Retail store unmanned shopping guide method and system based on large language model

Abstract

The invention discloses a retail store unmanned shopping guide method and system based on a large language model, and the method comprises the steps: carrying out the directional training of the large language model based on a pre-constructed structured knowledge base and customer data collection and arrangement, and providing information and technical support for intelligent shopping guide interaction; through personalized recommendation, an optimal product recommendation result is provided for customers; and through feedback optimization, the prediction capability of the model is continuously optimized, so that the system can continuously perform self-learning optimization, and the intelligent shopping guide capability is continuously enhanced. The method is widely applicable to retail stores of various brands, and particularly has a remarkable application prospect in the field of electronic product sales. And by deploying the unmanned shopping guide system, the retailer can reduce the labor cost and improve the store operation efficiency. In addition, the system has a highly intelligent recommendation capability and a quick response characteristic, so that the shopping experience of customers can be effectively improved, and the sales efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of sales technology, and in particular to an unmanned shopping guide method and system for retail stores based on a large language model. Background Art

[0002] With the rapid development of artificial intelligence technology, the retail industry is gradually evolving towards intelligent and unmanned directions. In traditional retail scenarios, sales staff need to introduce products to consumers face to face and provide shopping guide services. They need to understand a lot of product information and provide professional answers to various questions raised by different customers. However, manual shopping guides have problems such as high training costs, low efficiency, and inconsistent service quality. At the same time, modern consumers are increasingly inclined to self-service shopping and hope to obtain instant and accurate product recommendations and consulting services through smart devices. Therefore, the development of efficient, intelligent, and self-learning unmanned shopping guide methods is a demand of the future commercial retail industry. Summary of the invention

[0003] The purpose of the present invention is to provide an unmanned shopping guide method and system based on a large language model. By learning and processing a large number of sales scripts, product features and frequently asked questions, the system can provide customers with accurate and personalized product recommendations, improve the shopping experience, and ultimately achieve an efficient unmanned shopping guide mode.

[0004] To achieve the above purpose, the present invention designs an unmanned shopping guide method for retail stores based on a large language model. Based on the constructed structured knowledge base, the large language model is trained in a targeted manner to provide information and technical support for intelligent shopping guide interaction. The method includes: S1. Construct a structured knowledge base, including data collection and organization. The data collection (or knowledge base) includes several items such as product feature data, function description data, consumer feedback data, and common questions. Data organization includes storing the collected data in a structured form in the knowledge base for subsequent query and call. The structure of the knowledge base may include fields such as product ID, brand, model, specification, price, function description, user feedback, common questions and answers, etc.

[0005] S2. Targeted training of large language models. By introducing a large amount of sales tactics, product pros and cons analysis, and consumer demand in different scenarios, the system generates intelligent conversation capabilities based on natural language processing and responds to consumers' questions in real time. S3, smart shopping guide interaction: customers interact with the system through smart terminal devices in stores. The system uses a large language model to generate natural and coherent product recommendations and consultation answers based on the customer's voice and / or text input, and displays detailed information, promotions, usage suggestions, etc. of related products in real time; S4, Personalized recommendation: The system uses a large language model to screen and match existing products based on the needs or preferences entered by customers, such as desired product usage, price range, brand preference, etc., to provide customers with the best product recommendations; S5. Feedback optimization: The system records every customer interaction and recommendation result, and uses feedback data to optimize the model's predictive capabilities to continuously improve the system's intelligence level.

[0006] Furthermore, the product characteristic data includes several items such as the product's brand, model, specifications, price, color, applicable scenarios, main functions, etc.; the functional description data includes several items such as instructions for use, technical features, special functions (such as waterproof, fast charging), etc., which are descriptions of the core functions of each product; the consumer feedback data includes several items such as the product's advantages, disadvantages, and improvement suggestions, etc., and consumers' evaluation and usage experience of the product. These data come from the product's official evaluation page, third-party e-commerce platforms, etc.; the frequently asked questions include several items such as product comparison, usage scenario suggestions, and technical parameter explanations.

[0007] Furthermore, the sales talk includes the communication language and skills commonly used by retail staff when receiving customers, which are embedded in the model in the form of natural language, so that the system can simulate the real shopping guide communication method and improve the interactive experience; the sales talk also includes welcome talk, demand confirmation talk, product recommendation talk, purchase promotion talk, etc. The product advantages and disadvantages analysis includes a summary of each product's core performance, features, applicable population, and shortcomings; The consumer demands include that different consumer groups may focus on different scenario demands when purchasing.

[0008] Furthermore, the method for generating the intelligent dialogue capability includes: S21. Knowledge base preprocessing: First, the data on sales pitches, product advantages and disadvantages, and consumer needs are processed into structured text and converted into a format that the model can recognize. The data is embedded into the model in the form of tags and / or keywords, so that the model can quickly retrieve relevant information based on the consumer's questions. S22, Intent recognition and semantic matching: The model first identifies the consumer's question intent through natural language processing (NLP) technology; S23, Conversation generation capability: Based on the pre-trained and targeted fine-tuned models, sales scripts and product feature data are embedded into the language generation model; S24, Multi-turn dialogue management: The model also uses a multi-turn dialogue management module to ensure continuous and accurate responses when consumers have further questions or change their needs; Among them, step S24 is an optional step.

[0009] Furthermore, the intelligent shopping guide interaction method includes: S31. Customer demand collection: The system first collects customer input through the terminal device, including voice and / or text; S32, Intent recognition and keyword extraction: Through intent recognition technology, the system analyzes customer needs; S33, calling the large language model to generate recommendations or answers: After the intent and keywords are determined, the system calls the trained large language model to generate recommended content; S34. Real-time feedback and display: The generated recommendations and answers are transmitted to the terminal device in real time. The system displays several items such as product details, promotional activities or usage suggestions in a graphic and textual manner.

[0010] Furthermore, the large language model uses the following methods to ensure natural and coherent recommendations and answers: S331, context processing: If the customer's input involves multiple rounds, the model will call on the context information to maintain coherent communication; S332, Knowledge base retrieval and matching: The model retrieves content matching the keywords from the knowledge base and integrates several items of information such as the core features, advantages and disadvantages, and applicable scenarios of the product into the recommendation language; S333. Generate natural language dialogue: By processing the input requirements and knowledge base content, introduce preset sales scripts, and generate language that is closer to the style of real shopping guides, making the responses appear natural and smooth.

[0011] Furthermore, the personalized recommendation method includes: S41. Demand analysis and customer classification: The system first receives the customer's input and identifies the customer's intention and specific needs through the semantic analysis module. While analyzing the needs, the system divides the customers into two categories: new customers and old customers; S42, screening and matching strategy of large language model: Based on the demand analysis results, the system calls the large language model to screen and match existing products; S43, personalized recommendation for old customers: For old customers identified by the system, the personalized recommendation module will call the customer's historical preference data to make recommendations that are more in line with their habits; S44, display of recommendation results and comparison of multiple options: After the system screens out the products that best meet the customer's needs and preferences, the recommendation results are displayed on the terminal device, supporting multiple option comparison display to help customers evaluate products from multiple angles; Among them, steps S43 and S44 are optional steps.

[0012] Furthermore, the screening and matching process includes any number of the following steps S421 to S423, and there is no strict order: S421, label and keyword matching: the product data in the knowledge base has been labeled according to labels such as function, brand, price range, and purpose. The large language model will identify the keywords entered by customers and match them with the knowledge base labels to screen out products that meet the needs; S422, Multi-condition screening: For complex requirements, the model builds a multi-condition screening tree to gradually filter out products that do not meet the conditions from the knowledge base and retain the product set that meets all conditions; S423, Similarity matching and weighted recommendation: For fuzzy requirements, the large language model performs similarity matching based on product characteristics and sorts the results by weighted scores of camera performance to ensure the best recommendation effect; The personalized recommendation for old customers includes any number of the following steps S431 to S433, and there is no strict order: S431. Historical purchase records and access data: The system will extract the purchase records and historical browsing records of old customers, and conduct secondary screening based on the customer's preferences in product selection in different categories; S432. Fusion matching of demand and historical preferences: When screening, the large language model not only considers the customer's current demand, but also adds weight to the customer's historical preferences; S433, Intelligent recommendation optimization and intention expansion: For old customers who have not clearly expressed their specific needs, the system automatically generates further guiding questions through a large language model.

[0013] Furthermore, the feedback data types of the feedback optimization include any several items of interactive behavior data, recommendation result data, purchase conversion data, customer feedback and evaluation, customer portrait and preference data.

[0014] On the other hand, a retail store unmanned shopping guide system based on a large language model is a system constructed according to any of the above-mentioned retail store unmanned shopping guide methods based on a large language model, including a data collection and organization module, a large language training module, an intelligent shopping guide interaction module, a personalized recommendation module and a feedback optimization module that are sequentially and closed-loop connected, and the feedback optimization module connects the data collection and organization module to form a closed loop.

[0015] The advantages and beneficial effects of the present invention are as follows: the present invention provides an unmanned shopping guide method for retail stores based on a large language model, which can provide customers with real-time product recommendations and consulting services through intelligent terminal devices. The method not only solves the shortcomings of traditional manual shopping guides, but also improves the personalization and accuracy of shopping guide services through intelligent means, and has a wide range of market application value and technical advantages.

[0016] The present invention is widely applicable to various brand retail stores, especially in the field of electronic product sales. By deploying an unmanned shopping guide system, retailers can reduce labor costs and improve store operation efficiency. In addition, since the system has highly intelligent recommendation capabilities and fast response characteristics, it can effectively enhance customer shopping experience and improve sales efficiency. In the future, with the continuous advancement of large language model technology, the system is expected to be promoted in more retail scenarios, such as fashion, home furnishing and other industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of an unmanned shopping guide method for retail stores based on a large language model; Figure 2 It is a connection diagram of an unmanned shopping guide system for retail stores based on a large language model. DETAILED DESCRIPTION

[0018] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0019] Embodiment 1: The present invention designs a retail store unmanned shopping guide method based on a large language model, which is characterized in that the large language model is trained in a targeted manner based on a constructed structured knowledge base to provide information and technical support for intelligent shopping guide interaction, such as Figure 1 As shown, the method includes: S1. Construct a structured knowledge base, including data collection and organization. The data collection (or knowledge base) includes several items such as product feature data, function description data, consumer feedback data, and common questions. Data organization includes storing the collected data in a structured form in the knowledge base for subsequent query and call. The structure of the knowledge base may include fields such as product ID, brand, model, specification, price, function description, user feedback, common questions and answers, etc.

[0020] Regarding the structured processing of diversified product information: The retail industry has a wide variety of products, and information such as specifications, models, and prices is complex and changeable. How to efficiently structure and process and store them for subsequent training and reasoning of large language models is an important challenge for this system. It is necessary to ensure the accuracy and consistency of the data, especially since multiple versions or configuration items (such as color, memory capacity) of the same product may have different sales logic. At the same time, this information must be updated in real time to ensure that the product information obtained by consumers is the latest. This embodiment achieves efficient structured processing of product information by establishing a data standardization process and a dynamic update mechanism:

[0021] Product data standardization: Design a unified data format and labeling system to ensure that all product information (such as specifications, models, prices, etc.) can be stored in the same format so that the model can be called uniformly; Version and configuration management: For different versions and configurations of the same product (such as color, memory capacity), use associated fields to mark them in the database so that the system can identify the differences between different configurations; Real-time updates: Introduce an automated update mechanism to synchronize and update data in real time from the brand’s API interface or internal data source to ensure that customers have access to the latest product information.

[0022] S2. Targeted training of large language models. By introducing a large amount of sales tactics, product pros and cons analysis, and consumer demand in different scenarios, the system generates intelligent conversation capabilities based on natural language processing and responds to consumers' questions in real time. Regarding the targeted training and optimization of large language models: Large language models are usually general-purpose and require targeted training to adapt to specific tasks in retail scenarios, such as product recommendations and consumer Q&A. How to improve the model's accurate grasp of specific product features and sales tactics without reducing its ability to understand natural language is a technical challenge.

[0023] Targeted training requires a large amount of industry-related data, and it is also necessary to prevent the model from "forgetting" after multiple trainings (i.e., the model's ability to understand general language decreases). In addition, product features, promotional information, etc. may change frequently, and how to dynamically update while maintaining the model's efficiency is also a major challenge. This embodiment uses small sample targeted fine-tuning and continuous learning technology to adapt the model to specific retail scenario needs.

[0024] Targeted fine-tuning: Perform small sample fine-tuning on the large language model, using sales scripts, product features, and other content as fine-tuning data to improve the model's adaptability to retail scenarios without sacrificing natural language understanding capabilities.

[0025] Dynamic update mechanism: Design an automated training and update process to incrementally update the model with new product features and promotion information as needed to avoid frequent retraining.

[0026] S3, intelligent shopping guide interaction, customers interact with the system through the store's intelligent terminal devices. The system uses the large language model to generate natural and coherent product recommendations and consultation answers based on the customer's voice and / or text input, and displays detailed information, promotional activities, usage suggestions and other content of related products in real time; this embodiment is designed to collect voice and text input at the same time, or only accept any one of them for intelligent shopping guide interaction; Regarding the hardware compatibility and user experience of smart terminals: There may be many types of smart terminal devices in retail stores, including tablets, self-service terminals, etc. The system not only needs to be compatible and run on different hardware platforms, but also needs to ensure the smoothness of the graphical interface, the accuracy of voice recognition, and instant feedback of recommendation results. This places high demands on hardware performance and network bandwidth, especially when the store traffic is large, the system must have high concurrent processing capabilities. This embodiment builds a cross-platform compatible architecture and optimizes the terminal interface and interactive experience.

[0027] Cross-platform adaptation: Use cross-platform development tools to ensure that the system can run smoothly on different smart terminal devices and ensure a consistent user interface; Interface optimization and response speed improvement: Use lightweight UI design and front-end framework to reduce system resource usage; introduce cache technology and distributed architecture to ensure terminal response speed; High concurrent processing capability: During peak store traffic hours, the system supports high concurrent access through load balancing and distributed computing, ensuring a smooth real-time interactive experience for multiple users.

[0028] S4, Personalized recommendation: The system uses a large language model to screen and match existing products based on the needs or preferences entered by customers, such as desired product usage, price range, brand preference, etc., to provide customers with the best product recommendations; Regarding the recommendation and comparative analysis issues across product categories: When making a purchase, customers may want to compare products from multiple brands or categories (such as smartphones and smart watches). How to ensure that the large language model can perform comparative analysis across categories and give reasonable recommendations and explanations. The model needs to have sufficient reasoning ability to understand the differences and connections between different product categories, and provide accurate and logical comparative information when consumers ask cross-category questions. At the same time, the model also needs to adapt to the unique attributes of each product and be able to switch seamlessly between multiple categories. This embodiment introduces a cross-category recommendation strategy to enhance the reasoning ability of the model to achieve comparative analysis.

[0029] Category labeling: Add unique labels to each product category and build a "product attribute vector" so that the model can clearly identify the main features and differences of different categories during comparative analysis; Multi-task learning: Through multi-task learning technology, the model's adaptability in multi-category products is improved, so that when receiving cross-category demands, it can effectively identify and analyze the different advantages of each category; Product comparison function: Design a "comparative analysis" function in the system interface, allowing customers to select multiple categories at a time, displaying differentiation parameters in real time to help them make choices between multiple products.

[0030] S5. Feedback optimization: The system records every customer interaction and recommendation result, and uses feedback data to optimize the model's predictive capabilities to continuously improve the system's intelligence level.

[0031] Preferably, the product characteristic data includes several items such as the product's brand, model, specifications, price, color, applicable scenarios, main functions, etc.; the function description data includes several items such as instructions for use, technical features, special functions (such as waterproof, fast charging), etc., which are descriptions of the core functions of each product; the consumer feedback data includes several items such as the product's advantages, disadvantages, and improvement suggestions, etc., and consumers' evaluation and usage experience of the product, and these data come from the product's official evaluation page, third-party e-commerce platforms, etc.; the frequently asked questions include several items such as product comparison, usage scenario suggestions, and technical parameter explanations.

[0032] Preferably, the sales speech includes the communication language and skills commonly used by retailers when receiving customers, which has the characteristics of guidance, professionalism, and attractiveness, and is intended to help customers understand the product, meet their needs, and promote purchasing behavior; in the AI ​​shopping guide system, the sales speech will be embedded in the model in the form of natural language, so that the system can simulate the real shopping guide communication method and enhance the interactive experience; the sales speech in this embodiment also includes welcome speech (such as "Hello, welcome to our store! Do you want to know about mobile phones, tablets, or other products today?"), demand confirmation speech (such as "Which aspect do you prefer for mobile phones? For example, photo taking, game performance, or battery life?"), product recommendation speech (such as "Based on your needs, I recommend you to try Huawei Mate 50 Pro. Its photo taking effect is excellent in night scenes and outdoor scenes, and the battery life is also quite powerful."), and purchase promotion speech (such as "This product is on sale today, and you can enjoy a discount price. If you are satisfied with this phone, we can help you go through the purchase procedures at any time.") and other items;

[0033] The product advantages and disadvantages analysis includes a summary of the core performance, features, applicable groups and shortcomings of each product. These analysis contents enable the system to provide more professional and comprehensive information when answering consumer questions. For example, the advantages of Huawei Mate 50 Pro include excellent photo effects, excellent night scene performance, long battery life, support for fast charging and 5G networks, etc. The disadvantages include a heavy body, which may be a burden for users who hold it for a long time, and a high price, which is suitable for consumers with high performance requirements, etc.; The consumer demands include that different consumer groups may focus on different scenario demands when purchasing, such as taking travel photos, playing large-scale mobile games, office use, etc.

[0034] By analyzing these needs, the system can quickly match appropriate product recommendations when consumers raise corresponding scenario needs.

[0035] Preferably, the method for generating the intelligent dialogue capability includes: S21. Knowledge base preprocessing: First, the data on sales tactics, product advantages and disadvantages, and consumer needs are processed into structured text and converted into a format recognizable by the model; the data will be embedded in the model in the form of tags and / or keywords, so that the model can quickly retrieve relevant information based on consumer questions; in this embodiment, the data will be embedded in the model in the form of tags and keywords at the same time.

[0036] S22. Intent recognition and semantic matching: The model first identifies the consumer’s question intent through natural language processing (NLP) technology. For example, a customer may ask, “Is this phone suitable for taking photos?” The model will identify the keyword “photography” and associate it with the Mate 50 Pro’s photography function. S23. Ability to generate conversations: Based on a pre-trained and fine-tuned model, sales scripts and product feature data are embedded into the language generation model. When answering, the model extracts the best responses from similar questions or scripts and generates coherent and natural language based on the context.

[0037] S24. Multi-round dialogue management: The model also uses a multi-round dialogue management module to ensure continuous and accurate responses when consumers have further questions or change their needs. For example, if a consumer first asks "Recommend a mobile phone suitable for taking photos" and then asks "What about the battery life?", the system can retain the context and provide more complete recommendations.

[0038] Among them, step S24 is an optional step.

[0039] Preferably, the intelligent shopping guide interaction method includes: S31, customer demand collection: the system first collects customer input through the terminal device, including voice and / or text; the voice input will be converted into text through voice recognition technology (such as ASR, Automatic Speech Recognition), and the text processing module will perform basic cleaning and standardization processing based on the input content to ensure that the large language model receives clear and accurate demand information; this embodiment collects voice and text information at the same time (collect if available); S32, Intent recognition and keyword extraction: Through intent recognition technology, the system analyzes customer needs. For example, if a customer enters "recommend a mobile phone suitable for taking photos", the system will recognize "photography" and "mobile phone" as keywords, and classify the intention as "product recommendation". The accuracy of intent recognition and keyword extraction directly affects the matching degree of subsequent product recommendations. S33, calling the large language model to generate recommendations or answers: After the intent and keywords are determined, the system calls the trained large language model to generate recommended content; S34. Real-time feedback and display: The generated recommendations and answers are transmitted to the terminal device in real time. The system displays several items such as product details, promotional activities or usage suggestions in a graphic and textual manner. This module supports multimedia display, including product pictures, videos, user reviews, etc., to help customers understand the product more comprehensively.

[0040] Preferably, the large language model uses the following methods to ensure natural and coherent recommendations and answers: S331, context processing: If the customer's input involves multiple rounds, the model will call on the context information to maintain coherent communication. Context processing ensures that the customer's needs are remembered in each round of dialogue, thereby generating logical and coherent recommendations; S332, Knowledge base retrieval and matching: The model retrieves content matching the keywords from the knowledge base and integrates several items of information such as the core features, advantages and disadvantages, and applicable scenarios of the product into the recommendation language. For example, the keyword "photographing" will trigger the system to give priority to recommending models with excellent camera functions. S333. Generate natural language dialogue: By processing the input needs and knowledge base content, introducing preset sales scripts, and generating language that is closer to the real shopping guide style, the response appears natural and smooth; for example, if a customer asks "a mobile phone suitable for taking pictures", the system will answer "Huawei Mate 50 Pro is a good choice, and its ultra-sensitive imaging system performs particularly well in night scene shooting."

[0041] Preferably, the personalized recommendation method includes: S41. Demand analysis and customer classification: The system first receives the customer's input and identifies the customer's intention and specific needs (such as "a mobile phone suitable for taking photos" and "price within 3,000 yuan") through the semantic analysis module. While analyzing the needs, the system divides the customers into two categories: new customers and old customers. The new customers include customers who use the system for the first time or customers who do not have sufficient interaction data. The system mainly makes recommendations based on current input requirements; The old customers include customers with existing interaction data. The system will give priority to historical preferences and purchase records, and provide more personalized recommendations based on current input needs.

[0042] S42, screening and matching strategy of large language model: Based on the demand analysis results, the system calls the large language model to screen and match existing products; Regarding the problem of matching real-time semantic understanding with consumer demand: different customers may express their needs in a variety of ways. How to accurately understand the natural language input by customers through a large language model and respond and recommend in real time is a core technical problem. In addition, customer needs are often vague and may involve multiple product attributes (such as "low price, good performance"). The system needs to accurately interpret these fuzzy needs and quickly select the most matching recommendations from a variety of possible products. At the same time, the system needs to complete this complex processing process within milliseconds to ensure the real-time experience of consumers. This embodiment uses efficient semantic parsing and fast matching algorithms to accurately understand customer needs and respond in real time.

[0043] Semantic analysis: Combines hierarchical intent recognition and fuzzy matching to analyze the multi-layered meanings of customer needs. For example, for "low price and good performance", the model automatically extracts the two key attributes of "price" and "performance"; Fast matching algorithm: Build a multi-condition screening algorithm to quickly screen out the most matching products by weighting features (such as assigning different weights to requirements such as photography and battery life); Cache mechanism: For common needs, pre-store matching results to reduce the amount of calculation for each screening and improve response speed.

[0044] Preferably, the screening and matching process includes any number of the following steps S421 to S423, and there is no strict order: S421, label and keyword matching: The product data in the knowledge base has been labeled according to labels such as function, brand, price range, and purpose. The large language model will identify the keywords entered by customers (such as "photography", "game", "5G", etc.) and match them with the knowledge base labels to screen out products that meet the needs; S422, multi-condition screening: For complex requirements (such as "Huawei mobile phones suitable for taking photos and priced below 3,000 yuan"), the model constructs a multi-condition screening tree to gradually filter out products that do not meet the conditions from the knowledge base and retain the product set that meets all the conditions; S423, Similarity matching and weighted recommendation: For fuzzy requirements (such as "a mobile phone with good camera performance"), the large language model performs similarity matching based on product features and sorts the results by weighted scores of camera performance to ensure the best recommendation effect; The personalized recommendation for old customers includes any number of the following steps S431 to S433, and there is no strict order: S431. Historical purchase records and access data: The system will extract the purchase records and historical browsing records of old customers, and conduct secondary screening based on the customer's preferences for different product categories (such as preference for a certain brand or configuration). For example, if a customer has chosen the Huawei brand many times in the past, the system will give priority to recommending Huawei's latest products. S432. Fusion matching of demand and historical preferences: The large language model not only considers the customer's current demand when screening, but also weights the customer's historical preferences. For example, when an old customer inputs the demand for "5G mobile phone", the system will give priority to recommending Huawei products that meet 5G conditions and have high-scoring camera functions; S433. Intelligent recommendation optimization and intention expansion: For old customers who have not clearly expressed their specific needs, the system automatically generates further guiding questions through a large language model. For example, when an old customer inputs "Please recommend a mobile phone for me", the system will add questions such as "Do you have a brand preference? What features do you want to pay attention to?" to generate more personalized recommendations.

[0045] Preferably, the feedback data types of the feedback optimization include any several items of interactive behavior data, recommendation result data, purchase conversion data, customer feedback and evaluation, customer portrait and preference data.

[0046] The interactive behavior data includes all behaviors of customers recorded by the system during the interaction with the shopping guide system, including clicks, browsing, collections, click-through rates of recommended products, etc. These data reflect the customer's interest in different recommended content.

[0047] The recommendation result data includes the product list data tracked by the system for each recommendation, including the recommended product types, display order, recommendation reasons, etc. By analyzing the effects of the recommendation results, the system can improve the recommendation strategy and improve the recommendation accuracy of the model.

[0048] The purchase conversion data includes the customer's purchase decision data, that is, whether the customer purchases the recommended product and which recommended products the customer purchases. The purchase conversion data can directly reflect the effectiveness of the recommendation.

[0049] The customer feedback and evaluation includes satisfaction evaluation or feedback data provided by customers after the interaction, such as "whether the recommendation is accurate", "whether the recommended product meets the needs", etc. This feedback can be used to adjust the recommendation strategy and improve the customer experience.

[0050] The customer portrait and preference data include personalized information such as historical preferences, purchase records, usage scenarios, etc. of old customers. By analyzing the customer's historical behavior, the system can adjust the recommendation weight to meet the customer's needs more accurately.

[0051] The sources of feedback data include several items such as smart shopping guide interaction records, purchase record database, evaluation and rating system, and old customer archive.

[0052] Smart shopping guide interaction records: including all clicks, browsing, collections and other behavior records of customers in the smart shopping guide terminal. The system will automatically synchronize these data to the feedback optimization module after each interaction is completed.

[0053] Purchase record database: The system is connected to the sales database to compare the customer's final purchase data with the recommendation data to evaluate the success rate of the recommendation. This data can be used to train the model to improve the conversion rate.

[0054] Evaluation and rating system: At the end of the interaction, customers can rate the recommendation results and service satisfaction, or make suggestions for improvement. The rating and evaluation data will be automatically imported into the feedback optimization module to help the system adjust the recommendation strategy in a timely manner.

[0055] Old customer archive: contains information such as customers’ historical purchase preferences, common needs, and usage scenarios. The system continuously updates old customers’ behavioral preferences by connecting to the archive for personalized optimization of future recommendations.

[0056] Embodiment 2: The difference from Example 1 is that in this embodiment, when collecting or interacting with customer needs, only voice is collected, and data is embedded into the model in the form of keywords during knowledge base preprocessing.

[0057] Embodiment 3: The difference from Example 1 is that the personalized recommendation method described in this embodiment further includes: S43, personalized recommendation for old customers: For old customers identified by the system, the personalized recommendation module will call the customer's historical preference data to make recommendations that are more in line with their habits; S44. Display of recommendation results and comparison of multiple options: After the system selects the products that best meet the needs and preferences of customers, the recommendation results will be displayed on the terminal device, supporting multiple option comparison display to help customers evaluate products from multiple angles; the display content includes detailed parameters, special functions, discount information, etc. of the product, so that customers can clearly understand the differences between products; Steps S43 and S44 are independently optional or jointly optional steps.

[0058] Embodiment 4: A retail store unmanned shopping guide system based on a large language model, a system constructed according to any one of the retail store unmanned shopping guide methods based on a large language model (including the preferred method) in the aforementioned embodiments, such as Figure 2 As shown, it includes a data collection and organization module, a large language training module, an intelligent shopping guide interaction module, a personalized recommendation module and a feedback optimization module which are sequentially and closed-loop connected. The feedback optimization module connects the data collection and organization module to form a closed loop.

[0059] Regarding data security and privacy protection: Consumers' interaction data and store sales data involve privacy and business secrets. How to ensure that these data are not leaked during transmission, storage and processing is a key issue in system design. The data needs to be strictly encrypted and a sound access control mechanism needs to be provided to prevent unauthorized access. At the same time, the system must also comply with laws and regulations related to data privacy, such as GDPR, to ensure the security of commercial data while protecting consumer privacy. This embodiment uses encryption technology and permission management to comprehensively protect data security and user privacy.

[0060] Data encryption: During data transmission and storage, encryption protocols (such as TLS and AES encryption) are used to protect data and prevent sensitive information leakage; Access control: Set up a multi-level permission management mechanism to ensure that only authorized users can access specific data; and anonymize customer data to protect their privacy; Compliance management: The system strictly complies with data privacy regulations (such as GDPR), obtains user consent before storing customer interaction data, and conducts regular data security audits.

[0061] Regarding the scalability and maintainability of the system: With the increase in product types and customer interactions, the system needs to have good scalability to support more product data and user requests. How to expand while maintaining system performance is a major technical difficulty. The system needs to be designed with a modular structure to facilitate future functional expansion and maintenance. At the same time, with the continuous optimization of large language models, the system's training process and model update mechanism must also be efficient and automated. This embodiment constructs a modular system architecture and optimizes the model's training and update mechanism.

[0062] Modular design: The system is separated according to different functional modules (such as product data management, recommendation system, feedback optimization, etc.) to achieve flexible expansion and update; Automated training and updating: Feedback data is integrated into model training in real time through an automated pipeline, and model parameters are continuously optimized in an incremental manner.

[0063] The basic principle of the present invention is: based on the pre-built structured knowledge base and customer data collection and collation, the present invention conducts targeted training on the large language model to provide information and technical support for intelligent shopping guide interaction; through personalized recommendations, the best product recommendation results are provided to customers; through feedback optimization, the prediction ability of the model is continuously optimized, so that the system can continuously self-learn and optimize, and the intelligent shopping guide ability is continuously enhanced. Thus, a method and system for unmanned shopping guide in retail stores based on a large language model is provided, which aims to provide customers with automated product introductions and sales suggestions through intelligent terminal devices to promote the sales of retail store goods.

[0064] This method uses pre-entered sales scripts and product information, combined with real-time customer demand analysis, to generate suitable product recommendations and interact with customers through intelligent voice or text. The system does not require human intervention and can effectively solve problems such as insufficient staff and uneven service quality in traditional sales scenarios, improving retail store operating efficiency and customer experience, including

[0065] 1. Realization of unmanned shopping guide The present invention realizes unmanned shopping guide by combining a large language model with the specific needs of the retail industry. This model greatly reduces the dependence on manual services and can provide highly personalized recommendations and consultations during customer interaction.

[0066] 2. Real-time intelligent interaction The system can achieve real-time intelligent interaction through smart terminal devices, quickly respond to consumers' questions and provide high-quality product recommendations. This immediacy improves consumers' shopping experience and reduces shopping waiting time.

[0067] 3. Accurate personalized recommendations Combined with the powerful computing power of the large language model, the system can generate accurate product recommendations based on the customer's demand characteristics. Whether in terms of price, function or usage scenario, the system can provide the most suitable product information for customers.

[0068] 4. Large-scale data processing and model training By introducing a large amount of product information and consumer interaction data, the system can continuously optimize shopping guide services and improve the response accuracy and recommendation efficiency of the large language model.

[0069] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical principles of the present invention, several improvements and modifications can be made, such as the way of constructing the knowledge base, the selection and training method of the large language model, etc. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A retail store unmanned shopping guide method based on a large language model, characterized in that: Based on the constructed structured knowledge base, the large language model is trained in a targeted manner to provide information and technical support for intelligent shopping guide interaction. The method includes: S1. Building a structured knowledge base, including data collection and organization, wherein the data collection includes product feature data, function description data, consumer feedback data, and several items of frequently asked questions, and data organization includes storing the collected data in a structured form in the knowledge base; S2. Targeted training of large language models. By introducing sales tactics, product pros and cons analysis, and consumer needs in different scenarios, the system generates intelligent conversation capabilities based on natural language processing and responds to consumers' questions in real time. S3, smart shopping guide interaction, customers interact with the system through store smart terminal devices. The system uses a large language model to generate natural and coherent product recommendations and consultation answers based on the customer's voice and / or text input; S4, Personalized recommendation: The system uses a large language model to filter and match existing products based on the needs or preferences entered by customers, and provides customers with the best product recommendations; S5. Feedback optimization: the system records every customer interaction and recommendation result, and uses feedback data to optimize the model's predictive ability.

2. The method for unmanned shopping guide in retail stores based on a large language model according to claim 1, characterized in that: The product characteristic data includes several items including the product's brand, model, specifications, price, color, applicable scenarios, and main functions; the functional description data includes several items including instructions for use, technical features, and special functions; the consumer feedback data includes several items including the product's advantages, disadvantages, and improvement suggestions; the frequently asked questions include several items including product comparison, usage scenario suggestions, and technical parameter explanations.

3. The method for unmanned shopping guide in retail stores based on a large language model according to claim 1, characterized in that: The sales talk includes the communication language and skills commonly used by retail staff when receiving customers, which are embedded in the model in the form of natural language; the sales talk also includes several items of welcome talk, demand confirmation talk, product recommendation talk, and purchase promotion talk; The product advantages and disadvantages analysis includes a summary of each product's core performance, features, applicable population, and shortcomings; The consumer demands include that different consumer groups may focus on different scenario demands when purchasing.

4. The method for unmanned shopping guide in retail stores based on a large language model according to claim 1, characterized in that: The method for generating the intelligent dialogue capability includes: S21. Knowledge base preprocessing: First, the data on sales pitches, product advantages and disadvantages, and consumer needs are processed into structured text and converted into a format that the model can recognize. The data is embedded into the model in the form of tags and / or keywords, so that the model can quickly retrieve relevant information based on the consumer's questions. S22, Intent recognition and semantic matching: The model first identifies the consumer's question intent through natural language processing technology; S23, Conversation generation capability: Based on the pre-trained and targeted fine-tuned models, sales scripts and product feature data are embedded into the language generation model; S24, Multi-round dialogue management: Through the multi-round dialogue management module, ensure continuous and accurate responses when consumers have further questions or change their needs; Among them, step S24 is an optional step.

5. The method for unmanned shopping guide in retail stores based on a large language model according to claim 1, characterized in that: The method for intelligent shopping guide interaction includes: S31. Customer demand collection: The system first collects customer input through the terminal device, including voice and / or text; S32, Intent recognition and keyword extraction: Through intent recognition technology, the system analyzes customer needs; S33, calling the large language model to generate recommendations or answers: After the intent and keywords are determined, the system calls the trained large language model to generate recommended content; S34. Real-time feedback and display: The generated recommendations and answers are transmitted to the terminal device in real time. The system displays detailed product information, promotional activities and several usage suggestions in a graphic and textual manner.

6. The method for unmanned shopping guide in retail stores based on a large language model according to claim 5, characterized in that: The large language model uses the following methods to ensure natural and coherent recommendations and answers: S331, context processing: If the customer's input involves multiple rounds, the model will call on the context information to maintain coherent communication; S332, Knowledge base retrieval and matching: The model retrieves content matching the keywords from the knowledge base and integrates several items of information on the core features, advantages and disadvantages, and applicable scenarios of the product into the recommendation language; S333. Generate natural language dialogue: By processing the input requirements and knowledge base content, introduce preset sales scripts, and generate language that is closer to the style of real shopping guides, making the responses appear natural and smooth.

7. The method for unmanned shopping guide in retail stores based on a large language model according to claim 1, characterized in that: The personalized recommendation method includes: S41. Demand analysis and customer classification: The system first receives the customer's input and identifies the customer's intention and specific needs through the semantic analysis module. While analyzing the needs, the system divides the customers into two categories: new customers and old customers; S42, screening and matching strategy of large language model: Based on the demand analysis results, the system calls the large language model to screen and match existing products; S43, personalized recommendation for old customers: For old customers identified by the system, the personalized recommendation module will call the customer's historical preference data to make recommendations that are more in line with their habits; S44, display of recommendation results and comparison of multiple options: After the system screens out the products that best meet the customer's needs and preferences, the recommendation results are displayed on the terminal device, supporting multiple option comparison display to help customers evaluate products from multiple angles; Among them, steps S43 and S44 are optional steps.

8. The method for unmanned shopping guide in retail stores based on a large language model according to claim 7, characterized in that: The screening and matching process includes any number of the following steps S421 to S423, in no particular order: S421, label and keyword matching: the product data in the knowledge base has been labeled according to function, brand, price range, and purpose labels. The large language model identifies the keywords entered by the customer and matches them with the knowledge base labels to screen out products that meet the needs; S422, Multi-condition screening: For complex requirements, the model builds a multi-condition screening tree to gradually filter out products that do not meet the conditions from the knowledge base and retain the product set that meets all conditions; S423, Similarity matching and weighted recommendation: For fuzzy requirements, the large language model performs similarity matching based on product characteristics and sorts the results by weighted performance scores to ensure the best recommendation effect; The personalized recommendation for old customers includes any number of the following steps S431 to S433, in no particular order: S431, historical purchase records and access data: the system extracts the purchase records and historical browsing records of old customers, and performs secondary screening based on the customer's preferences in product selection in different categories; S432. Fusion matching of demand and historical preferences: When screening, the large language model not only considers the customer's current demand, but also adds weight to the customer's historical preferences; S433, Intelligent recommendation optimization and intention expansion: For old customers who have not clearly expressed their specific needs, the system automatically generates further guiding questions through a large language model.

9. The method for unmanned shopping guide in retail stores based on a large language model according to claim 1, characterized in that: The feedback data type of the feedback optimization includes any several items of interactive behavior data, recommendation result data, purchase conversion data, customer feedback and evaluation, customer portrait and preference data.

10. A retail store unmanned shopping guide system based on a large language model, characterized in that: The system constituted by the unmanned shopping guide method for retail stores based on a large language model according to any one of claims 1 to 9 comprises a data collection and organization module, a large language training module, an intelligent shopping guide interaction module, a personalized recommendation module and a feedback optimization module connected in sequence and in a closed loop, wherein the feedback optimization module connects the data collection and organization module to form a closed loop.

Images