Intelligent customer service model construction method special for second-hand car field and intelligent customer service robot

Abstract

The embodiment of the invention discloses an intelligent customer service model construction method special for the second-hand car field and an intelligent customer service robot, and the method comprises the steps: collecting multi-dimensional business data, carrying out the data cleaning operation of term unification and black word correction, and protecting the privacy of a user and the sensitive information of a vehicle through combining with a hierarchical desensitization mechanism. Instruction fine-tuning training is carried out based on the lightweight basic model, so that the model is accurately adapted to business scenes such as vehicle introduction and demand analysis, and human feedback reinforcement learning is adopted to optimize output preferences. And finally, high-efficiency service is realized by using a high-speed reasoning framework deployment model. According to the embodiment of the invention, the problems of insufficient service suitability, high cost of long cue words and risk of data leakage of a general model are effectively solved, the operation cost of an enterprise is reduced, the customer service response speciality and security are remarkably improved, the service suitability is accurately improved, the data security is effectively guaranteed, and the risk of leakage is avoided.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for constructing an intelligent customer service model and an intelligent customer service robot specifically for the used car industry. Background Technology

[0002] With the development of internet technology and the accelerating trend of online used car transactions, users' demand for obtaining used car information, making inquiries, and comparing options through new media channels (such as short video platforms, social media, and live streaming platforms) is growing. Traditional human customer service models are no longer sufficient to handle the massive, high-frequency, and personalized user inquiries; therefore, intelligent customer service systems are gradually being applied in this field.

[0003] Currently, the intelligent question-and-answer systems in the used car industry mainly employ the following technical solutions:

[0004] A static question-answering solution based on FAQ templates: The system pre-compiles frequently asked questions and answers, retrieves and returns the answers through keyword matching or simple similarity algorithms. This solution is simple to implement and low in cost, but it can only handle fixed questions, cannot understand personalized user needs, and cannot incorporate dynamic data such as real-time vehicle availability and pricing, resulting in poor flexibility.

[0005] Traditional task-oriented dialogue solutions based on intent classification and slot filling identify user intents using machine learning or deep learning models, extract slot information using rule-based or sequence labeling methods, and guide users to complete queries according to a pre-defined dialogue flow. While this approach can accomplish simple tasks to some extent, it requires a large amount of labeled data, involves cumbersome intent and slot configuration, and has limited ability to understand complex, multi-turn dialogues.

[0006] A simple question-answering solution that integrates with a general language model: This solution directly calls a general language model to generate answers, injecting some used car knowledge through prompts. While this solution possesses some natural language understanding and generation capabilities, it lacks effective integration with business data such as real-time vehicle inventory databases and valuation services. This can easily lead to factual errors (i.e., "illusions") and cannot guarantee the business accuracy and controllability of the answers.

[0007] Rule-based vehicle screening and valuation tools: Users fill in parameters through a form, and the system returns valuation or screening results based on preset rules or models. However, this approach is disconnected from natural language dialogue, resulting in a high user barrier, an inability to achieve dynamic interaction and intelligent explanation, and difficulty in answering high-level questions related to vehicle comparisons and applicable scenarios.

[0008] In summary, existing solutions generally suffer from the following problems in the context of intelligent Q&A for used cars:

[0009] It is difficult to combine real-time vehicle availability with business data for accurate recommendations;

[0010] Insufficient understanding of multi-round, progressive user needs;

[0011] Lack of structured domain knowledge support and reasoning ability;

[0012] It cannot achieve dynamic constraint solving and interpretable recommendation mechanisms;

[0013] There is a risk of large-scale model illusion, and the business is poorly controllable;

[0014] Contact information identification and sales lead generation are inefficient and inaccurate.

[0015] Therefore, there is an urgent need for an intelligent question-answering system that can deeply integrate domain knowledge, real-time business data, and large language models, and support multi-turn interaction, dynamic constraint solving, and risk warning, in order to improve the intelligence level and user experience of used car consulting services. Summary of the Invention

[0016] To address these issues, this invention provides a method for constructing an intelligent customer service model and an intelligent customer service robot specifically for the used car industry. This addresses the technical problems in used car intelligent Q&A scenarios, such as difficulty in integrating real-time vehicle data and business data for accurate recommendations, insufficient understanding of multi-round progressive user needs, lack of structured domain knowledge support and reasoning ability, inability to achieve dynamic constraint solving and interpretable recommendation mechanisms, the risk of large model illusion leading to poor business controllability, and low efficiency and accuracy in contact information identification and sales lead generation.

[0017] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0018] According to a first aspect of the present invention, a method for constructing an intelligent customer service model specifically for the used car industry is provided, the method comprising:

[0019] Construct and maintain a knowledge graph for the used car industry, which integrates vehicle brand, model, configuration, year, vehicle condition, market price, residual value, city traffic restriction policies and common fault information in a structured manner;

[0020] Establish a tool call interface framework to connect to and encapsulate a real-time vehicle source database, a vehicle valuation service engine, and a city policy query service;

[0021] Build a dialogue management module to receive natural language input from users on new media channels and maintain multi-turn dialogue contexts that include historical dialogues, identified user demand profiles, and results of invoked tools.

[0022] Based on the large language model and combined with the dialogue context, user intent recognition and slot filling are performed to dynamically construct and update a formal user demand profile, which includes budget range, target city, vehicle preference and risk tolerance constraints.

[0023] The user demand profile is transformed into an executable structured query constraint. The tool calls the real-time vehicle source database through the interface framework and uses an incremental update constraint solving algorithm to iteratively filter and sort the candidate vehicle source set by matching degree.

[0024] Based on the query results of the knowledge graph, the real-time vehicle source and valuation data returned by the tool call, and the risk assessment results of the candidate vehicle sources, the big language model is driven to generate natural language answers that include recommendation reasons and targeted risk warnings.

[0025] During the conversation, the contact information provided by the user is identified by combining pattern matching and semantic understanding. After format standardization and confidence assessment, it is associated with the user's demand profile to generate structured sales leads.

[0026] Furthermore, the incremental update constraint solving algorithm is as follows: In multi-round dialogues, when users add or modify constraints, the system performs incremental filtering and reordering based on the candidate vehicle source set selected in the previous round, rather than recalculating the entire vehicle source database.

[0027] Furthermore, the construction of user demand profiles is dynamic. The system continuously integrates newly identified slot information in multi-round dialogues and can cover or correct previously confirmed constraints.

[0028] Furthermore, the risk assessment results are generated based on the vehicle condition fields of the target vehicle source, including the completeness of accident records, maintenance records, and mileage reasonableness. These are calculated through a preset risk scoring model and transformed into natural explanatory suggestions by a large language model.

[0029] Furthermore, the confidence assessment specifically includes: when multiple contact method candidates or ambiguous formats are identified, the system initiates a natural language confirmation dialogue to verify with the user.

[0030] Furthermore, the generated natural language answer must simultaneously include at least one core parameter of the recommended vehicle source, a recommendation reason based on knowledge graph and real-time data, and a risk warning for specific vehicle conditions.

[0031] Furthermore, the method can be deployed and applied to customer service scenarios on short video platforms, social media, or live streaming platforms to directly respond to users' inquiries about purchasing used cars on the relevant platforms.

[0032] According to a second aspect of the present invention, a smart customer service robot specifically designed for the used car industry is provided, the smart customer service robot comprising:

[0033] The knowledge graph construction and query module is used to construct and maintain the knowledge graph in the used car field and provide a graph query interface;

[0034] The tool invocation and service integration module is used to encapsulate and invoke the vehicle source database, valuation engine, and policy query service interfaces;

[0035] The dialogue management and context maintenance module is used to process user input and maintain the state of multi-turn dialogues;

[0036] The intent recognition and user profile building module is used to identify user intents and extract and update user demand profiles.

[0037] The constraint solving and vehicle source screening module is used to transform user demand profiles into constraints and to iteratively screen and sort vehicle sources.

[0038] The answer and risk generation module is used to integrate knowledge, data, and risk information to generate natural language answers.

[0039] The sales lead identification and management module is used to identify and verify user contact information and generate structured sales leads.

[0040] Furthermore, the tool invocation and service integration module includes an intent-tool mapping strategy, which mandates that when a specific user intent is identified, the corresponding business tool must be invoked, and the numerical range and field integrity of the returned data must be verified.

[0041] Furthermore, the sales lead identification and management module is integrated into the dialogue process. When it detects that a user expresses a desire for in-depth consultation, it automatically guides the user to provide contact information and associates it with the user demand profile already constructed in the current dialogue after identification.

[0042] The embodiments of the present invention have the following advantages:

[0043] This invention, through multi-dimensional business data collection and data cleaning processes including terminology standardization and jargon correction, combined with a tiered anonymization mechanism, protects user privacy and sensitive vehicle information. Based on a lightweight foundational model, fine-tuning training is conducted to ensure the model accurately adapts to business scenarios such as vehicle introductions and demand analysis. Human feedback reinforcement learning is then employed to optimize output preferences. Finally, a high-speed inference framework is used to deploy the model for efficient service. This invention effectively solves the problems of insufficient business adaptability of general-purpose models, high costs of long prompts, and data leakage risks. While reducing enterprise operating costs, it significantly improves the professionalism and security of customer service responses, accurately enhances business adaptability, and effectively ensures data security, avoiding the risk of leakage. Attached Figure Description

[0044] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0045] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0046] Figure 1 A schematic diagram of the logical structure of an intelligent customer service robot specifically designed for the used car industry, provided as an embodiment of the present invention;

[0047] Figure 2 This is a flowchart illustrating a method for constructing an intelligent customer service model specifically for the used car industry, as provided in an embodiment of the present invention. Detailed Implementation

[0048] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0049] Large Language Model (LLM):

[0050] This refers to next-generation large-scale pre-trained language models such as DeepSeek, which possess natural language understanding and generation capabilities and are used for dialogue understanding, reasoning, and answer generation.

[0051] Used car intelligent Q&A robot:

[0052] The intelligent dialogue system, deployed on new media channels such as used car websites, apps, short video platforms, live streaming platforms, and social media, is used to answer users' questions about used car selection, evaluation, configuration, usage costs, and vehicle condition risks, and assists in the selection and recommendation of car models.

[0053] Used Car Knowledge Graph:

[0054] The multi-relationship graph structure is formed by structuring and modeling information related to used cars (brand, model series, configuration, year, vehicle condition, mileage, price, city, emission standards, license plate restrictions, common faults, etc.) to support question-answering reasoning.

[0055] Multi-turn Dialogue:

[0056] In the continuous back-and-forth dialogue between the user and the robot, the user's needs are often gradually clarified over multiple rounds, such as budget, preferred car model, intended use, and whether they accept a car with a tampered odometer.

[0057] Intent Recognition:

[0058] Map users' natural language questions to predefined business intents, such as "budget estimation", "recommended car model", "estimate residual value", "compare two cars", "inquire about annual inspection / traffic restriction policies", etc.

[0059] Slot Filling:

[0060] Structured parameters, such as budget range, city of purchase, acceptable vehicle age, mileage limit, fuel type, and transmission type, are extracted from user multi-turn conversations for subsequent decision-making or retrieval.

[0061] ToolCalling:

[0062] The large language model automatically decides whether to call external services, such as vehicle source retrieval API, valuation service API, and city policy query API, based on the current intent, and then fills the results back into the generated answer.

[0063] Constraint Solving and Candidate Set Shrinking Algorithm:

[0064] The algorithm transforms user needs into formal constraints (budget, mileage, vehicle age, emission standards, city license plate restrictions, etc.), filters and sorts the vehicle source set multiple times, gradually reduces the candidate set, and produces interpretable recommendation results.

[0065] Risk warning generation module:

[0066] Based on vehicle condition fields (accident records, whether it is a major accident vehicle, whether it is a flood-damaged / fire-damaged vehicle, missing maintenance records, etc.), it automatically generates targeted risk alerts and re-inspection suggestions.

[0067] Contact information extraction / Mobile number extraction:

[0068] This refers to the process of automatically identifying contact information such as mobile phone numbers proactively provided by users in multi-turn dialogue texts on new media platforms, and standardizing, verifying, and confirming the format. This process supports robust recognition of various non-standard writing styles in natural language (such as spaces, Chinese characters and numbers, and colloquial descriptions), which are then used to generate sales leads.

[0069] Currently, in used car e-commerce platforms and new media scenarios, intelligent customer service or robots for used car-related questions mainly adopt the following types of technical solutions:

[0070] 1. Static Question Answering Solution Based on FAQ Templates

[0071] Frequently asked questions are pre-compiled into FAQs. Using keyword matching or simple similarity algorithms, the most similar question-and-answer pairs are found from the FAQs and returned directly when a user asks a question.

[0072] Advantages: Simple to implement and low deployment cost.

[0073] shortcoming:

[0074] Only a limited number of fixed questions can be answered;

[0075] Unable to understand users' personalized needs;

[0076] Unable to link to real-time vehicle availability, pricing, and other dynamic data.

[0077] 2. A task-oriented dialogue solution based on traditional intent classification and slot filling.

[0078] Using traditional machine learning or small deep learning models, the intent of the user's single-turn input is classified, and slot identification (such as budget, city, brand) is performed using methods such as rule / CRF / BiLSTM. Then, a fixed dialogue script is designed according to the business process.

[0079] On some platforms, the chatbot can guide users to input information such as budget, city, brand, and intended use, and then call the vehicle sourcing interface to return some recommended models.

[0080] shortcoming:

[0081] A large amount of manually labeled data is required;

[0082] Intent / slot configuration is cumbersome and difficult to maintain;

[0083] They have a weak understanding of complex issues and cross-turn dialogues.

[0084] 3. A "general question-answering" solution that easily integrates with large language models

[0085] Some platforms are trying to directly connect to a general large language model, sending user questions directly to the model, which then generates natural language answers.

[0086] To supplement used car knowledge, a small amount of car knowledge or FAQ text is often added to the prompts.

[0087] shortcoming:

[0088] The model struggles to combine real-time vehicle data and specific vehicle condition information, making it prone to producing answers that "seem reasonable but are actually wrong" (i.e., illusions).

[0089] There is no formal knowledge graph or structured data fusion;

[0090] There is a lack of stable and controllable tool calling frameworks;

[0091] There is a lack of specific constraint solving and recommendation logic for the used car scenario.

[0092] 4. Rule-based vehicle screening / valuation tool

[0093] Some platforms also offer "valuation tools" or "intelligent car selection" functions, where users fill in parameters (license plate number, mileage, model year, etc.) in a form, and the backend calculates the price or filters car sources through rules or valuation models.

[0094] shortcoming:

[0095] Disconnected from natural language communication;

[0096] Users need to actively fill out relatively complex forms;

[0097] It cannot dynamically update parameters or explain the reasons for recommendations based on dialogue, nor can it answer higher-level questions (such as "Is it suitable for family use?" or "What are the differences between these two cars?").

[0098] In summary, existing solutions either remain at the FAQ and rule level or simply "connect a large model," lacking a systematic and engineered knowledge integration, tool invocation, and constraint solving mechanism for the used car scenario.

[0099] Compared to the solution proposed in this application, existing technologies have at least the following problems in the intelligent question-and-answer scenario for used cars:

[0100] 1. Difficulty in providing accurate Q&A and recommendations based on real vehicle data.

[0101] The general-purpose big data model is completely unaware of the actual inventory, real-time prices, and specific vehicle conditions of used cars. Most existing simple integration methods cannot reliably call business interfaces such as vehicle retrieval and valuation, resulting in:

[0102] The answer does not match the current vehicle listings on the platform;

[0103] The price range is inaccurate or lacks sufficient basis;

[0104] It is difficult to achieve a closed loop of "Q&A + real-time car purchase".

[0105] 2. Insufficient ability to understand users' multi-turn, progressive expressions.

[0106] Users' true needs are often gradually revealed through multiple rounds of dialogue:

[0107] "Budget is around 150,000 yuan" → "Preferably an SUV with more space" → "Mainly for commuting in the city and taking the kids out on weekends" → "I'm hesitant to buy a car that has been in an accident, so it's best if it's well-maintained."

[0108] Traditional intent classification and slot filling solutions typically parse statements in a single round, making it difficult to integrate historical dialogue information and build a complete and updatable "user demand profile," resulting in inaccurate and insufficiently explained recommendation results.

[0109] 3. Lack of knowledge graph and reasoning capabilities specific to the used car sector.

[0110] The used car market involves a wide range of knowledge: differences in vehicle configurations, residual value patterns, common faults, target audiences, and city-specific license plate / traffic restriction policies.

[0111] Existing solutions either do not build knowledge graphs or only use simple tags, which cannot support complex question answering, for example:

[0112] "For family use and commuting, with a budget of 80,000 to 100,000 yuan, and no traffic restrictions in Shanghai, should I recommend a joint venture or a domestic brand?"

[0113] "What is the approximate resale value of this car that has been driven for 3 years and 80,000 kilometers? What price will it drop to after another two years?"

[0114] 4. Lack of constraint solving and candidate set shrinking algorithms tailored to the characteristics of the used car business.

[0115] In real-world business scenarios, user constraints are often numerous and interconnected:

[0116] Budget, vehicle age, mileage, engine displacement, emission standards, city of purchase, license plate type, etc.

[0117] Existing simple screening methods cannot handle multiple rounds of dynamic changes and combination constraints, and lack fine scoring and ranking of "matching quality", resulting in poor interpretability of recommendation results.

[0118] 5. Insufficient risk warning and explanation capabilities

[0119] The key to making a decision about a used car lies in recognizing and interpreting the risks, such as the risks of accident vehicles, odometer tampering, and uncertainties caused by missing maintenance records.

[0120] Existing solutions typically only return simple labels (such as "non-accident vehicle", "accident-involved", or "vehicle in good condition"), lacking:

[0121] A comprehensive assessment of the risk fields;

[0122] Natural language explanations and suggestions regarding risks (such as recommendations for listing testing items and which dimensions to consider);

[0123] Explanation of uncertainties in the case of missing data.

[0124] 6. The illusion of a generalized, large-scale model and the issue of business controllability.

[0125] Using a general, large-scale model to answer used car questions is easy:

[0126] Fabricating vehicle specifications or prices out of thin air;

[0127] Ignoring urban policy restrictions;

[0128] The response was not in accordance with the platform's rules.

[0129] Furthermore, many systems lack robust tool invocation strategies and result verification mechanisms, resulting in poor business controllability.

[0130] To solve the following problem:

[0131] 1. How to integrate a large language model with multiple data sources such as vehicle source database, valuation service, and city policy database in the context of intelligent Q&A for used cars, and generate Q&A results and vehicle recommendations based on real data.

[0132] 2. How to continuously and dynamically build and update user demand profiles (budget, purpose, city, preferences, risk tolerance, etc.) in multi-round dialogues, and transform them into executable formal constraints for screening vehicle sources.

[0133] 3. How to construct a knowledge graph for the used car industry and combine it with a large language model for joint reasoning to achieve accurate Q&A on issues such as vehicle model differences, applicable scenarios, residual value, and usage costs.

[0134] 4. How to design a "multi-round shrinkage and sorting algorithm" for the set of candidate used cars, which gradually converges to a small number of highly matched candidate cars while continuously adding user constraints and preferences, and provides an explainable recommendation reason.

[0135] 5. How to reduce the illusion of a large language model by using tool call control, result verification and risk warning mechanisms, ensure that the answer content is consistent with business data and business rules, and improve the reliability and controllability of the system.

[0136] refer to Figure 1 This invention discloses an intelligent customer service robot specifically for the used car industry. The system includes: a knowledge graph construction and query module 1; a tool call and service integration module 2; a dialogue management and context maintenance module 3; an intent recognition and user profile construction module 4; a constraint solving and vehicle source screening module 5; an answer and risk generation module 6; and a sales lead recognition and management module 7.

[0137] The system structure and method flow of the present invention will be described in detail below with reference to a typical embodiment.

[0138] 1. System Overall Architecture

[0139] The intelligent Q&A customer service robot system for used cars of this invention mainly includes the following functional modules:

[0140] 1. Dialogue Management and Context Maintenance Module

[0141] Responsible for receiving natural language messages from users through new media channels;

[0142] Maintain the state of multi-turn dialogues, including historical turns, parsed user constraints, recommended vehicle sources, and risks that have been alerted;

[0143] Construct inputs containing the necessary context for large language models.

[0144] 2. Intent Recognition and Slot Filling Module

[0145] Based on a large language model or a lightweight intent classifier, determine the intent of the current round of user speech;

[0146] Identify and extract information such as budget (value + currency unit), city, brand / car series / model, purpose (family / commuting / business / off-road), whether fuel consumption is important, and whether you can accept accident vehicles.

[0147] Merge and update the new slot information with the conversation history.

[0148] 3. Used Car Knowledge Graph Construction and Query Module

[0149] Extract entities (brand, model, powertrain, configuration items, etc.) and relationships (such as "suitable for family use" and "higher fuel consumption on highways") from data sources such as car manufacturer configuration tables, car review articles, and maintenance knowledge bases.

[0150] Build a graph database or graph structure;

[0151] It provides graph query interfaces by vehicle type, configuration, usage scenario, etc., for use during the question and answer process.

[0152] 4. Vehicle Source and Valuation Tool Call Module

[0153] It connects to multiple backend services within the platform, such as the vehicle source database, valuation engine, and city rule services;

[0154] Based on the current dialogue intent and the collected slot information, automatically generate API call parameters;

[0155] The call results are uniformly encapsulated and returned to the large language model as factual data.

[0156] 5. Constraint Solving and Candidate Vehicle Source Shrinking Module

[0157] Transform "budget, city, vehicle age limit, mileage limit, emission standard, and transmission type" into formal constraints;

[0158] The vehicle database was filtered and sorted multiple times.

[0159] In multi-turn dialogues, when users add or modify constraints, the current candidate set is incrementally updated instead of being fully recalculated, thus improving performance.

[0160] Output the sorted list of candidate cars and calculate the matching score between each car and the user profile.

[0161] 6. Answer generation and risk warning module

[0162] The candidate vehicle source data, knowledge graph query results, valuation information, risk assessment results, and dialogue context are input into the large language model.

[0163] Guided model generation:

[0164] A direct answer to user questions;

[0165] Vehicle recommendations with "reasons for recommendation";

[0166] Risk warnings and re-inspection suggestions for situations such as accidents, missing maintenance records, and high mileage.

[0167] 7. Tool Invocation Strategy and Result Verification Module

[0168] Pre - define a "tool list" for the large - language model, such as [Vehicle Source Query], [Valuation Query], [City Policy Query], [Knowledge Graph Query], etc.;

[0169] Use a dedicated Prompt to constrain the model to necessarily or preferably call the corresponding tool under a specific intention;

[0170] Perform basic verification on the data returned by the tool (field integrity, numerical range rationality), and trigger a fallback strategy (such as re - calling, prompting that it is temporarily unavailable) when obvious anomalies occur.

[0171] 8. Contact Information Recognition and Lead Generation Module

[0172] Monitor the user's speech in real - time throughout the conversation. When it is detected that the user has the intention of "being willing to consult further and being willing to be contacted", guide the user to actively leave contact information (such as a mobile phone number);

[0173] Perform candidate identification of mobile phone numbers for the user input content, and comprehensively use:

[0174] Detection of pre - defined number patterns (regular expressions);

[0175] The large - language model's semantic understanding of the context (such as "This is my phone number", "The WeChat ID and phone number are the same", etc.);

[0176] Joint analysis of multi - round historical messages to identify whether the user is providing a new number, modifying the number, or just repeating;

[0177] Perform format standardization and verification on the identified mobile phone number, including length verification, segment legality verification, removing middle spaces / symbols, and restoring oral or Chinese -夹杂expressions such as "1 three 80××××××××" to a standard digital string;

[0178] When the recognition confidence is insufficient or there are multiple candidate numbers, the robot makes a secondary confirmation through natural language, for example: "To facilitate the consultant to contact you, confirm that your mobile phone number is 138××××5678?";

[0179] After the user confirms, associate the mobile phone number with the user portrait (budget, city, vehicle model requirements, etc.) that has been constructed in the current conversation, generate a structured sales lead, and push it to the lead system or CRM for subsequent manual follow - up.

[0180] Corresponding to the above - disclosed intelligent customer service robot dedicated to the used - car field, the embodiment of the present invention also discloses a method for constructing an intelligent customer service model dedicated to the used - car field. The following details a method for constructing an intelligent customer service model dedicated to the used - car field disclosed in the embodiment of the present invention in combination with the above - described intelligent customer service robot dedicated to the used - car field.

[0181] refer to Figure 2 This invention discloses a method for constructing an intelligent customer service model specifically for the used car industry. The method includes: constructing and maintaining a knowledge graph for the used car industry, wherein the knowledge graph integrates vehicle brand, model series, configuration, model year, vehicle condition, market price, residual value, city traffic restriction policies and common fault information in a structured manner.

[0182] Establish a tool call interface framework to connect to and encapsulate a real-time vehicle source database, a vehicle valuation service engine, and a city policy query service;

[0183] Build a dialogue management module to receive natural language input from users on new media channels and maintain multi-turn dialogue contexts that include historical dialogues, identified user demand profiles, and results of invoked tools.

[0184] Based on the large language model and combined with the dialogue context, user intent recognition and slot filling are performed to dynamically construct and update a formal user demand profile, which includes budget range, target city, vehicle preference and risk tolerance constraints.

[0185] The user demand profile is transformed into an executable structured query constraint. The tool calls the real-time vehicle source database through the interface framework and uses an incremental update constraint solving algorithm to iteratively filter and sort the candidate vehicle source set by matching degree.

[0186] Based on the query results of the knowledge graph, the real-time vehicle source and valuation data returned by the tool call, and the risk assessment results of the candidate vehicle sources, the big language model is driven to generate natural language answers that include recommendation reasons and targeted risk warnings.

[0187] During the conversation, the contact information provided by the user is identified by combining pattern matching and semantic understanding. After format standardization and confidence assessment, it is associated with the user's demand profile to generate structured sales leads.

[0188] Taking "a user inquiring through new media channels and requesting a car model recommendation" as an example, the system process includes, but is not limited to, the following steps:

[0189] 1.S1: Receive and parse the user's initial question.

[0190] A user typed: "My budget is around 100,000 yuan. I'd like to buy a reliable family car. Any recommendations?"

[0191] The dialogue management module records this round of messages, and the intent recognition module identifies it as the intent of "intelligent car selection recommendation".

[0192] The slot filling module identified: Budget ≈ 100,000, Use = Household, Other slots are currently empty.

[0193] 2.S2: Constructing preliminary user profiles and constraint sets

[0194] Preliminary constraints are formed:

[0195] Budget cap = 100,000 + tolerance (e.g., 10%)

[0196] Purpose = Household use;

[0197] User profiles record that users prioritize stability and affordability.

[0198] 3.S3: Invoke vehicle source retrieval and knowledge graph query

[0199] The tool calling module filters several brands / models (such as Sylphy, Corolla, Lavida, Excelle, etc.) from the vehicle source database based on budget constraints.

[0200] The knowledge graph module queries the knowledge tags of these car models in dimensions such as "family use", "stability", and "fuel consumption".

[0201] 4.S4: Candidate Vehicle Source Narrowing and Scoring

[0202] The constraint solving module performs the first round of sorting on the candidate vehicle source set:

[0203] The rating factors include: the degree to which the price deviates from the target budget, vehicle age, mileage, and completeness of maintenance records;

[0204] Generate a batch of candidate cars with high scores and calculate their matching degree with user profiles.

[0205] 5.S5: Multi-round supplementary constraints and iterative recommendation

[0206] Based on the current results, the robot interacts with the user in natural language: "Do you prefer joint ventures or domestic brands? How many years old can you accept in a car?"

[0207] A user might reply: "Preferably a joint venture, no more than 6 years old, and automatic transmission."

[0208] Slot filling module update constraints: vehicle age ≤ 6 years, transmission = automatic, brand attribute = joint venture.

[0209] The constraint solving module incrementally shrinks and re-scores the existing candidate set.

[0210] 6.S6: Risk Assessment and Alert Generation

[0211] For each candidate vehicle, the risk assessment logic comprehensively considers fields such as accident records, whether it was a major accident, whether it was flooded / fired, whether there are abnormal mileage markers, and whether the maintenance records are complete.

[0212] Output the risk level and recommended areas of concern;

[0213] The answer generation module naturally incorporates these risk warnings into the recommended text, for example:

[0214] "The vehicle has no record of major accidents, but its maintenance records are incomplete. We recommend focusing on checking for leaks in the chassis and engine."

[0215] 7.S7: Final Questions and Recommendations

[0216] The large language model generates a language that includes both candidate vehicle sources and knowledge graph information.

[0217] 1-3 key recommended models;

[0218] The core parameters of each vehicle (model year, mileage, price, emission standards, etc.);

[0219] Reasons for recommendation (e.g., space, fuel consumption, stability);

[0220] Risk warnings and re-inspection recommendations;

[0221] The output is displayed on the new media chat interface, allowing users to continue asking questions or compare other car sources.

[0222] S8: Contact Information Extraction and Clue Generation

[0223] Throughout the multi-round conversation, the conversation management module monitors whether the user expresses intentions such as "wanting to know more" or "wanting someone to contact them";

[0224] When this type of intent is detected, the answer generation module naturally guides the user to leave their contact information in the response, for example: "If it's convenient, you can tell me your mobile number, and a dedicated consultant will explain it to you in detail later."

[0225] When a user enters text containing a phone number, the contact information recognition and lead generation module:

[0226] By using number pattern matching and semantic analysis of a large language model, candidate mobile phone numbers are extracted from text;

[0227] Perform length and number segment verification to restore numbers containing non-standard writing such as spaces, hyphens, and Chinese numerals to the standard format;

[0228] If there is any ambiguity or insufficient confidence, an automatic confirmation conversation will be initiated to verify the final number with the user;

[0229] After the number is confirmed, information such as "phone number + user profile + current conversation summary + recommended vehicle source" is structured and written into the lead system to form a lead record that can be directly used by sales consultants, avoiding manual copying and input errors and reducing lead loss.

[0230] 3. Tool Invocation Control and Result Verification Details

[0231] 1. Intent-to-Tool Mapping Table

[0232] [Valuation Consultation] → Must use [Valuation Tools];

[0233] [Whether there are traffic restrictions / whether a license plate can be registered] → You must use the [City Policy Inquiry] function;

[0234] [Vehicle Configuration / Parameter Comparison] → Prioritizes [Knowledge Graph Query];

[0235]

Car Selection Recommendation

Vehicle Source Search

Knowledge Graph Query

[0236] 2. Automatic validation of call results

[0237] Regarding the returned vehicle source data:

[0238] Is the price range legal?

[0239] Are the vehicle's age and mileage within a reasonable range (e.g., there is no anomaly such as "500,000 kilometers in 1 year");

[0240] Regarding valuation data:

[0241] Does it deviate too much from the market average price of similar vehicles?

[0242] When an anomaly occurs, the strategy module driven by the large language model provides explanations or prompts to the user that the data is temporarily unavailable, ensuring that the business remains controllable.

[0243] The embodiments of the present invention have the following advantages:

[0244] 1. Achieve deep integration of "real data + intelligent question answering"

[0245] By constructing a tool call framework that links the vehicle source database, valuation service, city policy database, and large language model, the question and answer results no longer remain at the level of "general" experience and advice, but are based on real-time available vehicle sources and real prices, greatly improving the reliability and practicality of the answers.

[0246] 2. Supports dynamic requirement modeling and constraint solving in multi-turn dialogue scenarios.

[0247] The system continuously completes and updates user profiles in multi-round dialogues, automatically transforms natural language requirements into formal constraints, and gradually converges to high-matching vehicle sources through an iterative shrinking algorithm of the candidate set.

[0248] Compared to traditional single-round intent recognition solutions, it is better adapted to users' real communication habits, and the recommendation hit rate is significantly improved.

[0249] 3. Domain reasoning ability combining used car knowledge graphs

[0250] Joint reasoning based on knowledge graphs and large language models can solve complex problems such as "applicable groups", "car usage scenarios", "residual value trends" and "comparison of models in the same price range".

[0251] It reduces the illusion caused by a lack of structured knowledge in purely large model responses.

[0252] 4. Provide explainable vehicle recommendations and risk warnings.

[0253] The candidate vehicle matching score and risk assessment results are presented in natural language, so users not only know "which car is recommended", but also "why it is recommended" and "what to pay attention to".

[0254] Improving user trust helps shorten the decision-making cycle and increase conversion rates.

[0255] 5. Improve system controllability and reduce business risks caused by illusions.

[0256] Constraining and correcting the output of large language models is achieved through explicit tool call strategies and interface return result verification mechanisms.

[0257] Avoid fabricating key business information such as vehicle condition, price, and policies in the model, and reduce the impact of incorrect questions and answers on user decisions.

[0258] 6. Possesses good scalability and portability.

[0259] The core of this solution lies in the framework of "dialogue management + tool call + constraint solving + knowledge graph reasoning". The vehicle source database, valuation model, city rules, etc. can be replaced according to different platforms.

[0260] It has taken the lead in the used car vertical sector and can be further extended to new car sales, financial solution matching, after-sales maintenance consultation and other scenarios.

[0261] 7. Improve the accuracy of contact information identification and lead conversion efficiency.

[0262] By automatically identifying, standardizing, and confirming users' mobile phone numbers in multi-round dialogues on new media, this solution reduces the need for manual copying and input, significantly lowering the probability of incorrect numbers and missed clues.

[0263] By automatically linking mobile phone numbers with user profile information such as budget, city, and car model preferences, sales consultants can obtain a complete background of user needs before contacting them, thereby improving the hit rate and conversion efficiency of subsequent phone or offline communications.

[0264] This invention provides an integrated framework for "dialogue + tool invocation + constraint solving" for used car Q&A scenarios.

[0265] Transform the user's natural language requirements generated from multi-turn dialogues into formal constraints;

[0266] The framework integrates vehicle sourcing database, valuation tools, and policy database through a unified tool call system.

[0267] An incremental candidate vehicle source shrinkage algorithm is used to achieve dynamic updating and convergence of recommendation results.

[0268] This invention provides a joint reasoning mechanism combining a used car knowledge graph and a large language model.

[0269] Knowledge graphs are used to represent knowledge in areas such as vehicle configuration, applicable scenarios, residual value, and common faults.

[0270] By calling the graph query interface through a large language model, the structured results are combined with natural language generation to achieve high-quality answers to complex questions in the used car industry.

[0271] This invention utilizes risk assessment and natural language risk warning generation technologies.

[0272] A risk scoring model is built based on vehicle condition fields, which comprehensively evaluates information such as accident records, maintenance records, and mileage anomalies.

[0273] The risk score is converted into user-oriented explanatory suggestions and re-examination reminders using a large language model.

[0274] This invention employs a tool invocation strategy and result verification mechanism:

[0275] The corresponding tool is invoked or prioritized based on the intent type, and the returned data is validated for numerical range, field integrity, etc.

[0276] Triggering rollback strategies and interpretive processing for abnormal situations significantly reduces the impact of large language model illusions on business operations.

[0277] This application also features a method for building and dynamically updating multi-round conversational profiles for the used car business:

[0278] By continuously integrating new constraints and preferences in multi-round dialogues, a user profile that can be executed by machines can be constructed.

[0279] This profile serves as the core driver for selecting and ranking candidate vehicles, achieving high matching accuracy and explainable recommendations.

[0280] This application also features a robust method for identifying and verifying mobile phone numbers in new media dialogue scenarios:

[0281] To address the common non-standard number writing styles (such as mixing Chinese characters and numbers, colloquial sentence breaks, and adding "v" to express the same number) in scenarios such as short videos, live streaming, and social media, a mobile phone number recognition method combining pattern matching, semantic understanding, and multi-round context analysis is proposed.

[0282] Through confidence assessment and natural language secondary confirmation mechanisms, the contact information ultimately written into the lead system is ensured to be accurate and reliable, achieving an automated closed loop from dialogue to sales lead.

[0283] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for constructing an intelligent customer service model specifically for the used car industry, characterized in that, The method includes: Construct and maintain a knowledge graph for the used car industry, which integrates vehicle brand, model, configuration, year, vehicle condition, market price, residual value, city traffic restriction policies and common fault information in a structured manner; Establish a tool call interface framework to connect to and encapsulate a real-time vehicle source database, a vehicle valuation service engine, and a city policy query service; Build a dialogue management module to receive natural language input from users on new media channels and maintain multi-turn dialogue contexts that include historical dialogues, identified user demand profiles, and results of invoked tools. Based on the large language model and combined with the dialogue context, user intent recognition and slot filling are performed to dynamically construct and update a formal user demand profile, which includes budget range, target city, vehicle preference and risk tolerance constraints. The user demand profile is transformed into an executable structured query constraint. The tool calls the real-time vehicle source database through the interface framework and uses an incremental update constraint solving algorithm to iteratively filter and sort the candidate vehicle source set by matching degree. Based on the query results of the knowledge graph, the real-time vehicle source and valuation data returned by the tool call, and the risk assessment results of the candidate vehicle sources, the big language model is driven to generate natural language answers that include recommendation reasons and targeted risk warnings. During the conversation, the contact information provided by the user is identified by combining pattern matching and semantic understanding. After format standardization and confidence assessment, it is associated with the user's demand profile to generate structured sales leads.

2. The method for constructing an intelligent customer service model specifically for the used car industry as described in claim 1, characterized in that, The incremental update constraint solving algorithm is as follows: In multi-turn dialogues, when users add or modify constraints, the system performs incremental filtering and reordering based on the candidate vehicle source set selected in the previous round, rather than recalculating the entire vehicle source database.

3. The method for constructing an intelligent customer service model specifically for the used car industry as described in claim 1, characterized in that, The construction of user demand profiles is dynamic. The system continuously integrates newly identified slot information in multi-round dialogues and can cover or correct previously confirmed constraints.

4. The method for constructing an intelligent customer service model specifically for the used car industry as described in claim 1, characterized in that, The risk assessment results are generated based on the vehicle condition fields of the target vehicle, including the completeness of accident records, maintenance records, and mileage reasonableness. They are calculated through a preset risk scoring model and transformed into natural explanatory suggestions by a large language model.

5. The method for constructing an intelligent customer service model specifically for the used car industry as described in claim 1, characterized in that, The confidence assessment specifically includes: when multiple contact method candidates are identified or the format is ambiguous, the system initiates a round of natural language confirmation dialogue to verify with the user.

6. The method for constructing an intelligent customer service model specifically for the used car industry as described in claim 1, characterized in that, The generated natural language answer must simultaneously include at least one core parameter of the recommended vehicle source, a recommendation reason based on knowledge graph and real-time data, and a risk warning for specific vehicle conditions.

7. The method for constructing an intelligent customer service model specifically for the used car industry as described in claim 1, characterized in that, The method is deployed and applied to customer service scenarios on short video platforms, social media, or live streaming platforms to directly respond to users' inquiries about purchasing used cars on the relevant platforms.

8. An intelligent customer service robot specifically designed for the used car industry, characterized in that: The intelligent customer service robot includes: The knowledge graph construction and query module is used to construct and maintain the knowledge graph in the used car field and provide a graph query interface; The tool invocation and service integration module is used to encapsulate and invoke the vehicle source database, valuation engine, and policy query service interfaces; The dialogue management and context maintenance module is used to process user input and maintain the state of multi-turn dialogues; The intent recognition and user profile building module is used to identify user intents and extract and update user demand profiles. The constraint solving and vehicle source screening module is used to transform user demand profiles into constraints and to iteratively screen and sort vehicle sources. The answer and risk generation module is used to integrate knowledge, data, and risk information to generate natural language answers. The sales lead identification and management module is used to identify and verify user contact information and generate structured sales leads.

9. The intelligent customer service robot specifically designed for the used car industry as described in claim 8, characterized in that, The tool invocation and service integration module includes an intent-tool mapping strategy, which mandates that when a specific user intent is identified, the corresponding business tool must be invoked, and the numerical range and field integrity of the returned data must be verified.

10. The intelligent customer service robot specifically designed for the used car industry as described in claim 8, characterized in that, The sales lead identification and management module is integrated into the dialogue process. When it detects that a user expresses a desire for in-depth consultation, it automatically guides the user to provide contact information and associates it with the user demand profile already built in the current dialogue after identification.

Images