentage of responses containing at least one correct piece of evidence), human intervention rate, and user satisfaction.

[0044] Active learning combined with work orders: The final results of work orders are used as high-quality labels to write back to the training pool to enhance the model's ability to adapt to real-world fault scenarios.

[0045] 8. Actual Results Functionality: Top-1 question-answering accuracy ≥ 85% and evidence coverage ≥ 95% in a verification set of 200 real work orders; Performance: P95 response time <3 seconds (retrieval + generation) under 100 concurrent queries; monthly system availability ≥99.5%.

[0046] Security: All private documents are statically encrypted, and transmission encryption and access control comply with enterprise compliance requirements.

[0047] Implementation Results: After implementation, this solution can significantly improve the response efficiency of enterprise FAEs, reduce repetitive work, and systematically accumulate human experience back into the knowledge base, thereby shortening the average problem resolution time and improving customer satisfaction.

[0048] Example 2 (Vertical Industry of Terminal Intelligent Manufacturing) 1. Data sources: MES / SCADA / PLC alarm logs, SOP documents, maintenance work orders, and sensor timing data.

[0049] Example initial scale: approximately 2,000 documents and 500,000 time-series log entries.

[0050] 2. Standardized knowledge processing workflow 21. Data Acquisition: Periodically retrieve operational data and alarm logs from MES / SCADA via system interface; batch import SOPs and maintenance records.

[0051] 22. Preprocessing and semantic markup: Aggregate time-series alarms into event windows and generate natural language summaries (event summaries), and decompose SOPs into process units and label parameter thresholds.

[0052] Map alarm codes (alarm code → meaning → severity level) and map device serial numbers to physical locations and models.

[0053] 23. Knowledge Representation: Vectorize event summaries, process units, and historical work orders and store them in a vector database; construct a knowledge graph of "process-equipment-alarm-root cause-solution".

[0054] Weights are assigned to time attributes in the index to support time-aware retrieval.

[0055] 3. Time-aware retrieval and generation Time decay weight: A time weight function is introduced during the retrieval phase to improve the retrieval priority of recency events.

[0056] RAG generated content: The generated content consists of three parts: possible root causes (sorted by probability), suggested investigation steps (item by item), and required supplementary data or measurement items (if needed).

[0057] All conclusions should be accompanied by historical evidence (event summary ID, SOP paragraph ID, historical work order ID).

[0058] 4. Credibility and Takeover High-risk strategies: If the suggestion includes "shutdown", "replacement of critical components" or involves personnel safety, the system will force manual confirmation and will not allow automatic implementation.

[0059] Binary takeover model: During the model fine-tuning phase, a classifier for "whether manual takeover is needed" is trained in parallel. This classifier serves as an important component of the decision flow and is used to reduce the false positive rate.

[0060] 5. Model fine-tuning scheme 51. Data Construction: Historical work order pairs: 20,000 samples containing "alarm-root cause-action-result"; approximately 10,000 alarm-root cause tags.

[0061] 52. Multi-task training: The generation task (RAG) and the takeover classification task are trained in parallel; the loss function is the weighted sum of the generation loss and the classification loss.

[0062] 53. Temporal Enhancement: Add event splicing examples from different time windows to the training samples to enhance the model's ability to identify temporal correlations.

[0063] 54. Active learning priority: Prioritize the inclusion of low-confidence samples that affect production capacity (such as line stoppages and quality incidents) into the manually labeled queue. 6. Actual Results Fault location accuracy ≥80% (validation set), average work order processing time reduced by ≥20% compared to historical levels (statistical period comparison), high-risk cases are recommended to proceed to manual confirmation process 100%.

[0064] This embodiment can shorten the fault diagnosis time of the production line, reduce the misoperation caused by misjudgment, and precipitate the manufacturing site experience into reusable knowledge items, thereby improving the overall stability and production efficiency of the production line.

[0065] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for supporting AI technology for electronic components based on an enterprise's private knowledge base, characterized in that, Includes the following steps: Receive enterprise proprietary technical data or information from at least one data source, which may be any one or more of local upload, FTP / SFTP, enterprise email, ERP / PLM system, or third-party storage interface; The received data is processed by text extraction and optical character recognition, document structure parsing, table cell extraction and metadata annotation, and the data is segmented into searchable semantic units and the terminology is standardized. The semantic units are generated into vector representations and stored in a vector database. At the same time, entities and relations are extracted, and the enterprise's private knowledge graph is updated. When a user query is received, the query is vectorized and a set of candidate evidence fragments is retrieved based on a hybrid retrieval strategy of vector retrieval and inverted index. Using the obtained candidate evidence fragments as the retrieval context, a response is generated through a retrieval enhancement generation strategy, and the response is accompanied by the corresponding evidence fragment identifier and location information. A confidence assessment is performed on the answer, which is based on a weighted fusion of retrieval similarity, generation model confidence score, rule assertion verification, and source credibility. When the confidence level is lower than the preset threshold or a high-risk rule is triggered, the session will be upgraded to manual takeover according to the preset strategy. Collect user confirmations, ratings, work order results, and manual annotation feedback information, and use the annotated samples to update the knowledge base index or as fine-tuning data in the periodic model / retrieval strategy retraining process.

2. The method for AI technology support of electronic components based on an enterprise private knowledge base according to claim 1, characterized in that: The hybrid retrieval strategy includes first retrieving the top K semantic units in the vector database using KNN, then performing fine ranking on the above K results based on BM25 or inverted index, and finally generating the final top-N evidence fragments by fusing and ranking according to semantic similarity, BM25 score, document credibility, time weight and historical click rate through a learned ranking model.

3. The method for AI technology support of electronic components based on an enterprise private knowledge base according to claim 1, characterized in that: The knowledge graph construction includes entity recognition and synonym normalization based on a domain terminology list, extraction of entity types such as device model, pin definition, performance parameters, and fault codes, and identification of causal, substitution, and subordinate relationships between entities, supporting graph-based reasoning retrieval.

4. The method for AI technology support of electronic components based on an enterprise private knowledge base according to claim 1, characterized in that: The evidence-based generation and credibility steps further include explicitly limiting the response to "only using the given evidence" in the model generation prompts, and performing unit and numerical range verification when the generated results involve numerical values ​​or parameters. If the verification fails, the answer will be marked as low trust and manual intervention will be triggered, or a "manual confirmation required" response will be returned to the user.

5. The method for AI technology support of electronic components based on an enterprise private knowledge base according to claim 1, characterized in that: The manual takeover includes pushing the session context, matched evidence fragments, confidence scores, and recommended action sets to the human collaboration panel. The panel allows human FAEs to edit the answers and write the standardized solutions that have been manually confirmed back into the knowledge base as new knowledge entries and as labeled samples for subsequent model fine-tuning.

6. The method for AI technology support of electronic components based on an enterprise private knowledge base according to claim 1, characterized in that: The feedback and closed-loop learning also includes stratifying and filtering the collected feedback samples according to error type and confidence level, prioritizing the import of low-confidence samples that have been labeled as errors by users or humans into the active learning pool, and periodically triggering retrieval strategy fine-tuning or model fine-tuning in a micro-batch manner after manual labeling.

7. The method for AI technology support of electronic components based on an enterprise private knowledge base according to claim 1, characterized in that: Data security and compliance processing is performed on incoming or queried data, including sensitive information identification and on-demand anonymization, static / transmission encryption, role-based access control and audit logging, and the system is deployed in a hybrid cloud manner with private deployment or dedicated line access.

8. An AI technology support system for electronic components based on an enterprise private knowledge base, as described in any one of claims 1-7, characterized in that: It includes a data acquisition module for collecting proprietary technical data from local uploads, FTP / SFTP, corporate emails, ERP / PLM, or external storage interfaces. The preprocessing module, connected to the data acquisition module, is used to perform OCR, document structure parsing, table cell extraction and terminology standardization on the acquired data and output semantic units and metadata. Vectorization service and vector database are used to generate vector representations of the semantic units and store them for semantic retrieval. A knowledge graph engine is used to extract entities and relationships from the semantic units and maintain the enterprise's private knowledge graph. The retrieval and reordering module, connected to the vector database and inverted index, is used to perform vector retrieval, BM25 / inverted index retrieval, and reordering based on fusion features to output a set of candidate evidence. The generation and credibility module is used to generate an answer based on the candidate evidence obtained from the retrieval through retrieval enhancement, and to inject evidence citations and confidence scores into the answer; The strategy engine is used to determine whether to respond automatically or escalate the session to manual takeover based on the confidence score and preset rules. The multi-channel access and session management module supports multiple access channels such as WeChat for Enterprise, WeChat Official Accounts, web customer service, ticketing system and email, and maintains multi-turn session contexts; The human collaboration panel displays the conversation context, candidate evidence, and confidence level, allowing human intervention, editing of responses, and writing human confirmation results back to the knowledge base; The feedback and learning module is used to collect user evaluations, work order results and manual annotations, and is responsible for importing the annotated samples into the active learning or fine-tuning process to update the model or retrieval strategy. The security and operations module is used to implement static and transmission encryption of data, identification and desensitization of sensitive information, role-based access control and audit logs, and supports private or hybrid cloud deployment; The system is configured to return the answer to a user's query and simultaneously provide the corresponding evidence fragment identifier and location information.

Images