Big model-based operator work order operation method and system

Abstract

The application discloses an operator work order operation method and system based on a large model, and belongs to the technical field of artificial intelligence and customer service work order management. The method comprises the following steps: performing real-time voice transcription, noise filtering and dialogue role separation on user incoming call voice, and outputting structured dialogue text; generating a structured work order abstract in combination with a large language model; performing two-way hybrid retrieval in a customer service knowledge base and a historical work order knowledge base to obtain relevant business knowledge and similar historical work orders; performing work order processing scene discrimination and work order intelligent distribution; for fault and / or complaint work orders, performing root cause positioning and attribution analysis and generating a root cause analysis report; and automatically mining question and answer pairs from completed work order data and iteratively updating the knowledge base. The application constructs an automatic solution covering the whole link of work order processing, can significantly improve the processing efficiency and accuracy of work orders, reduces the labor cost, and has strong business expansion capability and self-adaptive iteration capability.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and customer service work order management technology, and in particular to a method and system for operator work order operation based on a large model. Background Technology

[0002] With the rapid development of large language model technology, intelligent interaction is profoundly reshaping the operational model of the customer service industry. Telecom operators' customer service volume is massive and the scenarios are complex, with hotlines handling a huge number of user requests daily, covering various scenarios such as business inquiries, fault reports, package purchases, and complaints / suggestions. The user base is large, the frequency of requests is high, and the issues are fragmented. The traditional work order processing model of human agents "communicating, recording, and organizing simultaneously" is no longer sufficient to meet the dual requirements of efficient operation and high-quality service.

[0003] Currently, customer service work orders from telecom operators are still primarily handled manually, facing multiple practical challenges throughout the process: First, efficiency bottlenecks are prominent. Manual recording and information organization consume a significant amount of agent time, easily leading to call queuing and work order backlogs during peak business hours and regional fault outbreaks, prolonging problem resolution cycles, and diverting agent attention, thus impacting the user's real-time communication experience. Second, the quality and stability of work orders are insufficient. The complexity of telecom operator business rules and tariff systems, coupled with differences in agent proficiency and recording standards, can easily result in omissions of key information, non-standard problem descriptions, and inconsistent wording for similar requests, increasing communication costs in subsequent dispatch and processing stages. Third, the accuracy of problem classification and prioritization is limited. Manual judgment relies heavily on personal experience, making it difficult to consider multiple dimensions such as user emotional intensity, problem impact scope, and fault urgency level, easily leading to misclassification and inaccurate prioritization, resulting in untimely responses to high-risk requests and potential service compliance risks. Fourth, the efficiency of the review process is low. The quality inspection process relies on manual verification of each item, which requires a large investment of manpower, has a long processing cycle, and makes it difficult to establish a standardized knowledge system. The repeated handling of similar issues further exacerbates the operational burden.

[0004] Therefore, there is an urgent need to design a method and system that can automate the entire process of operator work orders to address the shortcomings and deficiencies in the existing technologies. Summary of the Invention

[0005] The present invention aims to provide a method and system for operator work order operation based on a large model, so as to solve the technical problems of low efficiency, unstable quality, limited classification accuracy, and low efficiency of review and transfer in the existing technology of operator customer service work order processing.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A large-scale model-based operator work order operation method includes the following steps: Step 1: Record the user's incoming voice in real time, and perform speech-to-text processing, noise filtering, and dialogue role separation on the recorded voice to output structured dialogue text; Step 2: Based on the structured dialogue text, and combining the prompt word template with the large language model, automatically generate a structured task summary; Step 3: Based on the structured dialogue text and structured work order summary, retrieve relevant business knowledge and similar historical work orders from the customer service business knowledge base and the historical work order knowledge base; Step 4: Based on the structured work order summary, relevant business knowledge, and similar historical work orders, determine the work order processing scenario and assign the work order; Step 5: For fault-related work orders and / or complaint-related work orders, based on the structured work order summary, relevant business knowledge, and similar historical work orders, perform root cause localization and attribution analysis, and generate a root cause analysis report; Step 6: Automatically mine question-answer pairs based on completed work order data, and use them to update the customer service business knowledge base and / or the historical work order knowledge base.

[0007] Furthermore, in step 1, the specific method for outputting the structured dialogue text is as follows: Real-time incoming call voice streams are transcribed using the Whisper end-to-end speech recognition engine. The audio data is segmented based on speaker log technology to distinguish and label the dialogue roles of agents and users. VAD (Voice Endpoint Detection) technology is used to filter out invalid silent segments and background noise during a call; The large model is used to perform intelligent error correction and sentence smoothing on the original transcribed text, removing interjections and repetitive sentences to generate a standardized and complete dialogue text.

[0008] Furthermore, in step 2, the specific method for generating the structural chemical single abstract is as follows: The structured dialogue text is semantically segmented into paragraphs, and the core key statements of the dialogue are roughly screened using extractive summarization technology. Construct thought chain prompts containing knowledge of operator business domains to guide the large model to output JSON format work order summaries according to a five-segment structure: customer demand - problem phenomenon - measures taken - unresolved issues - emotional intensity. The document uses a textual entailment model to compare the work order summary with the structured dialogue text to perform a factual consistency check.

[0009] Furthermore, the elements in the structured chemical order summary include: appeal category, core issue, user intent, sentiment tag, involved product / package, fault phenomenon, and urgency level.

[0010] Furthermore, step 3 specifically includes: Keywords are extracted from the structured dialogue text and structured order summary using a large model, and an Elasticsearch BM25 sparse search query is constructed for retrieval and recall in the customer service business knowledge base. The structured work order summary and dialogue context are encoded into a dense vector using a sentence embedding model, and semantic similarity retrieval and recall are performed in the historical work order knowledge base based on the FAISS vector engine. The two recall results are sorted using the Cross-Encoder reordering model, and the business knowledge items and similar historical work orders most relevant to the current request are output. When the overall score of the recall results is lower than the preset threshold, the query rewriting mechanism is triggered. The user's colloquial expression is generalized into a standard question through a large model and then the retrieval is re-executed.

[0011] Furthermore, in step 4, the specific steps for determining the work order processing scenario and assigning the work order are as follows: The structured chemical order summary is further classified by large model, matched with a preset skill group routing table, and the corresponding processing department and professional skill group are determined. For work orders whose model classification confidence is lower than a preset threshold, the department to which the work order belongs is determined by majority voting based on similar historical work orders returned by the retrieval. Dynamically adjust work order priorities by combining emotion intensity tags, user level, and the scope of the issue's impact. For inquiries that do not require dispatching, trigger automatic responses from the knowledge base or guide users to self-service; for work orders that require offline processing, generate dispatch suggestions and push them to the agent's workbench.

[0012] Furthermore, in step 5, the specific steps for performing root cause localization and attribution analysis are as follows: The structured work order summary, related business knowledge, and time, region, and product dimension data of similar historical work orders are integrated into an event graph. Common cause analysis of batch failure work orders was performed using causal discovery algorithms to identify key factors. For complaint tickets, we use a large model thinking chain reasoning and combine it with the user's historical interaction records to dig out the underlying reasons. Generate a root cause analysis report using a structured format of phenomenon-proximal cause-root cause-improvement suggestion.

[0013] Furthermore, step 6 specifically includes: Based on completed work order data, a large model is used to match and extract user questions and agent answers to generate candidate question-answer pairs. The SimCSE contrastive learning model is used to vectorize candidate question-answer pairs, and the DBSCAN density clustering algorithm is combined to aggregate and classify semantically similar user questions to generate representative standard questions. High-value samples are screened through proactive learning and then manually reviewed and calibrated. After the review is approved, the data is updated to the business knowledge base and the vector retrieval index is iterated synchronously; if the review fails, the process returns to the root cause analysis step to continue analyzing the root cause of the problem.

[0014] Furthermore, step 6 also includes: real-time monitoring of the retrieval hit rate and usage rate of each entry in the customer service business knowledge base and the historical work order knowledge base, and automatically marking, updating or removing knowledge entries that have no recall records for a long time or have become invalid.

[0015] This invention also provides a large-scale operator work order operation system for implementing the above-mentioned operator work order operation method, including: The text dialogue extraction module is used to record users' incoming voice messages in real time, and to process the recorded voice messages into text, filter noise and separate dialogue roles, and output structured dialogue text. The work order summary generation module is used to automatically generate a structured work order summary based on the structured dialogue text, combined with prompt word templates and a large language model; The knowledge base retrieval module is used to retrieve relevant business knowledge and similar historical work orders from the customer service business knowledge base and the historical work order knowledge base based on the structured dialogue text and the structured work order summary. The work order dispatch module is used to perform work order processing scenario identification and intelligent work order dispatch based on the structured work order summary, relevant business knowledge and similar historical work orders. The root cause analysis module is used to perform root cause localization and attribution analysis and generate root cause analysis reports for fault-related work orders and / or complaint-related work orders. The knowledge evolution module is used to automatically mine question-answer pairs based on completed work order data and update the customer service business knowledge base and / or the historical work order knowledge base. Compared with the prior art, the present invention discloses the following technical effects: 1. Improve Service Quality: This invention leverages large-model customized prompt word engineering, multimodal processing technology, and semantic similarity retrieval algorithms to comprehensively improve the accuracy and standardization of work order processing from the data source, information processing, and fault analysis. It abandons the traditional, crude approach of manual recording and analysis, using speech recognition, large-model thought chain reasoning, and extractive summarization technology to standardize call text and accurately extract core work order information. A verification mechanism eliminates information omissions, expression deviations, and large-model illusions, ensuring the completeness, authenticity, and consistency of work order information. Simultaneously, it integrates dual-path retrieval algorithms, adaptive rewriting, and causal reasoning technology to efficiently adapt to complex scenarios such as users' colloquial requests and fragmented faults, accurately completing knowledge matching and in-depth fault tracing. This solves the shortcomings of traditional work order judgments—one-sidedness, superficial tracing, and high error rates—significantly improving the accuracy of operator work order processing and overall service quality.

[0016] 2. Simple Operation: This invention constructs a fully automated operation system, eliminating the need for manual participation, organization, verification, and operation at each step, greatly simplifying the overall operation process of customer service work orders. During daily work order processing, the system can automatically complete caller ID transcription, structured dialogue text processing, work order element extraction, standard summary output, business knowledge matching, work order priority determination, and dispatch suggestion push. This eliminates the need for agents to manually record, fill out forms, perform keyword searches, manually identify dispatch departments, and manually review fault causes. Simultaneously, the system automatically outputs standardized work order summaries, similar handling solutions, and structured root cause reports, pushing them to the agent's workbench in an intuitive, organized, and unified format. This significantly reduces the operational threshold and analysis difficulty for agents, making the overall operation simple and efficient, greatly simplifying the daily operation, processing, review, and optimization of operator customer service work orders.

[0017] 3. Reduced Labor Costs: This invention replaces traditional manual operations with end-to-end automated processing, effectively reducing a large amount of repetitive and basic manual work, and significantly lowering the operator's customer service manpower and knowledge base maintenance costs. Traditional work order models require a significant investment of agent staff for call recording, work order filling, information organization, and knowledge retrieval. Simultaneously, dedicated maintenance personnel are needed to regularly manually organize, proofread, and update the business knowledge base, resulting in high manpower investment, low work efficiency, and high iteration costs. The system, through large-scale model thinking chain reasoning, semantic clustering, automatic accumulation, and dynamic replacement mechanisms, automatically standardizes completed work order issues, automatically extracts high-frequency fault issues, and automatically marks and removes outdated knowledge. This compresses the traditional bi-weekly manual knowledge base iteration cycle to daily automatic iteration, eliminating the need for extensive manual organization and proofreading. Simultaneously, end-to-end automated processing significantly reduces the ineffective working hours of agents, freeing up manpower to focus on user problem solving and high-end service scenarios, effectively reducing redundant manpower costs and knowledge base maintenance costs.

[0018] 4. Strong Scalability: This invention is optimized based on operator business knowledge, fault dictionaries, package rules, and operation and maintenance specifications. It is not limited to a single business type and can comprehensively adapt to all types of work order scenarios, including operator broadband faults, package inquiries, service changes, billing complaints, and service anomalies. It boasts broad scenario coverage and strong universal adaptability. The overall system architecture is modular, lightweight, and decoupled, with each functional module operating independently yet efficiently interconnected. It can be flexibly expanded to include new functions, new business scenarios, and new knowledge categories according to business needs. Furthermore, relying on a knowledge evolution mechanism, the system can autonomously adapt to iterative user demands, new terminal equipment faults, and new service package rules without large-scale system architecture reconstruction. It can undergo long-term adaptive iterative upgrades, possessing strong business expansion capabilities and long-term technological evolution capabilities, adapting to the continuous intelligent operation upgrade needs of operators. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the operator work order operation method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the operator work order operation method provided in this embodiment of the invention; Figure 3 This is a diagram illustrating the architecture of a carrier work order operation system provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention provides a carrier work order operation method based on a large model, such as... Figure 1-2 As shown, it includes the following steps: Step 1: Record the user's incoming voice in real time, complete voice-to-text conversion, noise filtering and dialogue role separation, and output standardized, high-purity structured dialogue text.

[0023] Specifically, based on the Whisper end-to-end speech recognition engine, streaming speech transcription is achieved, keeping the overall word error rate below 8%. Combined with speaker log technology, precise turn-by-turn segmentation is performed on dual-channel or mono audio data, clearly distinguishing and marking the dialogue roles of agents and users. At the same time, VAD speech endpoint detection technology is used to filter invalid silent segments and background noise during the call, and a large model is used to intelligently correct errors and smooth sentences in the original transcribed text, automatically removing interjections, repetitive sentences and identifying erroneous content, generating standardized and complete dialogue text.

[0024] This invention controls the single-call transcription delay to within 2 seconds, replacing the traditional manual recording operation, effectively improving the readability and standardization of the dialogue text, and providing high-quality standardized input data for subsequent modules such as work order summary generation and knowledge evolution.

[0025] Step 2: Based on standardized dialogue text, combined with customized prompt word templates and a fine-tuned large language model, automatically generate a structured bill summary containing elements such as request category, core issue, user intent, emotion tag, involved products / packages, fault phenomenon, and urgency level.

[0026] Specifically, the dialogue text is first semantically segmented into paragraphs, and core key statements are roughly screened using extractive summarization technology. Then, thought chain prompts incorporating operator business domain knowledge are constructed to guide the large model to output a structured JSON-formatted work order summary in five segments: "Customer Request—Problem Phenomenon—Measures Taken—Unresolved Issues—Emotional Intensity (Level 1-5)". To prevent model illusion, a textual implication model is used to compare the summary content with the original dialogue text, performing factual consistency checks to ensure zero deviation of key information. The summary generation time of this invention is less than 3 seconds, and the key information coverage rate can reach over 95%, unifying work order description standards and completely eliminating problems such as information omissions, non-standard expressions, and content differences caused by differences in agent recording.

[0027] Step 3: Knowledge base retrieval steps.

[0028] The knowledge base includes a customer service business knowledge base and a historical work order knowledge base. The business knowledge base includes business knowledge related to hotline issues and common inquiry questions; the historical work order knowledge base is a knowledge base transformed from historical work order records accumulated in the hotline work order database. Both knowledge bases need to be continuously iterated and updated over time.

[0029] During the retrieval process, core keywords are first extracted based on standardized dialogue text and structured work order summaries. A dual-path hybrid retrieval and re-ranking architecture is then constructed. Semantic analysis retrieves relevant business knowledge from the customer service business knowledge base and similar work orders from the historical work order knowledge base, providing a basis for intelligent work order dispatch. Next, a large model is used to extract keywords from the dialogue text and work order summaries, constructing an Elasticsearch BM25 sparse retrieval system to recall relevant business knowledge (Elasticsearch BM25 is a widely used relevance scoring function in search engines). Simultaneously, the work order summary and dialogue context are encoded into dense vectors using a sentence embedding model. Semantic similarity retrieval is then performed in the historical work order knowledge base using the FAISS vector engine to obtain another retrieval result. Finally, the two retrieval results are accurately ranked using a Cross-Encoder re-ranking model, outputting the most relevant business knowledge items (FAQ, operation manual) and the top 3 similar historical work orders. Simultaneously, an intelligent query rewriting mechanism is configured to generalize and rewrite the original user requests with low recall rates through a large model (e.g., rewriting "optical modem red light" as "optical modem optical signal loss fault handling"), adapting to standardized business scripts, and significantly improving the accuracy of knowledge and work order recall.

[0030] This step provides a reliable basis for intelligent order dispatching and agent processing of work orders, effectively shortening user call duration, with an average call duration reduction of 25%.

[0031] Step 4: Based on the work order summary information, retrieved business knowledge, and similar historical work order data, the system achieves work order processing scenario identification and intelligent work order assignment. First, a large model is used to finely classify the intent of the work order summary, matching it with a preset skill group routing table to initially obtain the corresponding processing department and professional skill group. If the model classification confidence is insufficient, the system uses a majority voting method to determine the assigned department based on historical similar work orders returned by knowledge retrieval. Simultaneously, the system dynamically adjusts work order priority by combining emotion intensity tags, user level, and the scope of the problem's impact. For consultation-type questions that do not require assignment, the system directly triggers automatic responses from the knowledge base or guides self-service; for work orders requiring offline maintenance by the responsible department, the system generates assignment suggestions and pushes them to the agent's workbench, supporting manual selection for one-click confirmation or fully automatic assignment. This invention can improve the accuracy of work order assignment to 88%, reduce the misassignment rate of urgent work orders by 60%, and ensure that high-risk requests reach professional agents within 5 minutes, effectively avoiding service compliance risks and improving work order processing efficiency.

[0032] Step 5: For key work orders such as those related to faults and complaints, intelligent root cause localization and attribution analysis of faults and complaints is achieved by combining work order summaries, related business knowledge, and historical work order data. Specifically, work order summaries, related business knowledge, and time, region, and product dimension data of similar historical work orders are first integrated into an event graph. Then, a causal discovery algorithm is used to perform common cause analysis on batch of fault work orders to identify key factors. For simple complaint work orders, a large-scale model thinking chain reasoning is used, combined with user historical interaction records, to uncover the underlying causes caused by billing rule conflicts, service process breakpoints, etc. The analysis results are presented in a structured report of "phenomenon-proximate cause-root cause-improvement suggestion" and are automatically synchronized to the knowledge evolution module, realizing batch fault root cause localization. The speed is reduced from the traditional hourly processing time to the minute level, and the accuracy rate of repeat complaint warnings reaches 80%, driving continuous optimization of business rules and processes, and realizing the transformation of customer service work orders from "passive processing" to "proactive prevention".

[0033] Step 6: Automatically mine new questions and answers from massive daily user call and completed work order data to form standardized question-answer pairs, enabling automated iteration, updating, and optimization of the knowledge base, and building a closed-loop knowledge accumulation system. For completed work orders, a large model is used to pair and extract "user question - agent answer" to generate candidate question-answer pairs. The SimCSE (Simple Contrastive Learning of Sentence Embeddings) contrastive learning model is used to vectorize the candidate question-answer pairs. Combined with DBSCAN density clustering, semantically similar user questions are aggregated and categorized to generate representative standard questions. For high-frequency consultation questions and typical failure cases, high-value samples are actively selected through learning and pushed to manual review and calibration to ensure accurate and compliant knowledge. After review and approval by business experts, the knowledge base is updated in batches, and the vector retrieval index is iterated simultaneously. If the expert review fails, the system returns to the root cause analysis module to continue analyzing the root cause of the problem and updating the knowledge question-answer pairs based on the analysis results until the business expert review is approved. Simultaneously, the system monitors the retrieval hit rate and usage rate of each item in the customer service business knowledge base and historical work order knowledge base in real time, automatically marking, updating, or removing knowledge items with no recall records for a long time or that have become invalid. Ultimately, it can achieve an average monthly addition of more than 200 effective business knowledge points, reduce the duplication rate of similar issues by 35%, and compress the knowledge iteration cycle from 2 weeks to 1 day, truly realizing a closed loop of work order processing and real-time knowledge accumulation.

[0034] This invention also provides a large-scale operator work order operation system, such as... Figure 3 As shown, it includes: The text dialogue extraction module is used to record users' incoming voice messages in real time, and to process the recorded voice messages into text, filter noise and separate dialogue roles, and output structured dialogue text. The moment a user initiates a call through the customer service hotline and the system connects the call, the system automatically triggers the text dialogue extraction process. The system queries the customer information database, retrieving basic user information, package details, terminal device model, and historical service records in real time based on the caller's number and relating them to the current conversation context. Simultaneously, the Whisper streaming speech recognition engine continuously transcribes the real-time incoming call audio stream. Using a speaker log algorithm, it performs accurate turn-taking segmentation of mono or stereo audio, automatically distinguishing and labeling the "customer" and "agent" roles, generating the original transcribed text content between the customer service representative and the user. Combined with VAD (Voice Endpoint Detection) technology, it removes useless information such as silences during call intervals, ambient noise, and invalid background noise, retaining only valid human voice segments. Based on this, the system uses large-model-specific prompts to intelligently correct errors, deduplicate, and smooth the original transcribed text, removing interjections, repetitive expressions, and incorrectly identified characters. Finally, it outputs a structured dialogue text with clear roles, standardized sentences, and no redundant noise.

[0035] The work order summary generation module is used to automatically generate a structured work order summary based on the structured dialogue text, combined with prompt word templates and a large language model; After acquiring the structured dialogue text, the system automatically triggers the work order summary generation process. First, the long-sequential dialogue text is semantically segmented into paragraphs. Extractive summarization technology is used to filter redundant information and perform a coarse screening of key information, removing ineffective content such as small talk, confirmations, and irrelevant chatter, retaining core semantic content strongly related to user requests, fault symptoms, user emotions, and urgency. The system then uses a command-based fine-tuning of a large language model combined with multi-level prompt word templates, concatenating key statements with ontology information and inputting them into the large model to generate a structured JSON work order summary.

[0036] The work order summary information includes: request category, core issue, user intent, emotion tag, involved product / package, fault phenomenon, and urgency level. The prompt word template is designed as a five-stage thought chain guide. The example prompt word is: "You are a customer service work order assistant of an operator. Please generate a JSON-formatted work order summary based on the following dialogue. Analyze according to 'Customer Request - Problem Phenomenon - Measures Taken - Problem to be Solved - Emotion Intensity (1-5 levels)' and output key fields such as request category, core issue, user intent, emotion tag, involved product / package, fault phenomenon, and urgency level." For example, if a user reports the problem "Broadband is not working, and the optical modem has a red light," the generated summary would be: {"Request Category":"Broadband Fault Report","Core Issue":"Optical Modem LOS Red Light","User Intent":"Restore Internet Access","Emotion Tag":4,"Involved Product":"Home Broadband 100M","Fault Phenomenon":"Optical Modem Optical Signal Loss","Urgency Level":"Urgent"}. To prevent model illusion, the system introduces a fact consistency verification module based on the text implication model, which compares the generated summary with the original dialogue item by item to ensure the accuracy of key information such as account information, failure time, device model, and failure scenario.

[0037] The knowledge base retrieval module is used to retrieve relevant business knowledge and similar historical work orders from the customer service business knowledge base and the historical work order knowledge base based on the structured dialogue text and the structured work order summary. Specifically: First, a large model is used to extract keywords such as "broadband," "optical modem," and "LOS red light" from structured dialogue text and work order summaries. An Elasticsearch BM25 sparse search query is then constructed to recall candidate knowledge items from the business knowledge base. Simultaneously, the summary and dialogue context are encoded into dense vectors using a sentence embedding model. Based on the FAISS vector engine, semantic similarity is used to retrieve historical work orders relevant to the user's request. Then, the two retrieved results are precisely ranked using a Cross-Encoder re-ranking model, outputting the most relevant business knowledge (e.g., "Optical Modem LOS Red Light Fault Handling Manual") and the top 3 historically similar work orders. When the combined score of the two retrieved results is low, query rewriting is triggered: the large model is invoked to generalize the user's colloquial expression into a standard question, such as rewriting "optical modem red light" as "optical modem optical signal loss fault handling," and the search is re-executed to improve the model's recall rate.

[0038] The work order dispatch module is used to perform work order processing scenario identification and intelligent work order dispatch based on the structured work order summary, relevant business knowledge and similar historical work orders. The specific steps are as follows: First, the large model refines the classification of the work order summary by business type, fault phenomenon, and user request attributes, and matches it with the system's preset business skill group routing table to initially identify the corresponding processing department and professional skill group. For example, if the identification is "Broadband Fault Report > Optical Modem Problem > LOS Red Light", the system's skill group routing table is matched to obtain the candidate processing department as "Network Department > Delivery Center". For difficult work orders with insufficient model classification confidence or unclear scenario boundaries, the system retrieves the top 3 similar historical work orders and uses majority voting to determine the optimal processing department, avoiding misassignment due to errors in a single model. At the same time, the system dynamically assigns three levels of priority to work orders—ordinary, urgent, and extremely urgent—by combining multiple dimensions such as user emotional intensity level, user star rating, fault impact range, and whether it is a group fault issue, so as to prioritize the processing of high-risk work orders. For simple inquiries that do not require dispatch, the system triggers an automatic response from the knowledge base or guides users to self-service. For work orders that require offline maintenance by professional departments, such as equipment failures, network anomalies, user complaints, and business anomalies, the system combines departmental skill matching, staff availability, and the urgency of the work order to generate dispatch suggestions that include the handling department, the person handling the task, and the handling time limit. These suggestions are then pushed to the agent's workbench and can be confirmed manually with one click or automatically assigned by the system.

[0039] The root cause analysis module is used to perform root cause localization and attribution analysis and generate root cause analysis reports for fault-related work orders and / or complaint-related work orders. Specifically, for assigned fault work orders, the system continuously executes a root cause analysis process to pinpoint the core causes of the problem. The specific steps are as follows: The system gathers current work order summaries, dialogue text content, related business knowledge, and historical similar work order data in real time, extracting multi-dimensional structured elements such as time, region, equipment model, package type, and fault characteristics to dynamically construct a work order fault event graph. For batches of concentrated fault work orders from the same time period and region, a causal discovery algorithm is used to perform correlation and causal statistical analysis on multi-dimensional features to identify common root causes such as regional network anomalies, data center equipment failures, terminal batch defects, and batch errors in backend configuration. For individual faults and user complaints, the system relies on large-scale model thinking chain reasoning prompts to break down the surface fault phenomenon, direct causes, intermediate related factors, and deep business root causes layer by layer, assisting in identifying hidden problems that are difficult to detect manually, such as process breakpoints, rule conflicts, configuration anomalies, and service omissions. Finally, an analysis report is generated according to a standardized format of "fault phenomenon - direct causes - core root causes - business impact - optimization suggestions". For example, in response to multiple LOS red light work orders for optical modems in a certain area, the system found that the common cause was a faulty OLT equipment board and generated a root cause analysis report.

[0040] After a fault work order is closed, the knowledge evolution module will continuously execute the knowledge evolution closed-loop process to realize the automatic accumulation, organization, iteration and replacement of business knowledge. Specifically, the knowledge evolution module periodically traverses all completed work orders and their corresponding call text data daily. It extracts "user question - agent standard answer" pairings using a large model to generate candidate knowledge base materials. Then, it uses the SimCSE contrastive learning model to vectorize the knowledge and combines it with the DBSCAN density clustering algorithm to automatically cluster similar questions with similar semantics but different colloquial expressions, generating standardized question formats and corresponding standard answers. For high-frequency consultation questions and typical fault cases, it actively learns to select high-value samples for manual review and calibration, ensuring accurate and compliant knowledge. Approved standardized questions and answers and fault handling cases are entered into the knowledge base, and the text retrieval index and vector retrieval index are updated synchronously. If the review fails, it returns to the root cause analysis module to continue analyzing the root cause of the problem, updating the knowledge pairs based on the analysis results until expert review is passed. Simultaneously, this invention also monitors the retrieval hit rate and usage rate of each item in the knowledge base in real time, automatically marking, updating, or removing knowledge items that are long-term invalid, have outdated rules, or are obsolete in certain scenarios.

[0041] The operator work order operation method and system based on a large model provided by this invention can significantly reduce the workload of agents in recording data, free up manpower to focus on problem solving, shorten the work order processing time, and improve service response speed and processing accuracy. It has important practical value for optimizing the operator's customer service operation system and improving user satisfaction.

[0042] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for operator work order operation based on a large model, characterized in that, Includes the following steps: Step 1: Record the user's incoming voice in real time, and perform speech-to-text processing, noise filtering, and dialogue role separation on the recorded voice to output structured dialogue text; Step 2: Based on the structured dialogue text, and combining the prompt word template with the large language model, automatically generate a structured task summary; Step 3: Based on the structured dialogue text and structured work order summary, retrieve relevant business knowledge and similar historical work orders from the customer service business knowledge base and the historical work order knowledge base; Step 4: Based on the structured work order summary, relevant business knowledge, and similar historical work orders, determine the work order processing scenario and assign the work order; Step 5: For fault-related work orders and / or complaint-related work orders, based on the structured work order summary, relevant business knowledge, and similar historical work orders, perform root cause localization and attribution analysis, and generate a root cause analysis report; Step 6: Automatically mine question-answer pairs based on completed work order data, and use them to update the customer service business knowledge base and / or the historical work order knowledge base.

2. The operator work order operation method according to claim 1, characterized in that, In step 1, the specific method for outputting the structured dialogue text is as follows: Real-time incoming call voice streams are transcribed using the Whisper end-to-end speech recognition engine. The audio data is segmented based on speaker log technology to distinguish and label the dialogue roles of agents and users. VAD (Voice Endpoint Detection) technology is used to filter out invalid silent segments and background noise during a call; The large model is used to perform intelligent error correction and sentence smoothing on the original transcribed text, removing interjections and repetitive sentences to generate a standardized and complete dialogue text.

3. The operator work order operation method according to claim 1, characterized in that, In step 2, the specific method for generating the structural chemical single abstract is as follows: The structured dialogue text is semantically segmented into paragraphs, and the core key statements of the dialogue are roughly screened using extractive summarization technology. Construct thought chain prompts containing knowledge of operator business domains to guide the large model to output JSON format work order summaries according to a five-segment structure: customer demand - problem phenomenon - measures taken - unresolved issues - emotional intensity. The document uses a textual entailment model to compare the work order summary with the structured dialogue text to perform a factual consistency check.

4. The operator work order operation method according to claim 3, characterized in that, The elements in the structured chemical order summary include: appeal category, core issue, user intent, sentiment tag, involved product / package, fault phenomenon and urgency level.

5. The operator work order operation method according to claim 1, characterized in that, Step 3 specifically includes: Keywords are extracted from the structured dialogue text and structured order summary using a large model, and an Elasticsearch BM25 sparse search query is constructed for retrieval and recall in the customer service business knowledge base. The structured work order summary and dialogue context are encoded into a dense vector using a sentence embedding model, and semantic similarity retrieval and recall are performed in the historical work order knowledge base based on the FAISS vector engine. The two recall results are sorted using the Cross-Encoder reordering model, and the business knowledge items and similar historical work orders most relevant to the current request are output. When the overall score of the recall results is lower than the preset threshold, the query rewriting mechanism is triggered. The user's colloquial expression is generalized into a standard question through a large model and then the retrieval is re-executed.

6. The operator work order operation method according to claim 1, characterized in that, In step 4, the specific steps for determining the work order processing scenario and assigning the work order are as follows: The structured chemical order summary is further classified by large model, matched with a preset skill group routing table, and the corresponding processing department and professional skill group are determined. For work orders whose model classification confidence is lower than a preset threshold, the department to which the work order belongs is determined by majority voting based on similar historical work orders returned by the retrieval. Dynamically adjust work order priorities by combining emotion intensity tags, user level, and the scope of the issue's impact. For inquiries that do not require dispatching, trigger automatic responses from the knowledge base or guide users to self-service; for work orders that require offline processing, generate dispatch suggestions and push them to the agent's workbench.

7. The operator work order operation method according to claim 1, characterized in that, In step 5, the specific steps for performing root cause localization and attribution analysis are as follows: The structured work order summary, related business knowledge, and time, region, and product dimension data of similar historical work orders are integrated into an event graph. Common cause analysis of batch failure work orders was performed using causal discovery algorithms to identify key factors. For complaint tickets, we use a large model thinking chain reasoning and combine it with the user's historical interaction records to dig out the underlying reasons. Generate a root cause analysis report using a structured format of phenomenon-proximal cause-root cause-improvement suggestion.

8. The operator work order operation method according to claim 1, characterized in that, Step 6 specifically includes: Based on completed work order data, a large model is used to match and extract user questions and agent answers to generate candidate question-answer pairs. The SimCSE contrastive learning model is used to vectorize candidate question-answer pairs, and the DBSCAN density clustering algorithm is combined to aggregate and classify semantically similar user questions to generate representative standard questions. High-value samples are screened through proactive learning and then manually reviewed and calibrated. After the review is approved, the data is updated to the business knowledge base and the vector retrieval index is iterated synchronously; if the review fails, the process returns to the root cause analysis step to continue analyzing the root cause of the problem.

9. The operator work order operation method according to claim 8, characterized in that, Step 6 further includes: real-time monitoring of the retrieval hit rate and usage rate of each entry in the customer service business knowledge base and the historical work order knowledge base, and automatically marking, updating or removing knowledge entries that have no recall records for a long time or have become invalid.

10. A large-scale operator work order operation system, used to implement the operator work order operation method according to any one of claims 1-9, characterized in that, include: The text dialogue extraction module is used to record users' incoming voice messages in real time, and to process the recorded voice messages into text, filter noise and separate dialogue roles, and output structured dialogue text. The work order summary generation module is used to automatically generate a structured work order summary based on the structured dialogue text, combined with prompt word templates and a large language model; The knowledge base retrieval module is used to retrieve relevant business knowledge and similar historical work orders from the customer service business knowledge base and the historical work order knowledge base based on the structured dialogue text and the structured work order summary. The work order dispatch module is used to perform work order processing scenario identification and intelligent work order dispatch based on the structured work order summary, relevant business knowledge and similar historical work orders. The root cause analysis module is used to perform root cause localization and attribution analysis and generate root cause analysis reports for fault-related work orders and / or complaint-related work orders. The knowledge evolution module is used to automatically mine question-and-answer pairs based on completed work order data and update the customer service business knowledge base and / or the historical work order knowledge base.

Images