A method and system for processing shopping guide information

Abstract

The application provides a shopping guide information processing method running in a commodity sales system, which comprises the following steps: an intelligent shopping guide device in the commodity sales system acquires body data confirmed by a customer and generates body analysis data of the customer; a first recommendation intelligent agent generates first commodity recommendation information of the customer according to the body analysis data and commodity information accessed in the commodity sales system; a second effect generation intelligent agent generates first use effect generation data of the recommended commodity according to the body analysis data and the first commodity recommendation information; and first shopping guide recommendation strategy data is generated on the shopping guide side according to user label data of the customer, the first commodity recommendation information, the first use effect generation data and interactive feedback information of the customer in the commodity sales system. The embodiment of the application improves the precision of shopping guide service and the customer shopping guide experience during commodity sales.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for processing shopping guide information. Background Technology

[0002] In today's internet-driven information age, in the sales of goods, especially clothing, sales guidance is the core link connecting customers and products, and the fitting experience directly determines customer satisfaction and the efficiency of sales guidance. Currently, the traditional sales guidance and fitting model still has many limitations. In offline scenarios, customers need to repeatedly try on different styles and sizes of clothing, which not only consumes a lot of time, but also leads to problems such as limited fitting room resources and high costs for cleaning and maintaining clothing after trying it on. In online scenarios, customers cannot intuitively perceive the effect of clothing on their bodies, resulting in a high return rate for apparel. The return rate for apparel on ordinary e-commerce platforms is generally between 25% and 60%, and in live-streaming scenarios, it can even reach 80% to 90%. With the rapid penetration of artificial intelligence (AI) technology in the e-commerce field, technologies such as AI-powered intelligent sales guidance and virtual fitting are gradually being applied to product sales systems, attempting to solve the pain points of traditional fitting guidance. Among them, AI-enabled virtual fitting has become the "last mile" in improving the online retail experience of apparel.

[0003] In the prior art, patent application number 202411347596.2, entitled "Virtual Fitting Method and Fitting Mirror," discloses a virtual fitting method and fitting mirror. The virtual fitting method specifically includes: S10: acquiring image information in front of the fitting mirror; S20: determining whether the image information includes person information; S30: if the image information includes person information, acquiring the person's needs information, including mirror-viewing needs and fitting-up needs; S40: if the person's needs information is fitting-up needs, taking a picture of the person to acquire their appearance information; S50: acquiring the person's fitting-up requirements for clothing, and based on the person's appearance information and fitting-up requirements, retrieving clothing from an e-commerce platform to create clothing combinations for the person; S60: displaying the matching effect on the fitting mirror. The existing technology enables the basic functions of virtual try-on and clothing matching through a fitting mirror, which reduces the number of times customers actually try on clothes to some extent. However, this solution only focuses on the presentation of the fitting effect and does not form a complete closed loop of sales strategy, thus failing to provide targeted service guidance for sales staff.

[0004] In conclusion, how to fully leverage the advantages of AI technology to improve the accuracy of shopping guide services and enhance the customer shopping experience has become a crucial issue to be addressed in the current merchandise sales sector. Summary of the Invention

[0005] This invention provides a method and system for processing shopping guide information to improve the accuracy of shopping guide services and the customer shopping experience during product sales.

[0006] In a first aspect, embodiments of the present invention provide a method for processing shopping guide information, which operates in a product sales system. The method includes: an intelligent shopping guide device in the product sales system acquiring body shape data confirmed by a customer and generating body shape analysis data for the customer; a first recommendation agent generating first product recommendation information for the customer based on the body shape analysis data and product information accessed in the product sales system; a second effect generation agent generating first usage effect generation data for the recommended product based on the body shape analysis data and the first product recommendation information; and generating first shopping guide recommendation strategy data on the shopping guide side based on the customer's user tag data, the first product recommendation information, the first usage effect generation data, and the customer's interactive feedback information in the product sales system.

[0007] Optionally, in the method for processing shopping guide information in this embodiment of the invention, the body data includes at least one of the following: customer's body shape data, body part proportion and shape data, and facial feature data.

[0008] Optionally, in the method for processing shopping guide information in this embodiment of the invention, the first recommendation agent generating the customer's first product recommendation information based on the body shape analysis data and the product information accessed in the product sales system includes: the first recommendation agent generating a first product recommendation list from the product information accessed in the product sales system based on the customer's body shape tags and / or style tags in the body shape analysis data, through a person-product matching rule; and / or the first recommendation agent generating a first product recommendation list based on a styling formula library, combined with the product information accessed in the product sales system, and through the aesthetic agent scoring results of the body shape analysis data after a virtual model tries on the clothes; and / or the first recommendation agent generating a first product recommendation list by combining the product information accessed in the product sales system, and after the body shape analysis data is used to try on clothes on a virtual model, retrieving the collaborative scoring of the designer agent and / or stylist agent and / or fashion agent.

[0009] Optionally, in the method for processing shopping guide information according to embodiments of the present invention, the second effect generation agent generates the first usage effect generation data of the recommended product based on the body analysis data and the first product recommendation information, including: matching a wearing model based on the body analysis data and loading and generating first image data of the wearing model wearing the recommended product; face detection and alignment: using a face detection and key point alignment model to detect the facial region in the first image data, and achieving alignment through face masking type; feature replacement: using a face swapping model to perform facial feature transfer, preserving the facial structure of the customer target image in the body analysis data; detail optimization: achieving a natural transition of the facial region in the first image data through edge expansion and blurring processing, and combining a face restoration model to restore the skin texture of the facial region in the first image data; super-resolution enhancement: improving the output resolution of the facial region in the first image data based on a diffusion transformer model and a variational autoencoder model, ensuring consistent skin color in the facial region in the first image data, and outputting the first usage effect generation data.

[0010] Optionally, in the method for processing shopping guide information according to embodiments of the present invention, the second effect generation agent generating the first usage effect generation data of the recommended product based on the body analysis data and the first product recommendation information includes: triggering a try-on rendering request on the intelligent shopping guide device side; the second effect generation agent issuing a try-on instruction to the rendering engine on the shopping guide side, directly calling the locally stored body analysis data, completing the first try-on rendering locally to obtain the first basic image data; extracting the first desensitized body features from the body analysis data on the shopping guide side; uploading the first desensitized body features, the first product recommendation information, and the rendering requirement information to the cloud-based detail rendering engine for the second try-on rendering to obtain the second detail enhancement image data; and fusing the first basic image data and the second detail enhancement image data to generate the first usage effect generation data of the recommended product.

[0011] Optionally, in the method for processing shopping guide information according to the embodiments of the present invention, the step of generating first shopping guide recommendation strategy data on the shopping guide side based on the customer's user tag data, first product recommendation information, first usage effect generation data, and the customer's interaction feedback information in the product sales system includes: the first shopping guide state machine in each shopping guide stage state, triggering the transition conditions of the shopping guide state machine in real time according to the customer's user tag data, first product recommendation information, first usage effect generation data, and the customer's interaction feedback information in each shopping guide stage in the product sales system, and generating corresponding first shopping guide recommendation strategy data according to the prompt word template.

[0012] Optionally, in the method for processing shopping guide information according to embodiments of the present invention, the step of generating first shopping guide recommendation strategy data on the shopping guide side based on the customer's user tag data, first product recommendation information, first usage effect generation data, and the customer's interaction feedback information in the product sales system includes: acquiring shopping guide stage status information, the customer's user tag data, first product recommendation information, first usage effect generation data, and the customer's interaction feedback information in the product sales system in real time on the shopping guide side, and outputting the customer's shopping guide recommendation strategy data in the corresponding shopping guide stage according to the third shopping guide intelligent agent.

[0013] Optionally, in the method for processing shopping guide information in this embodiment of the invention, the product information accessed in the product sales system includes: product information entered in the product sales system, and product information obtained from at least one external enterprise resource planning data interface accessed by a brand owner.

[0014] Optionally, the method for processing shopping guide information in this embodiment of the invention further includes: a fourth product analysis intelligent agent analyzing the sales data of the product sales system and the customer's interactive feedback information in the product sales system, and feeding back the analysis results data to the product analysis big model on the enterprise resource planning side accessed by the brand owner.

[0015] Secondly, embodiments of the present invention provide a product guidance system, comprising: a body shape data acquisition and analysis module, configured in an intelligent shopping guide device within a product sales system, and configured to acquire body shape data confirmed by a customer and generate body shape analysis data of the customer; a first product recommendation module, configured to use a first recommendation agent to generate first product recommendation information for the customer based on the body shape analysis data and product information accessed in the product sales system; a first effect generation module, configured to use a second effect generation agent to generate first usage effect generation data of the recommended product based on the body shape analysis data and the first product recommendation information; and a shopping guide recommendation strategy data generation module, configured to generate first shopping guide recommendation strategy data for the shopping guide based on the customer's user tag data, the first product recommendation information, the first usage effect generation data, and the customer's interactive feedback information in the product sales system.

[0016] The technical solution of this invention acquires the body shape data confirmed by the customer through an intelligent shopping guide device in a product sales system, generating the customer's body shape analysis data. A recommendation agent generates product recommendation information for the customer based on the body shape analysis data and product information accessed in the product sales system. An effect generation agent generates usage effect generation data for the recommended products based on the body shape analysis data and product recommendation information. Based on the customer's user tag data, product recommendation information, usage effect generation data, and the customer's interactive feedback information in the product sales system, shopping guide recommendation strategy data is generated on the shopping guide side. This enables professional shopping guide recommendation strategy data to be generated based on the user's actual shopping scenario, participation status, and user's tag data during product sales, allowing shoppers to receive precise shopping guide services and purchase products that better meet their needs. Furthermore, the effect generation agent generates effect generation data for the recommended products during customer use, greatly improving the customer's shopping guide experience and achieving greater accuracy in the shopping guide service. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for processing shopping guide information according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the structure of the commodity sales system provided in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the interface interaction for generating shopping guide recommendation strategy data on the shopping guide side of the product sales system provided in Embodiment 3 of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0021] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0022] Furthermore, the timing of the steps in the following method embodiments is merely an example and not a strict limitation.

[0023] In embodiments of the present invention, the body shape data confirmed by the customer is obtained through an intelligent shopping guide device in the product sales system, generating the customer's body shape analysis data; a recommendation agent is provided to generate product recommendation information for the customer based on the body shape analysis data and product information accessed in the product sales system; and an effect generation agent generates usage effect data of the recommended products based on the body shape analysis data and product recommendation information; finally, shopping guide recommendation strategy data is generated on the shopping guide side. This enables the generation of professional shopping guide recommendation strategy data based on the actual shopping scenario, participation status, and user's own tag data during product sales, thereby providing shoppers with precise shopping guide services; moreover, the effect generation agent generates effect data of the recommended products when used by the customer during the shopping process, greatly improving the customer's shopping guide experience and also achieving greater accuracy in the shopping guide service.

[0024] Example 1 Figure 1This is a flowchart illustrating a method for processing shopping guide information according to Embodiment 1 of the present invention. The method operates within a merchandise sales system adapted to various scenarios, including offline physical apparel stores, online e-commerce apparel platforms, and live-streaming apparel shopping guides. It incorporates hardware and software modules such as intelligent shopping guide devices, a multi-AI intelligent agent cluster, a cloud rendering engine, and an Enterprise Resource Planning (ERP) data interface. This system provides end-to-end functionality including body data collection, intelligent product recommendation, virtual try-on effect generation, adaptive generation of shopping guide strategies on the shopping guide side, and linked analysis of merchandise data. It addresses the technical pain points of traditional apparel shopping guides, such as cumbersome try-on processes, poor experience, inaccurate service, and high return rates, achieving intelligent, precise, and closed-loop shopping guide services, thus facilitating improved service from sales staff. In this embodiment, the method for processing shopping guide information specifically includes the following steps: Step 100: The intelligent shopping guide device in the product sales system acquires the body shape data confirmed by the customer and generates the customer's body shape analysis data. In the product sales system of this embodiment, the intelligent shopping guide device is a customized terminal adapted to multiple scenarios. In offline scenarios, an intelligent shopping guide mirror / all-in-one shopping guide machine equipped with a light camera, a high-definition RGB acquisition module, and a touch interactive screen is used. The hardware parameters are configured as follows: light camera acquisition frame rate of 30fps, acquisition distance range of 0.5m-2m, high-definition RGB module with 12 million pixels, and touch interactive screen size of 43 inches, supporting 10-point touch. Preferably, in online scenarios, the intelligent shopping guide device can use the camera acquisition module built into a mobile phone / personal computer (PC). The mobile phone is compatible with Android 8.0 and above, iOS 12.0 and above systems, and uses the intelligent shopping guide APP or web application to obtain the body data confirmed by the customer. The device integrates functions such as data authorization verification, body information collection, data anonymization, and analysis modeling. It fully complies with the privacy authorization rules of the Personal Information Protection Law. The data collection and analysis process is only initiated after the customer actively clicks to confirm authorization and signs the body data usage agreement, thus preventing unauthorized information acquisition and protecting customer privacy.

[0025] Shape data acquisition adopts a non-contact passive acquisition mode. The device has a built-in data calibration module. First, environmental calibration is performed to eliminate interference from light and distance. The calibration formula is as follows: ,in For actual body size data, The original data collected by the device, This is the light calibration factor (value range 0.95-1.05, 0.95 for strong light, 1.05 for weak light, and 1.00 for natural light). The distance calibration coefficient is 1.02 for 0.5m, 1.00 for 1m, and 0.98 for 2m. This formula can stably control the acquisition accuracy within ±0.3cm, which is far higher than the accuracy standard of conventional acquisition equipment.

[0026] Preferably, in this embodiment, the body data collected in this step includes at least one of the following: customer's body shape data, body part proportion and shape data, and facial feature data. The data collection dimensions can be flexibly selected according to the needs of the sales guide scenario. The data details and collection methods for each dimension are shown in the table below: Shape data type Specific data items Data types Acquisition / parse method Body shape and size data Height, weight, shoulder width, chest circumference, waist circumference, hip circumference, arm length, leg length, clothing length (fitting size), pant length (fitting size) Quantized numerical type Non-contact data acquisition via optical cameras, real-time calculation by edge computing units. Body part proportions and morphological data Shoulder shape, body shape, leg shape, body length ratio, waist-to-hip ratio, bust-to-waist ratio Qualitative labeling AI shape recognition models analyze collected data and generate labels based on threshold classification. Facial feature data Face shape, distance between facial features, facial contour, skin tone levels, and coordinates of facial key points. Characteristic parameter type High-definition face capture module acquires data, and facial landmark detection model extracts features. The body proportion data is generated through quantitative formulas, with the core proportion calculation formula being: , , Waist-to-hip ratio The ratio of bust to waist circumference; the formula for body length proportions: AI-based body shape recognition models categorize body types based on proportional thresholds, for example: Determined to be apple-shaped body type. Determined to be an hourglass body type. The body type was identified as pear-shaped, achieving standardized and quantitative classification of body type labels and avoiding subjective errors from manual judgment. Facial feature data was extracted using a 68-point facial landmark detection model, outputting a landmark coordinate matrix to provide accurate parameter support for subsequent facial feature transfer.

[0027] After the customer completes the authorization submission of their body data, the edge computing unit built into the smart shopping guide device (computing power configuration: 8-core ARM architecture processor, 8GB RAM) initiates the data preprocessing process. First, it performs noise reduction, completion, and normalization on the collected raw body data: a Gaussian filtering algorithm is used to remove collected noise, and the filtering formula is... ( (Value 1.2); for missing local shape data, a neighborhood interpolation algorithm is used to complete the interpolation, and the interpolation formula is... ,in The neighboring points are weighted, and n is taken as 8 neighboring points; then the processed standardized shape data is transmitted to the system's built-in shape analysis model.

[0028] This body shape analysis model is trained based on a lightweight Convolutional Neural Network (CNN). The network structure consists of 4 convolutional layers and 2 fully connected layers, with the number of parameters controlled to within 1.5 million. It is deployed on a lightweight server, either locally or in the cloud, within the smart shopping guide device. The parsing time for a single data entry is ≤0.8 seconds, enabling rapid data parsing and tag generation. The final output is personalized body shape analysis data for each customer, formatted in JSON (JavaScript Object Notation). This data includes four main modules: a quantitative body shape parameter table, multi-dimensional body shape tags, body shape adaptation suggestions, and core proportion parameters. This facilitates subsequent intelligent agent parsing and provides a clear view of the customer's body shape characteristics, offering crucial data support for product recommendations and virtual try-on. Furthermore, the body shape analysis data generated in this step is stored locally using AES-256 encryption. During anonymized transmission, MD5 hashing is used to remove personal identification information, retaining only body shape feature parameters, thus balancing data efficiency and privacy security.

[0029] Step 101: The first recommendation agent generates the customer's first product recommendation information based on the body shape analysis data and the product information accessed from the product sales system. In the product sales system of this embodiment, the first recommendation agent is the core AI agent module, built on a fusion architecture of deep learning and rule engine, and has functions such as accurate matching of people and goods, intelligent filtering of clothing, and collaborative scoring by multiple agents. In this embodiment, this step first completes the full-domain access of product information, and then generates accurate first product recommendation information through various matching logics, realizing the diversification and accuracy of recommendation logic. The first recommendation agent adopts a pre-training + fine-tuning training mode, relying on massive clothing and body shape data to complete the training. The specific training technical solution is as follows: First recommended intelligent agent training technology solution I. Construction of Training Dataset A multi-dimensional labeled dataset was constructed, with a total of 1.2 million data sets, covering four main categories: body feature data, product feature data, matching result labels, and outfit rating data. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio. 1. Body Feature Subset: Contains 500,000 sets of customer body data, covering various body parameters, body type tags, and proportion parameters, and simultaneously annotates body shape-suitable styles and style preferences; 2. Product Feature Subset: Contains 400,000 sets of apparel product data, covering different categories, fits, sizes, styles, and fabrics, and annotating the product's suitable body shape and wearing scenarios; 3. Matching tag subset: Based on manual annotation by professional stylists, the body shape-product matching relationship is assigned a score of 0-1. A score ≥0.7 is a positive sample, and <0.7 is a negative sample; 4. Outfit Rating Subset: Contains 300,000 sets of virtual try-on effect rating data, labeled with aesthetic ratings and professional matching ratings. The dataset has undergone anonymization to remove personal privacy information, and normalization has been used to map all feature data to the [0,1] interval for easy model training.

[0030] II. Model Training Architecture and Process We employ a fusion architecture of a lightweight BERT (Bidirectional Encoder Representations from Transformers) pre-trained model and a Multilayer Perceptron (MLP). The base pre-trained model loads open-source pre-trained weights from the apparel domain, freezes 80% of the bottom-level parameters, and only fine-tunes the top-level feature extraction layer and fully connected layer. The training framework is built on PyTorch, and the hardware environment is an NVIDIA RTX 4090 graphics card. The training parameters are set as follows: batch size = 64, learning rate = 2e-5, the optimizer is AdamW, the weight decay coefficient = 1e-4, the number of training epochs = 50, and the early stopping strategy is set to stop training if the validation set loss does not decrease for 10 consecutive epochs to prevent overfitting.

[0031] The training process consists of three steps: The first step is basic feature training. Input shape feature and product feature data, and train the model to extract core related features. The second step is matching accuracy training. Based on the cosine similarity matching branch, manually labeled matching scores are used as loss labels, and the mean squared error loss function is employed. Optimize model matching accuracy; The third step is multi-logic fusion training, which integrates rule matching, aesthetic scoring, and professional scoring branches, and uses a weighted loss function for joint optimization to improve the adaptability of recommendations in multiple scenarios.

[0032] III. Model Validation and Optimization The validation metrics used are precision, recall, and matching accuracy. The test set precision is required to be ≥92%, recall ≥90%, and matching accuracy ≥91%. To address the issue of mismatched body shapes during training, data augmentation was used to expand the sample of less common body shapes. 100,000 sets of augmented data were generated through feature interpolation, and the model was fine-tuned. After the model training was completed, it was quantized and compressed to 180MB to adapt to edge deployment. The inference time was ≤1.2 seconds, meeting the requirements of real-time recommendation.

[0033] First, the product sales system completes multi-dimensional product information access. The accessed product information includes two categories: one is product information entered internally by the product sales system, covering basic data such as style, size, color, fabric, fit, style, price, inventory, and wearing scenarios of clothing sold in stores / platforms, which is uniformly entered and updated in real time by the merchant's backend; the other is product information obtained from the Enterprise Resource Planning (ERP) data interface accessed by at least one brand owner. This information is connected to the brand owner's ERP system through a standardized API interface, synchronizing the brand owner's full-category product data, inventory dynamics, new product information, supply chain data, etc. in real time, breaking down product data silos, realizing the integrated access of the entire product pool, and providing sufficient product data support for intelligent recommendations.

[0034] The first recommendation agent retrieves the shape analysis data generated in step 100, combines it with the accessed global product information, and generates the first product recommendation information through one or more of the following combined logics to adapt to the needs of different shopping guide scenarios: The first recommendation logic is based on matching people and products according to body shape tags. The first recommendation agent extracts body shape tags, style tags, and core proportion parameters from the body shape analysis data to construct a user's body shape feature vector. At the same time, product information is transformed into product feature vectors. The intelligent agent invokes the system's built-in user-product matching rule library. This rule library is generated based on clothing and fashion industry standards and trained using over 100,000 clothing data points. It calculates the matching score between users and products using a cosine similarity algorithm, as shown in the following formula: The matching score ranges from 0 to 1, with a higher score indicating a higher degree of compatibility.

[0035] The intelligent agent sets a matching threshold of 0.7, filters products with a score ≥ 0.7, sorts them by matching degree from high to low, and generates a first product recommendation list. The list fields are specified in the table below to ensure that the recommended products fit the customer's body shape and avoid problems such as unsuitable size or incorrect fit. List fields Field Description Data Example Product ID Unique Product Identifier SP202603001 Product Name full name of clothing style High-waisted slimming A-line skirt Suitable size Match the customer's optimal size M size Matching score Cosine similarity calculation results 0.92 Outfit Highlights The core advantage of fitting the body shape Corrects wide hips and optimizes waist-to-hip ratio The second recommendation logic is based on a combination of outfit formulas and aesthetic agent scoring. The first recommendation agent uses the system's built-in outfit formula library, which includes over 1200 standardized outfit formulas covering single-item combinations, layering, and scenario-based outfits. These formulas are categorized into four scenarios: commuting, dating, casual daily wear, and formal work. Each formula includes rules for item combination, fit matching, and color matching. Combining body shape analysis data with comprehensive product information, the system first creates an initial product combination using the outfit formulas. Then, the recommended products are loaded into a virtual model to replicate the customer's body shape, generating initial try-on data. Subsequently, the aesthetic agent is used to quantitatively score the try-on effect across multiple dimensions. The aesthetic agent scoring system and weights are shown in the table below. The total score is the sum of the scores for each dimension multiplied by their weights, with a maximum score of 100. Rating Dimensions Scoring Criteria Weighting percentage Score range Fit The clothing fits the customer's body shape perfectly, without feeling too tight or too loose. 35% 0-35 points Color coordination The clothing colors complement the skin tone and body shape, without making the wearer appear darker or bulkier. 25% 0-25 points Stylistic consistency The degree to which clothing style matches the customer's style tag 20% 0-20 points Visual slimming effect Clothing's ability to flatter the figure and optimize body proportions. 20% 0-20 points The intelligent agent sets a scoring threshold of 80 points, removes products and outfit combinations with a total score below 80 points, filters high-scoring and high-quality content, and generates a first product recommendation list to ensure the aesthetics and practicality of the recommended products.

[0036] The third recommendation logic: multi-professional intelligent agent collaborative scoring and screening. After the first recommendation intelligent agent completes the initial product screening, it synchronizes the virtual try-on effect data to a professional collaborative module composed of a designer intelligent agent, a stylist intelligent agent, and a fashion intelligent agent. The three types of intelligent agents independently score from different dimensions, and the scores are all out of 100. Then, a comprehensive collaborative score is calculated according to a preset weighted formula, as follows: ,in Rate the designer agent. To rate the stylist's AI agent, Rate the fashionable intelligent agent.

[0037] The three types of AI agents have detailed scoring dimensions: The designer AI agent focuses on the fit of clothing design, tailoring, fabric drape, and the customer's body shape, emphasizing whether the fit conforms to the body's curves and whether the tailoring optimizes body flaws; the stylist AI agent focuses on the overall layering of the outfit, the suitability of individual items, and the coordination of the upper and lower body proportions, avoiding cluttered or unbalanced outfits; the fashion AI agent focuses on trend attributes, scene suitability, and popularity, aligning with current fashion trends. The AI ​​agents have a comprehensive scoring threshold of 85 points. Based on the comprehensive score, items are sorted from highest to lowest, low-scoring items are removed, and a final product recommendation list is generated, achieving professional-level outfit recommendations and enhancing the professionalism and customer acceptance of the recommended content.

[0038] The first product recommendation information generated in this step is personalized and accurate, abandoning the traditional generalized recommendation model and perfectly matching the customer's body characteristics. It also supports custom adjustments to the number of recommendations and the categories of recommended products, adapting to different scenarios such as quick recommendations from offline shopping guides and refined recommendations from online sources, laying the foundation for subsequent virtual try-on and shopping guide strategy generation.

[0039] Step 102: The second effect generation agent generates the first usage effect generation data of the recommended product based on the shape analysis data and the first product recommendation information. In the product sales system of this embodiment, the second effect generation intelligent agent is the core module for generating virtual try-on effects. It integrates technical solutions such as image rendering, face migration, detail optimization, super-resolution enhancement, and edge-cloud collaborative rendering. Preferably, it can generate highly realistic and high-definition first-use effect generation data of recommended products through the following two independent and switchable implementation methods, intuitively presenting the real effect of customers wearing recommended products, and solving the pain points of cumbersome offline try-on and inability to perceive the effect of wearing the product online.

[0040] In this embodiment, the preferred second effect generating agent includes three sub-models: image generation, face transfer, and super-resolution enhancement, all of which employ specialized training schemes, as detailed below: I. Training of the Virtual Try-on Image Generation Sub-model 1. Dataset Construction: A dataset of 700,000 clothing and body shape images was constructed, including images of different body shapes and clothing styles. Human body key points, clothing outlines, and occlusion relationships were annotated. High-resolution training images (≥1080P) were included, covering multiple angles (front, side, back) and considering different skin tones and body shapes to avoid generating uniform effects. 2. Training Architecture: The model was built based on the ControlNet diffusion model, incorporating human pose control and body contour control modules. The training framework was TensorFlow, and the hardware environment consisted of an NVIDIA A100 graphics card. Training parameters included: batch size = 32, learning rate = 1e-4, DDIM sampler, sampling steps = 25, and training epochs = 60. 3. Training Process: Using body shape features and product features as input conditions, the model was trained to generate clothing and body shape fitting images. The loss function employed was perceptual loss + pixel loss + contour loss to optimize image restoration accuracy. The generated images achieved a clothing fit error ≤2% and a body shape restoration accuracy ≥95%.

[0041] II. Face Transfer and Optimization Sub-model Training 1. Dataset Construction: A dataset of 300,000 anonymized facial images was collected, covering different face shapes, skin tones, and expressions. 68 facial landmarks were annotated, and samples were categorized by age group and gender to ensure feature diversity. 2. Training Architecture: Based on the lightweight StyleGANv3 model, combined with the GFPGAN face restoration module, the underlying facial feature extraction layer was frozen, and the transfer and restoration layers were fine-tuned to prevent facial distortion. 3. Training Process: Adversarial loss + feature matching loss were used for training to optimize facial feature transfer accuracy. The requirements were that the facial similarity after transfer ≥90%, with no unnatural distortions, and a noise removal rate ≥98% after skin restoration. 4. Validation Metrics: Facial naturalness score ≥4.8 (out of 5), skin tone consistency error ≤3ΔE.

[0042] III. Training of Super-Resolution Enhancement Sub-Model 1. Dataset Construction: Construct 200,000 pairs of low-resolution to high-resolution images, covering clothing textures and facial details, to simulate low-quality images generated by rendering; 2. Training architecture: Based on the diffusion transformer + variational autoencoder fusion architecture, a 4x super-resolution enhancement is achieved. Training parameters: learning rate = 5e-5, batch_size = 16, number of training rounds = 40. 3. Training objectives: Upgrade image resolution to 4K, preserve clothing details and facial texture, achieve a peak signal-to-noise ratio (PSNR) ≥ 32dB and a structural similarity (SSIM) ≥ 0.92 after super-resolution.

[0043] The second effect generation agent is the first implementation method for generating virtual try-on effects: single-end effect generation based on image migration and optimization. It is suitable for scenarios with stable network environments and moderate rendering speed requirements. The entire rendering process takes ≤3 seconds. The specific process is as follows: The first step is body matching and basic image generation. The second step is to generate an intelligent agent based on the customer's body analysis data, including core parameters such as height, shoulder width, and waist circumference, using the Euclidean distance formula. The system matches a standard wearing model with a body shape and proportion that are highly compatible with the model library, with a matching error of ≤1cm. The recommended clothing in the first product recommendation information is loaded into the wearing model model to generate basic first image data with a resolution of 1080P. This image data fully presents the overall shape, style and color of the clothing after it is worn, and fits the customer's body shape characteristics. The second step is face detection and alignment. The mature MTCNN face detection and 68-point key point alignment model is used to accurately identify the facial region in the first image data and locate key points such as eyebrows, eyes, mouth, nose, and contours. The face masking technology is used to achieve accurate selection and angle alignment of the facial region. The rotation alignment error is ≤2°, eliminating the rendering incongruity caused by posture deviation. The third step is facial feature replacement. The lightweight StyleGAN face-swapping model is called to transfer the facial features in the customer's body analysis data to the model's face in the first image data. The customer's own facial structure and facial feature proportions are strictly preserved. The feature transfer similarity is ≥90%, avoiding facial deformation, incongruity and distortion, and achieving a high degree of restoration of "body shape + face". The fourth step is to optimize the details. By using edge expansion algorithm and Gaussian blur processing, a natural transition between the facial area and the model's body and clothing is achieved. The blur radius is set to 1.5px to eliminate splicing marks. At the same time, the GFPGAN face repair model is used to finely repair the texture and texture of the facial skin, remove noise and blemishes, and restore the real skin color and texture. The fifth step is super-resolution enhancement of the output. A super-resolution network is built based on the diffusion transformer model and the variational autoencoder model to enhance the resolution of the facial region and clothing details in the first image data. This is achieved through a magnification formula. Upgrading image resolution from 1080P to 4K using color calibration formulas ( For skin tone calibration coefficient, (For color difference compensation value), calibrate the skin tone consistency of the face, neck, and clothing, with a color difference ≤3ΔE. The final output generates the first usage effect data, which is in high-definition PNG image or MP4 short video format and can be directly displayed on the shopping guide.

[0044] In this embodiment, the second effect generating agent is another preferred implementation of the first implementation method for generating virtual try-on effects (single-end effect generation based on image transfer and optimization), as follows: A model is matched to the body shape analysis data, and the first image data of the model wearing the recommended product is generated. Face detection and alignment: The RetinaFace-ResNet50 model is used to detect the facial region in the first image data, and alignment is achieved through jonathandinu masking. Feature replacement: The inswapper_128.onnx model is used for facial feature transfer to preserve the facial structure of the customer target image in the body shape analysis data. Detail optimization: Natural transition is achieved through 32-pixel edge expansion and 16-pixel blurring, and the skin texture of the facial region in the first image data is restored by combining the GPEN-BFR-512.onnx model. Super-resolution enhancement: The output resolution is increased to 2048+ based on the DiT (DiffusionTransformer) model and the VAE (Variational Autoencoder) model to ensure the skin color consistency of the facial region in the first image data, and the first usage effect generation data is generated.

[0045] Preferably, the second effect generating intelligent agent is a second implementation method for generating virtual try-on effects: edge-cloud collaborative rendering, which is suitable for scenarios where offline intelligent shopping guide devices have insufficient local computing power and pursue the ultimate rendering details. The specific process is as follows: The first step is to trigger a local rendering request. After the customer confirms and views the fitting effect on the smart shopping guide device, the fitting rendering request is triggered, and the second effect is to generate a smart agent to respond quickly. The second step is local basic rendering. The second effect generation intelligent agent sends a try-on command to the local rendering engine on the intelligent shopping guide device. It directly calls the locally encrypted body analysis data, without uploading the original body data to the cloud. The first try-on rendering is quickly completed through the local lightweight rendering model to generate the first basic image data. This data focuses on presenting the overall wearing effect of the clothing. The rendering time is controlled within 1 second to ensure response speed. The third step is to extract desensitized features. On the smart shopping guide device, the body analysis data is desensitized, the core body feature parameters are extracted, privacy-related information is removed, and the first desensitized body feature is generated to prevent the original privacy data from being transmitted. The fourth step is cloud-based detail rendering. The first desensitized body features, the first product recommendation information, and the detail rendering requirements are uploaded to the high-performance detail rendering engine in the cloud. The cloud uses powerful computing power to perform fine rendering, optimizing details such as the texture, wrinkles, luster of clothing fabrics, and facial micro-expressions, and generating the second detail enhancement map data. The fifth step is image fusion output, which involves pixel-level fusion of the locally generated first base image data and the cloud-generated second detail enhancement image data, balancing rendering speed and detail quality, and finally generating the first usage effect data of the recommended product. This method not only ensures data privacy and security, but also achieves high-quality try-on effect presentation.

[0046] The first-use effect data generated in this step has high fidelity and realism. It can dynamically display the front, side, and back effects of customers wearing recommended products, and supports zooming in to view details, allowing customers to intuitively perceive the effect of the clothing on their bodies. At the same time, the data is adapted to the display requirements of the sales guide, making it easier for sales guides to combine the effect data to provide services, thereby improving the customer's shopping experience and decision-making efficiency.

[0047] Step 103: Based on the customer's user tag data, the first product recommendation information, the first usage effect generation data, and the customer's interaction feedback information in the product sales system, generate the first sales guide recommendation strategy data on the sales guide side. In the product sales system of this embodiment, the sales guide side is a dedicated operating terminal for sales guides, including a PC-based management backend, a mobile sales guide APP and its control interface, etc. It has the function of receiving data and synchronizing information with intelligent sales guide devices, and can also generate and display sales guide strategies and interactively input information for sales guides. Preferably, this step can integrate multi-dimensional data, generate the first sales guide recommendation strategy data through two adaptive logics, and simultaneously link with the brand's ERP system to achieve data feedback, forming a closed loop for sales guide services.

[0048] First, the system retrieves customer user tag data, which forms a profile of the customer's historical behavior in the product sales system. This data includes historical purchase records, clothing preferences, price range, shopping frequency, categories of interest, and historical feedback. This data, along with body shape analysis data, first-recommendation information, and first-use effect data, forms a multi-dimensional data support. Simultaneously, the system collects real-time interactive feedback information from customers in the product sales system, including likes / favorites / skips of recommended products, satisfaction ratings of fitting effects, size modification requests, style preference adjustments, inquiries, and purchase intentions. This comprehensive data captures customers' shopping needs and real-time attitudes.

[0049] Based on the above multi-dimensional data, the first-level shopping guide recommendation strategy data is generated in the following two ways to adapt to the service needs of different shopping guide scenarios: The first generation method: Generation based on a process-oriented strategy using the first sales guide state machine. The first sales guide state machine includes six standardized sales guide stages: welcoming guests, product recommendation, try-on display, objection handling, order closing, and after-sales guidance. It presets the transition conditions for each stage, forming a closed-loop sales guide process. The stages and transition conditions are shown in the table below: Sales guidance stage status Current state core actions Flow triggering conditions Post-transfer status Welcome and reception Greet customers and guide them to authorize the collection of body data. Customer completes body data authorization submission Product Recommendations Product Recommendations Display the first recommended product information Customers click to view the fitting effect Fitting Showcase Fitting Showcase Demonstrating the first use effect and generating data Customer raises questions / expresses intention Handling of objections Handling of objections Answer customer questions and adjust recommendations Customers alleviate concerns and confirm their purchase intention facilitating order placement Throughout the entire shopping guide service process, the first shopping guide state machine synchronizes customer user tag data, initial product recommendation information, initial usage effect generation data, and interactive feedback information at each stage in real time. It dynamically determines whether the state transition conditions are met and automatically completes the stage switch. Simultaneously, the state machine has 36 built-in prompt word templates, categorized and stored according to the shopping guide stage. These templates cover shopping scripts, key recommendation points, explanation logic, objection handling, and styling suggestions. For different shopping guide stages after the transition, the matching prompt word template is invoked, and corresponding first shopping guide recommendation strategy data is generated based on customer body characteristics and product information. This strategy data includes standardized shopping scripts, core recommended selling points, key points for explaining the fitting effect, targeted styling suggestions, and objection handling solutions, guiding shopping guides to provide standardized and precise services. It is especially suitable for novice shopping guides, reducing service difficulty and improving service efficiency.

[0050] The second generation method: intelligent strategy generation based on a third-party intelligent sales agent. This third-party intelligent sales agent is a professional sales guide module driven by a large language model. Trained on massive amounts of apparel sales scenario data, it possesses functions such as real-time data analysis, adaptive strategy generation, and personalized script optimization. On the sales guide side, the third-party intelligent agent acquires real-time information on the sales guide stage status, customer user tag data, first product recommendation information, first usage effect generation data, and customer interaction feedback. Through deep learning algorithms, it quickly analyzes the customer's core needs, purchase intentions, and potential concerns, abandoning the limitations of standardized templates and adaptively generating personalized first-party sales guide recommendation strategy data. This strategy data aligns with the customer's personality preferences and real-time status, resulting in more flexible scripts and more accurate recommendations. It also highlights key service precautions, adapting to experienced sales guides to provide refined and personalized services, improving the targeting and conversion rate of sales guide services.

[0051] In this embodiment, the preferred training scheme for the shopping guide intelligent agent (first shopping guide state machine, third shopping guide intelligent agent) is as follows: I. Third-Party Shopping Agent Training Solution 1. Dataset Construction: A dataset of 800,000 dialogues in clothing sales scenarios was constructed, covering all scenarios including greeting, product consultation, fitting explanation, objection handling, and closing the sale. Customer intent, sales strategies, and script suitability were annotated, and professional content such as body shape recommendations, fitting effect explanations, and size matching were integrated. The samples cover customers with different personalities (introverted, straightforward, and indecisive). 2. Training Architecture: Based on the Qwen-7B lightweight large model, LoRA (Low-Rank Adaptation) was used for fine-tuning. Rank=8, alpha=16. The backbone model was frozen, and only the LoRA layer was fine-tuned to reduce training costs. 3. Training Parameters: Learning rate=3e-4, batch_size=128, number of training epochs=30, Adam optimizer was used, and cross-entropy loss was used as the loss function, focusing on the accuracy of sales scripts and scenario suitability. 4. Model optimization: To address the issue of insufficient professionalism in sales pitches, a clothing and styling knowledge base distillation was added to improve professional explanation capabilities. Reasoning time is ≤0.5 seconds, and the accuracy of speech adaptation is ≥94%.

[0052] II. First-level sales guide state machine training optimization scheme The first-stage sales guide state machine adopts a rule-based + reinforcement learning fusion optimization model: 1. Basic rule construction: Based on 100,000 sets of real sales guide process data, it mines the flow patterns of each stage and constructs an initial state flow rule base; 2. Reinforcement learning training: It uses customer satisfaction and order conversion rate as reward values, with a reward function... 3. Iterative optimization: Based on real shopping guide interaction data, continuous iteration is carried out to achieve a state transition accuracy rate of ≥96%, adapting to different shopping guide process scenarios and reducing the error rate of novice shopping guide services.

[0053] In addition, in this embodiment, this step simultaneously initiates the data analysis and feedback process of the fourth product analysis intelligent agent: the fourth product analysis intelligent agent captures real-time sales data from the entire product sales system, including recommended product click-through rate, trial rate, purchase conversion rate, return rate, repurchase rate, etc., and integrates customer interaction feedback information in the system to construct a data analysis indicator system. The core indicator calculation formula is as follows: , , By using cluster analysis and association rule algorithms to mine product suitability, customer pain points, best-selling categories, pattern optimization suggestions, and size adjustment directions, standardized analysis results data is generated in CSV format compatible with the brand's ERP system. This analysis results data is then transmitted in real-time to the brand's ERP system's large-scale product analysis model via HTTPS encrypted data interface through the brand's external ERP standardized API. This provides data support for the brand's product development, inventory allocation, new product planning, marketing strategy formulation, and pattern iteration optimization, achieving two-way linkage between terminal sales data and brand operations management. This improves the entire product sales and sales service system, effectively reducing brand inventory backlog and product return rates.

[0054] In this embodiment, the preferred embodiment discloses the training technology solution for the fourth commodity analysis agent as follows: I. Dataset Construction A dataset of 600,000 sets of product sales and interaction feedback was constructed, covering product sales data, customer interaction data, return feedback data, and inventory data. Product compatibility issues, pattern defects, pain points, and best-selling tags were labeled. Data cleaning was completed, abnormal data was removed, missing values ​​were added, and feature encoding was used to transform categorical data into numerical features.

[0055] II. Model Training and Architecture The model employs a fusion of XGBoost and cluster analysis. 1. Training process: First, clustering algorithms are used to classify products and customers. Then, an XGBoost regression model is used to predict product suitability and best-selling trends. Training parameters: learning rate = 0.1, number of decision trees = 100, maximum depth = 6; 2. Loss function: A logarithmic loss function is used to optimize analysis accuracy, achieving an accuracy of ≥89% for analyzing reasons for product returns and pattern issues; 3. Deployment optimization: The model is lightweight, supporting real-time data stream analysis. Single batch data analysis time is ≤2 seconds, and analysis results can be fed back to the brand's ERP system in real time.

[0056] All of the above-mentioned intelligent agents adopt an offline training + online fine-tuning mode. After deployment, a fine-tuning iteration is performed monthly based on newly added real business data to continuously optimize model performance and ensure the long-term stability and adaptability of the technical solution. Those skilled in the art can complete the training and deployment of intelligent agents without creative labor based on the above training scheme and achieve the technical effects described in the patent.

[0057] The first guide recommendation strategy data generated in this step, compared to the traditional guide service model that relies solely on experience, realizes the data-driven, precise, and intelligent nature of guide services. This not only improves the service efficiency and professionalism of guides, but also allows customers to receive personalized guide services tailored to their needs, speeding up shopping decisions, reducing clothing return rates, and enabling guide data to empower the brand and improve operational management efficiency.

[0058] Example 2: To more clearly reveal the structure of the commodity sales system involved in this invention, Figure 2 A schematic diagram of the structure of the commodity sales system implemented in Embodiment 2 of the present invention is provided.

[0059] Embodiment 2 of the present invention discloses a product sales system 200, which includes a body data acquisition and analysis module 201, a first product recommendation module 202, a first effect generation module 203, and a sales guide recommendation strategy data generation module 204. The body data acquisition and analysis module 201 is configured in an intelligent sales guide device 205 within the product sales system, and is configured to acquire body data confirmed by the customer and generate the customer's body analysis data. The first product recommendation module 202 is configured to use a first recommendation agent to generate first product recommendation information for the customer based on the body analysis data and product information accessed in the product sales system. The first effect generation module 203 is configured to use a second effect generation agent to generate first usage effect generation data for the recommended product based on the body analysis data and the first product recommendation information. The sales guide recommendation strategy data generation module 204 is configured in a sales guide terminal 206 within the product sales system, and is configured to generate first sales guide recommendation strategy data for the sales guide terminal 206 based on the customer's user tag data, the first product recommendation information, the first usage effect generation data, and the customer's interactive feedback information within the product sales system.

[0060] The following section provides a complete example of the product sales system 200 and its constituent modules: body data acquisition and analysis module 201, first product recommendation module 202, first effect generation module 203, sales guide recommendation strategy data generation module 204, intelligent sales guide device 205, and sales guide terminal 206. Using a physical mid-to-high-end women's clothing store as the actual application scenario, this example illustrates the system's implementation. This store primarily sells commuter and light luxury style women's clothing. After deploying this product sales system, it achieves a full-link sales guide service encompassing intelligent body data collection, accurate product recommendation, high-fidelity virtual try-on, and intelligent generation of sales guide strategies. In this example, the complete business process is exemplified by Ms. Zhang purchasing spring commuter clothing. The hardware deployment of each module is as follows: the intelligent sales guide device 205 is a 43-inch intelligent fitting mirror in the store's fitting area; the sales guide terminal 206 is a 10.1-inch industrial touchscreen tablet held by the sales guide; the system is deployed on an Alibaba Cloud ECS server; the edge computing unit is integrated into the intelligent fitting mirror; and the brand's ERP system is Yonyou NC Cloud, which connects with this system via standardized APIs.

[0061] In this embodiment, the product sales system 200 adopts a three-level distributed architecture of cloud-edge-device. The device side consists of intelligent shopping guide devices 205 and shopping guide terminals 206, which are responsible for data collection, interactive display, and local lightweight computing. The edge side consists of the edge computing unit built into the intelligent shopping guide device, which is responsible for shape data preprocessing, local rendering, and low-latency data interaction. The cloud side consists of Alibaba Cloud ECS servers (configured with 8 cores, 32GB of memory, and NVIDIA T4 graphics cards), which are responsible for intelligent agent model operation, massive product data storage, cloud-based high-precision rendering, data analysis, and feedback.

[0062] The system's overall operation logic is as follows: edge data acquisition → edge preprocessing → cloud intelligent computing → edge display and strategy output → data feedback to the brand's ERP system. Each module interacts with encrypted data via 5G Wi-Fi. Edge-to-cloud data transmission uses HTTPS + AES-256 dual encryption, and edge-side local data storage uses the national standard SM4 encryption algorithm to ensure the security of customer privacy data and the brand's business data. The system supports 24 / 7 uninterrupted operation, can handle ≥5 users simultaneously on a single device, and has an average edge-to-cloud interaction latency of ≤150ms, fully meeting the business needs of real-time sales guidance in offline stores.

[0063] The product information accessed by this system includes two categories: first, the self-owned product information entered by the store backend (a total of 2000+ SKUs, covering tops, bottoms, skirts, accessories, etc.); and second, product information obtained from the ERP systems of 3 partner brands through API interfaces (a total of 1500+ SKUs). The product information is synchronized every 10 minutes to ensure the real-time availability of inventory, price, and new product information, and to achieve unified management and access to the entire product pool.

[0064] Intelligent shopping guide device 205 and body data acquisition and analysis module 201 (a) Intelligent shopping guide equipment 205 In this embodiment, the intelligent shopping guide device 205 is a customized intelligent fitting mirror. Its hardware configuration includes: a 43-inch 4K touch screen (supporting 10-point touch), an RGB dual-lens depth camera (collecting data at 30fps, 12 million pixels), a light scanning module (scanning accuracy ±0.2cm), an edge computing unit (Rockchip RK3588, 8-core CPU + 6-core GPU, 8GB RAM), a human infrared sensing module (sensing distance 0.5-3m), a dual-channel microphone, a built-in Android 14.0 operating system, pre-installed merchandise sales system client-side program, and integration of all functions of the body data acquisition and analysis module 201. The device is deployed in the fitting area of ​​the store and supports standing non-contact body data collection, thereby obtaining body data of users authorized and confirmed by customers.

[0065] (II) Shape Data Acquisition and Analysis Module 201 The body data acquisition and analysis module 201 is the core functional module of the intelligent shopping guide device 205. It consists of four sub-units: data authorization unit (not shown), body acquisition unit (not shown), data preprocessing unit (not shown), and body analysis unit (not shown). The module is implemented using modular programming, developed based on Python + OpenCV, and runs on the edge computing unit of the intelligent shopping guide device. The total time for body data acquisition and analysis for a single user is ≤5 seconds, which perfectly matches the shopping guide rhythm of offline stores.

[0066] In this example, Ms. Zhang entered the store's fitting area. After the infrared sensor module detected her body, the smart fitting mirror's touchscreen automatically woke up and displayed the "Body Data Collection Authorization Agreement." After Ms. Zhang clicked "Confirm Authorization," the module started the body data collection process. The collected body data includes three categories: body shape dimensions, body part proportions and shapes, and facial features. The specific collection and analysis process is as follows: Body shape and size data acquisition: The light scanning module performs a 360° non-contact scan on Ms. Zhang to collect core dimensions such as height, shoulder width (38cm), chest circumference (88cm), waist circumference (66cm), hip circumference (92cm), arm length (62cm), and pant length (98cm). After the scan data is calibrated for light and distance, the size accuracy is controlled within ±0.3cm. Body proportion and shape data analysis: Through the built-in body recognition model, the waist-to-hip ratio is calculated to be approximately 0.72 (66 / 92) and the bust-to-waist ratio to be approximately 1.33 (88 / 66). The body type is identified as pear-shaped, the shoulder type as standard, the leg type as straight, and the style matching tags as commuter style and light luxury style. Facial feature data acquisition: RGB binocular cameras captured Ms. Zhang's frontal facial image. Through a 68-point facial key point detection model, feature data such as face shape (oval), skin tone (cool white), and facial feature proportions (standard proportions of the three courts and five eyes) were extracted to generate a 512-dimensional facial feature vector.

[0067] The three types of data mentioned above are processed by the data preprocessing unit (Gaussian filtering for noise reduction and neighborhood interpolation for completion) and then transmitted to the body analysis unit. This unit generates Ms. Zhang's body analysis data based on a lightweight CNN model. The data is in JSON format and includes three core parts: a quantified size table, a body label set, and clothing matching suggestions. For example, the clothing matching suggestions are labeled as "suitable for high-waisted A-line bottoms and H-line tops, avoid tight-fitting pencil skirts and low-waisted pants, and suitable for cool-toned fair skin with Morandi colors, navy blue, and burgundy." The body analysis data is encrypted and stored in the edge computing unit, and at the same time, a desensitized feature vector is generated and uploaded to the cloud to provide core data support for subsequent product recommendations and virtual try-on.

[0068] First Product Recommendation Module 202 The first product recommendation module 202 is deployed on the cloud server of the product sales system 200. It consists of a product data access unit, an intelligent agent operation unit, and a recommendation list generation unit. The module is developed based on the PyTorch framework. The core of the module is the first recommendation intelligent agent (based on the Llama3-8B apparel vertical domain fine-tuning model). The module supports real-time data docking with the brand's ERP system to ensure the comprehensiveness and real-time nature of product information.

[0069] In this example, after Ms. Zhang's body shape analysis data is uploaded to the cloud, the first product recommendation module 202 initiates the recommendation process. First, it retrieves commuter-related product information (a total of 800+ SKUs) from the entire product pool (stores + 3 partner brands) through the product data access unit. Then, the first recommendation agent combines the body shape analysis data and uses a combination of human-product rule matching and multi-agent collaborative scoring logic to generate the first product recommendation information. The human-product matching uses the cosine similarity algorithm to calculate the matching degree between user features and product features. The calculation formula is as follows: In the formula: The matching degree between users and products, with a value range of [0,1]. For the user's i-th shape / style feature value; is the i-th pattern / style feature value of the product; n is the total number of feature dimensions, in this embodiment n=36.

[0070] The specific recommendation process is as follows: Matching of people and goods according to rules: The intelligent agent extracts features such as Ms. Zhang's pear-shaped body, commuting style, and fair skin to construct a 36-dimensional user feature vector. The matching degree with all products in the domain is calculated by the cosine similarity algorithm mentioned above. A total of 50 products with a matching degree ≥ 0.8 are selected to generate a preliminary recommendation list. Multi-agent collaborative scoring: Products from the initial recommendation list are loaded onto a virtual model with a similar physique to Ms. Zhang for a fitting. The results are then simultaneously sent to a designer agent (not shown), a stylist agent (not shown), and a fashion agent (not shown) for collaborative scoring. The scoring weights are 40% for the stylist, 35% for the designer, and 25% for the fashion agent. The overall score is calculated using the following formula: Ten products with an overall score of ≥85 were selected and the final top product recommendation information (recommendation list) was generated. The list includes product ID, name, suitable size, reason for recommendation, and styling suggestions. For example, the top recommended product is "Navy Blue High-Waisted A-Line Skirt (Brand: XX, SKU: TC20260301, Size M, Matching Degree 0.95, Overall Score 92). Reason for recommendation: The high-waisted design enhances the waistline, the A-line shape flatters wide hips, navy blue is suitable for cool-toned fair skin, and it is suitable for commuting style. It looks even better when paired with a silk shirt from the same brand."

[0071] After the first product recommendation information is generated, it is simultaneously pushed to the smart shopping guide device 205 (fitting mirror touch screen) and the shopping guide terminal 206 (shopping guide tablet). Ms. Zhang can directly click on the fitting mirror to view product details and virtual try-on effect, and the shopping guide can view the complete recommendation information and matching suggestions on the tablet.

[0072] First Effect Generation Module 203 The first effect generation module 203 is deployed across the smart shopping guide device 205 (edge ​​side) and the cloud server. It consists of a local rendering unit, a cloud rendering unit, an image fusion unit, and a face optimization unit. The module is developed based on Unity3D+Diffusers. The core runs the second effect generation intelligent agent (composed of three sub-models: image generation, face migration, and super-resolution enhancement). It supports the real-time generation of high-fidelity virtual try-on effects. The generated effect image resolution is 4K, the face similarity is ≥95%, and the clothing restoration accuracy is ≥98%.

[0073] In this example, Ms. Zhang clicks on the fitting mirror to view the virtual try-on effect of the top-recommended "navy blue high-waisted A-line skirt". The module uses a method of edge-cloud collaborative rendering as the main approach and image migration and optimization as a secondary approach to generate the first usage effect data. The specific process is as follows: Local base rendering: The second effect generation agent sends a try-on command to the local rendering unit of the smart shopping guide device 205, calls the locally stored Ms. Zhang's body analysis data, and quickly generates the first base image data (1080P, taking 0.8 seconds) to present the basic upper body shape and overall effect of the dress. Desensitized feature extraction: While rendering locally, the first desensitized body features are extracted from the body analysis data (only body parameters are retained, and face and identity information are removed), and the feature vector dimension is 1024. Cloud-based detail rendering: The desensitized body features, product SKU, and rendering requirements (4K, front / side / back, natural lighting effects) are uploaded to the cloud rendering unit. The high-performance rendering engine generates a second detail enhancement map (4K, taking 1.2 seconds) to restore the fabric texture (suit fabric), drape, pleat details, and fit of the dress to Ms. Zhang's body shape. Image fusion and optimization: The image fusion unit performs pixel-level fusion of the first base image data and the second detail enhancement image data. The face optimization unit uses the RetinaFace face detection model + inswapper face swapping model to transfer Ms. Zhang's facial features to the try-on image. The face transition is achieved through edge expansion (20 pixels) + Gaussian blur (10 pixels). Then, super-resolution enhancement is performed based on the DiT+VAE model to ensure skin color consistency (color difference ≤ 2ΔE). Finally, the first usage effect generation data is generated (3 4K high-definition try-on images + 15-second dynamic try-on video).

[0074] The data is displayed in real time on the smart fitting mirror touch screen. Ms. Zhang can zoom in to view fabric details and switch between different angles to see the fitting effect. At the same time, the data is synchronized to the sales assistant's terminal 206, and the sales assistant can provide Ms. Zhang with professional explanations based on the fitting effect.

[0075] Shopping guide module 206 and shopping guide recommendation strategy data generation module 204 (a) Sales guide end 206 In this embodiment, the sales guide terminal 206 is an industrial-grade touch tablet held by the sales guide. The hardware configuration includes: a 10.1-inch 2K touch screen, a Snapdragon 870 processor, 8GB of RAM, 512GB of storage, a 5G module, and Bluetooth 5.3. It is pre-installed with a merchandise sales system sales guide program, which supports real-time data synchronization with the intelligent sales guide device 205, sales guide strategy display, and customer information entry. The tablet has a built-in barcode scanner module, which can quickly scan product barcodes to retrieve details. The battery life is ≥8 hours, which fully meets the needs of sales guides in offline stores.

[0076] (II) Shopping guide recommendation strategy data generation module 204 The shopping guide recommendation strategy data generation module 204 is the core functional module of the shopping guide terminal 206. It is deployed on the tablet locally and in the cloud collaboratively. It consists of a data acquisition unit, a shopping guide state machine unit, a third shopping guide intelligent agent unit, and a strategy output unit. The module is developed based on Java and Python. It can integrate two shopping guide strategy generation methods and link with the fourth product analysis intelligent agent to realize the accurate generation of shopping guide strategies and the reverse feedback of sales data.

[0077] In this example, the module first collects multi-dimensional data: Ms. Zhang's user tag data (silver card member of the store, 35 years old, 3 previous purchases, all in workwear, average transaction value of 800-1500 yuan, style preference for light luxury workwear), first product recommendation information, first usage effect generation data, and interaction feedback information (Ms. Zhang said after viewing the fitting effect, "The skirt has a good fit, I want to see the matching top, but I'm worried the color will make me look older"). Based on the above data, the module simultaneously activates the first sales guide state machine and the third sales guide intelligent agent to generate targeted first sales guide recommendation strategy data, providing real-time service guidance for sales staff. The following are detailed technical implementation schemes, model training details, and specific strategy examples for the two generation methods: (I) Technical Implementation and Strategy Generation of the First Shopping Guide State Machine 1. Generation scheme and training details of the first shopping guide state machine In this embodiment, the first sales guide state machine is generated by a fusion of Finite State Machine (FSM) and Q-Learning reinforcement learning. Unlike the PPO algorithm in Embodiment 1, this solution uses the completion rate of sales guide services and customer satisfaction as the core reward values. It optimizes the state transition rules through iterative training to adapt to the dynamic sales guide scenario in offline stores. The specific generation and training steps are as follows: Step 1: Define the State and Action Space: Based on the entire offline women's clothing sales process, we divide the space into 6 core states, each corresponding to 5 standardized actions. See the table below for details: Status coding Shopping guide status name Core execution actions State triggering conditions S1 Welcome and reception Greeting, identity verification, authorization guidance Customers enter the sensing range of the smart shopping guide device S2 Demand mining Scene inquiry, style confirmation, budget communication Customer completes body data authorization collection S3 Product Recommendations Product description, reasons for recommendation, and pairing suggestions Shape analysis data generation, first product recommendation information output S4 Fitting Showcase Explanation of fitting effect, angle switching, and detail display Customers click to view the virtual try-on effect of recommended products. S5 Handling of objections Q&A, address concerns, and alternative recommendations Customers raise questions / concerns related to the product S6 facilitating order placement Promotional information, sales guidance, and cross-selling The customer had no significant objections and approved of the product / fitting effect. Step 2: Construct an initial state transition rule base: Based on 50,000 sets of real offline women's clothing sales process data, mine the transition logic between each state, use IF-THEN structured logic to write initial rules, store them in the module's local SQLite database, and the rules cover more than 90% of common sales scenarios; Step 3: Q-Learning Reinforcement Learning Training: Define the state transition reward function, quantify the rationality of state transitions, optimize the transition threshold through iterative training, and improve the accuracy of the state machine response. in, Customer satisfaction (0-1 points, calculated based on interaction feedback and behavioral data). Service completion score (0-1 point, calculated based on the completeness of the sales guide's actions). Service time (normalized to 0-1 points, longer service time results in lower scores). Training hyperparameter settings: learning rate. Discount factor Greedy probability The training process iterated for 100,000 rounds, and the accuracy of state transitions was verified every 10,000 rounds. Finally, the trigger thresholds such as "customer stay ≥10 seconds" and "product views ≥3 items" were optimized to "customer stay ≥8 seconds" and "product views ≥2 items", which greatly improved the real-time performance of state transitions. Step 4: Build a prompt word template library: Build 30 standardized prompt word templates according to 6 sales guide statuses. The templates include the script, key points of the explanation, and response plan. They support dynamic replacement of variables such as customer name, product name, and body characteristics to realize personalized strategy generation.

[0078] After training, the first sales guide state machine has a state transition accuracy of ≥97%, can capture customer interaction feedback in real time, quickly complete state switching, and call matching templates to generate compliant and accurate sales guide strategy data.

[0079] 2. State machine policy generation in this example In this example, based on Ms. Zhang's interaction feedback of "worried that the color would make her look older and wanting to match it with a top," the sales guide state machine determines that the current state is S5 objection handling. After the trigger condition is successfully matched, the state machine calls the exclusive prompt word template for the S5 objection handling stage, and combines it with Ms. Zhang's user tags, body data, and product recommendation information to complete the variable replacement. The generated sales guide strategy data is as follows: Salesperson's script: Ms. Zhang, you have excellent taste! This navy blue is our new Morandi navy blue style this year. It's softer than traditional navy blue, and it looks especially elegant on fair skin, not at all aging, and it really brightens the complexion. You'd like to see a matching top, right? I recommend this off-white silk blouse from the same brand (SKU: CS20260305). It also has a high-waisted design, and it goes perfectly with this skirt. It's a formal yet gentle look for commuting. As you can see in the fitting photos, it's especially suitable for your pear-shaped figure.

[0080] Key points: 1. Emphasize the Morandi navy blue color, which suits cool-toned fair skin and doesn't make you look old; 2. Recommend pairing it with a high-waisted silk shirt to enhance the commuting outfit; 3. Consider the suitability for pear-shaped figures and highlight the advantages of the cut.

[0081] Further response: If the customer is still concerned about the color, you can recommend the same style in light gray (SKU: TC20260302) and show the effect of trying on the light gray; if the customer approves of the combination, guide them to try on the actual product in person.

[0082] (II) Training process and strategy generation of the third-party shopping guide agent 1. Training details of the third-party shopping guide agent In this embodiment, the third shopping guide agent is based on Tongyi Qianwen Qwen-1.8B-Chat and undergoes lightweight fine-tuning. Unlike the LoRA fine-tuning in Embodiment 1, this solution uses Prefix Tuning, optimizing only the model's prefix embedding layer and freezing the backbone model parameters. This balances training efficiency and inference speed, and is suitable for local lightweight deployment on the shopping guide platform. The detailed training process and parameters are as follows: Step 1: Construct a dedicated training dataset: Collect real sales dialogues and service cases from offline mid-to-high-end women's clothing stores to construct a dedicated dataset of 600,000 offline women's clothing sales guides. The labeled fields are "customer characteristics - interaction feedback - sales guide strategy". Among them, pear-shaped body, commuting style, and light luxury consumption related samples account for 30%. After desensitization, deduplication, and cleaning, the dataset is divided into a training set (540,000 sets) and a validation set (60,000 sets) in a 9:1 ratio. Step 2: Model Initialization and Tuning Configuration: Load the Tongyi Qianwen Qwen-1.8B-Chat pre-trained model, freeze all backbone network parameters, add a trainable prefix embedding layer in the input layer, and set the prefix length. Hidden Dimensions Gradient updates are performed only on the weights of that layer; Step 3: Training Parameters and Loss Function Settings: A training environment is set up based on the PaddlePaddle framework. The core training parameters are detailed in the table below. The loss function uses a weighted fusion of dialogue generation loss and policy adaptation loss to ensure both fluency of the dialogue and that the policy aligns with customer characteristics. Training parameters Parameter values Parameter Description Batch size 32 Number of samples input in a single training session Learning rate (lr) 5e-5 Prefix layer weight update rate Optimizer AdamW Adaptive moment estimation optimizer with weight decay Weight decay (wd) 1e-4 Preventing model overfitting Number of training epochs 20 Number of iterations for training the full dataset Maximum sequence length 512 Maximum number of tokens in input and output text Formula for calculating strategy adaptation loss: ,in This is a matching indicator function for strategies and customer characteristics. A value of 1 is assigned to a successful match, and a value of 0 is assigned to a failed match. N is the number of samples. Step 4: Model Quantization and Deployment: After training, the model is compressed using 4-bit INT4 quantization technology, reducing the model size from 7GB to 1.8GB. The inference memory usage is ≤2GB. The model is deployed locally on the shopping guide terminal 206. The time for a single round of strategy inference is ≤0.3 seconds, and the strategy generation accuracy is ≥95%, which fully meets the needs of offline real-time shopping guide.

[0083] 2. Strategy generation of the third-party shopping guide agent in this example In this example, the module inputs multi-dimensional data into the third-party intelligent sales agent in a standardized Prompt format. The Prompt format is tailored to offline sales scenarios, accurately inputting core customer information and needs. Prompt: You are a top sales associate at a mid-to-high-end women's clothing store. Your current customer is Ms. Zhang, 35 years old, a silver card member, pear-shaped body, fair skin, and a preference for a sophisticated yet casual style. Her historical average transaction value is 800-1500 yuan. She is currently viewing a 4K try-on effect of a navy blue high-waisted A-line skirt and has given feedback that "the fit is good, but I'd like to see what tops to pair it with. I'm worried the color might make me look older." Please generate a targeted sales strategy, including communication scripts, styling suggestions, and handling of objections. Your communication should be natural and friendly, suitable for the offline sales environment, and tailored to the customer's body type and style preferences.

[0084] The third-party intelligent sales agent generates personalized sales strategy data based on the trained model, complementing the state machine strategy and better meeting the needs of humanized communication, as detailed below: Communication script: Sister Zhang, this navy blue midi skirt fits you perfectly! For pear-shaped figures, this high-waisted A-line style is incredibly flattering, especially for the hips. Don't worry about the color at all; it's a soft, muted Morandi navy, not a dark or somber shade. It makes fair skin look even brighter and instantly elevates your style. This color looks especially sophisticated on someone over 30, and doesn't make you look old at all! For styling, I highly recommend our off-white silk high-waisted shirt. The fabric is incredibly comfortable, and it's a perfect match for commuting. Plus, the shirt has a small V-neck, which also flatters the neckline. I'll show you the actual item; wearing both together looks even better than trying them on virtually.

[0085] Styling suggestions: 1. Core outfit: Off-white silk high-waisted shirt + navy blue high-waisted A-line skirt (a suitable outfit for commuting and everyday work); 2. Alternative outfit: Light apricot knitted cardigan + navy blue skirt (suitable for slightly cool spring and autumn weather); 3. Accessory pairing: Delicate gold necklace + beige loafers to enhance the sense of understated luxury; Handling Objections: 1. Color Concerns: Show samples of Morandi navy blue and traditional navy blue fabric on-site for comparison, allowing customers to visually experience the difference; 2. Matching Concerns: Show actual samples of matching tops and skirts and guide customers to the fitting room to try them on; 3. Price Concerns: Remind customers that silver card members enjoy a 10% discount, and that purchasing two sets together can be combined with the store's promotion of 200 RMB off for purchases over 2000 RMB.

[0086] (III) Integration and Output of Sales Guide Strategy Data The module integrates the standardized strategies generated by the first sales guide state machine and the personalized strategies generated by the third sales guide intelligent agent, extracts core scripts, key points of explanation, matching suggestions, and objection handling solutions, and generates the final first sales guide recommendation strategy data. This data is displayed on the sales guide's 206 tablet in a visual split-screen interface. The interface is divided into four sections: "Script Recommendation, Matching Suggestions, Objection Handling, and Customer Information". Sales guides can directly refer to the scripts for communication or make flexible adjustments to achieve an organic combination of standardized and personalized services.

[0087] (iv) Data analysis and feedback of the fourth commodity analysis agent In this example, Ms. Zhang ultimately purchased a navy blue midi skirt and a beige silk blouse. This sales data and interaction feedback data were uploaded to the cloud in real time. The fourth product analysis agent (based on the XGBoost+K-Means clustering model) initiated full-dimensional data analysis. The analysis process is as follows: Data collection: Collect sales data such as sales volume of Morandi color series products in stores in the past 30 days, purchase preferences of customers of different body types, and conversion rate of commuter outfit matching, as well as customer evaluation and feedback data on color, fit and style; Data analysis: K-Means clustering was used to categorize customers into three body types: pear-shaped, hourglass-shaped, and straight-cut. Statistical analysis showed that pear-shaped customers had a 78% conversion rate for high-waisted A-line bottoms and a 35% repurchase rate for Morandi color commuter wear, which was much higher than other color schemes. Results Feedback: Standardized analysis results in CSV format are generated and pushed to the brand's ERP system's large-scale product analysis model in real time via API interface. Feedback suggestions include "increasing production and inventory of high-waisted A-line bottoms in Morandi color schemes, adding light gray matching skirts, and optimizing the combination strategy for commuter tops and bottoms." This provides data support for the brand's product development, inventory allocation, and marketing optimization, achieving a two-way closed loop between terminal sales guides and brand operations.

[0088] In this embodiment, the first product recommendation module 202 in the preferred product sales system 200 can further include a full lifecycle data access unit (not shown), a multi-dimensional adaptability scoring module (not shown), a dynamic weight adjustment unit (not shown), and an ERP reverse empowerment module (not shown), constructing a closed-loop system of "data collection - scoring modeling - recommendation generation - R&D feedback," as specifically implemented below: Full lifecycle data access unit: Collects product data from production to after-sales service through multiple interfaces, covering four core stages. The data access specifications and collection methods are shown in the table below: Data phase Core data items Data types Data collection method Update frequency Production data Fabric composition (cotton / silk / linen ratio), cutting process, pattern iteration record, quality inspection pass rate Numeric / Text Brand ERP system API interface integration Synchronization upon goods entering the warehouse Sales data Purchase conversion rate for customers of different body types, size fit rate, return reason classification (fit / fabric / color), and positive review keywords. Numeric / Enumeration Real-time database crawling of product sales system Real-time updates After-sales data Deformation rate after 3 / 6 / 12 months of wear, washing durability, repair rate, and long-term user reviews. Numeric / Text After-sales management system + user questionnaire feedback Monthly updates Supply chain data Inventory turnover rate, replenishment cycle, regional best-selling distribution, and stockout warnings. Numeric / Enumeration Supply Chain Management System API Interface Integration Daily updates After data is received, it is processed in a standardized manner: text data (such as reasons for returns and keywords for positive reviews) is converted into numerical features using the Term Frequency – Inverse Document Frequency (TF-IDF) algorithm; missing data is filled in using neighborhood interpolation; and abnormal data (such as extreme values) is removed using the 3σ principle to ensure data quality.

[0089] Multi-dimensional Adaptability Scoring Module: Constructs a four-dimensional scoring system of "body fit + quality reliability + scenario matching + supply chain health" to replace the traditional single-body matching logic. The core calculation formula is as follows: Body fit ( Based on the original patent's human-product matching rules, combined with customer body shape analysis data and product pattern parameters, cosine similarity is calculated, with a value range of 0-100 points; Quality Reliability ): Calculated based on production-side quality inspection pass rate (weight 30%), after-sales deformation rate (weight 40%), and return rate (weight 30%), using the following formula: The value ranges from 0 to 100. Scene matching degree ( ): Combining customer scenario demand tags with product scenario adaptation parameters (such as the crispness of commuter clothing and the elasticity of sportswear), the score is calculated through rule library matching, with a value range of 0-100. Supply chain health ): Calculated based on inventory turnover rate (weight 40%), replenishment cycle (weight 30%), and stockout risk (weight 30%), the formula is as follows: The value ranges from 0 to 100.

[0090] Dynamic weight adjustment unit: In conjunction with the first recommendation agent, it dynamically adjusts the weights of the four-dimensional score based on user tags, scenario requirements, and market trends. The core adjustment rules are as follows: For customers with high spending power (average order value ≥ 2000 yuan): Improve Weighting to 35%, reduced Weighted at 10%, high-quality products will be prioritized for recommendation; For customers making urgent purchases (scenario tagged "instant outfit"): Improve The weighting is set at 25% to ensure that recommended products have sufficient inventory and no stock shortages. For seasonal transition periods: Improvement With a weighting of 25%, products suitable for the current season and scenario are recommended first (e.g., breathable fabrics are recommended for summer and warm fabrics are recommended for winter). For the new product launch period: Improvement The weighting is set at 45%, highlighting the suitability of the new product for the customer's body shape.

[0091] ERP Reverse Empowerment Module: The fourth product analysis intelligent agent generates three types of core analysis reports based on four-dimensional scoring data and customer feedback, and feeds them back to the brand owner's ERP-side product analysis model through a standardized API interface. Layout optimization report: targeting For products with low ratings and a high concentration of return reasons, provide suggestions for adjusting the fit (e.g., "Pear-shaped customers have a high return rate for a certain style of pants, so we suggest increasing the pant length by 2cm and widening the hip circumference by 1cm"). Quality Improvement Report: Targeting For products with low ratings, provide suggestions for fabric / process improvement (e.g., "This shirt has low washability; it is recommended to replace it with wrinkle-resistant fabric"). Supply Chain Adjustment Report: Targeting For products with low ratings, provide inventory allocation suggestions (e.g., "A certain dress is selling well in the southern region, it is recommended to increase the replenishment quantity").

[0092] This improved solution builds a four-dimensional scoring system based on the entire product lifecycle data, breaking through the limitations of traditional methods that only focus on body fit. It achieves multi-dimensional, precise recommendations based on "body fit + quality + scenario + supply chain," addressing industry pain points such as "good fitting effect but poor long-term user experience" and "recommended products out of stock." A dynamic weight adjustment mechanism is proposed to adapt to the different needs of different customers, scenarios, and time periods, enhancing the personalization and flexibility of recommendations. Furthermore, it strengthens deep integration with brand owners' ERP systems, transforming recommendation data into decision-making basis for product development and supply chain optimization, forming a closed loop of "sales-data-R&D" and enhancing the core competitiveness of brand owners.

[0093] In this embodiment, the product sales system 200 can preferably further add a service effect data collection unit (not shown), a multi-dimensional effect evaluation model (not shown), a strategy iteration optimization unit (not shown), and a personalized strategy distribution unit (not shown) to the data generation module 204 for the shopping guide recommendation strategy. This constructs a closed-loop system encompassing "data collection - evaluation modeling - iterative optimization - precise distribution," collecting only the process and results data of the shopping guide service without involving any user privacy information. The specific implementation is as follows: Service Effectiveness Data Collection Unit: Deployed in the backend of the sales guide device 206 and the product sales system 200, it collects full-process effectiveness data on a "single sales guide service" basis, constructing a standardized dataset D = {D1, D2, D3, D4}, with each data item defined as shown in the table below: Data categories Core data items Data types Data collection method Value range / example Strategy Basic Information (D1) Strategy type, sales guide stage, trigger timing, associated product ID Enumeration / String The system automatically associates Strategy type: Order-closing / Demand-deepening; Sales guidance stage: Try-on demonstration / Objection handling Service process data (D2) Customer response time, number of interaction rounds, and completeness of strategy execution Numeric / Boolean Automatic recording and manual confirmation on the sales guide platform Response time ≤ 3 seconds; Interaction rounds: 2 rounds; Execution integrity: Yes Service outcome data (D3) Instant conversion results, associated sales, and customer satisfaction ratings Boolean / Numerical Sales system integration + sales guide rating entry Conversion result: Yes / No; Related amount: 599 yuan; Satisfaction rating: 4.8 out of 5. Environment variable data (D4) Store hours, customer traffic density, and merchandise inventory status Enumeration / Numerical The system automatically captures Time of day: Weekend afternoon; Crowd density: Medium; Inventory status: Sufficient / Scarce Data collection follows the principle of "minimum necessity," recording only objective data related to the service and satisfaction ratings marked by sales guides. It does not collect sensitive information such as customers' personal identities or privacy behaviors. Data storage is encrypted and accessible only to authorized personnel.

[0094] Multi-dimensional effect evaluation model: Employing a fusion architecture of "weighted scoring + cluster analysis," it evaluates the effectiveness of sales strategies based on collected data across three core dimensions: "conversion efficiency, interaction quality, and adaptability," generating a comprehensive strategy effectiveness score (S, 0-100 points). The specific implementation is as follows: Dimensional metrics definition: Conversion efficiency dimension (weight 40%): includes instant conversion rate (C1) and cross-selling rate (C2), the formula is S1 = 60×C1 + 40×C2 (C1 = number of orders converted per service / number of services, C2 = cross-selling amount / main product amount). Interaction quality dimension (weight 30%): includes customer response rate (C3) and average response time (C4), the formula is S2 = 50×C3 - 10×(C4 / 5) (C3 = number of services with a response / total number of services, C4 is in seconds, normalized to within 5 seconds). Adaptability dimension (weight 30%): includes satisfaction score (C5) and strategy execution completeness (C6), the formula is S3 = 80 × (C5 / 5) + 20 × C6 (C6 = 1 indicates complete execution, 0 indicates no execution); The overall score is calculated as follows: S = 0.4×S1 + 0.3×S2 + 0.3×S3, and is divided into 5 levels according to the score: Excellent (S≥90), Good (80≤S<90), Satisfactory (70≤S<80), Needs Optimization (60≤S<70), and Ineffective (S<60). Model optimization: A clustering model was trained based on 100,000 sets of historical service performance data to explore the performance patterns of different strategy types in different scenarios (time period, customer flow, sales guidance stage) and form a scenario-performance mapping relationship library for subsequent strategy iteration.

[0095] The strategy iteration and optimization unit works in conjunction with the first and third sales guide state machines to iterate strategies based on evaluation results. The core logic is as follows: Excellent strategies (S≥90): Incorporate them into the "high-quality strategy library" and extract the core elements of the strategy (such as the script structure, triggering timing, and recommendation logic) as templates for generating new strategies. For example, solidify the combination logic of "discount prompts + inventory scarcity" of the "high conversion order-driving strategy" into a template. Good / Acceptable strategy (70≤S<90): Perform local optimization, for example: Strategies for addressing insufficient interaction quality: Optimize the conciseness of the language, break long sentences into shorter ones, and reduce the customer's comprehension cost; Strategies for insufficient fit: Adjust the triggering time from "triggered immediately after trying on" to "triggered after the customer views the details"; Possible optimization / failed strategies (S < 70): Analyze the reasons for failure (such as scenario mismatch, awkward wording), pause the automatic triggering of the strategy, or reconstruct the core logic, for example, change the "failed associated recommendation strategy" from "forced accessory recommendation" to "scenario-based matching suggestions"; Iteration frequency: Batch iteration is performed weekly, and the strategy template library is updated monthly based on new data to ensure that the strategy adapts to changes in market and customer needs.

[0096] Personalized Strategy Distribution Unit: Based on the sales associate's personal service data (years of service, stage of expertise, historical conversion data), this unit enables precise distribution of strategies. The core logic is as follows: Salesperson profile building: Collect service data of salespersons and generate three types of tags: "preferred strategy type, advantageous sales stage, and suitable customer flow type". For example, "preferred strategy for deepening needs, advantageous stage: objection handling, suitable for high customer flow scenarios". Strategy matching and distribution: Based on the sales associate profile and the current service scenario (time period, customer flow, sales stage), match suitable strategies from the optimized strategy library, for example: For novice shoppers, in low-traffic scenarios, during the fitting and display stage, we recommend a "standardized, high-success-rate strategy," complete with execution steps. Experienced sales associates + high-traffic scenarios + order closing stage --> We recommend "personalized and flexible strategies" to allow sales associates to adjust their sales pitches independently; Distribution effectiveness feedback: Record the sales staff's feedback on the use of the strategy (adoption rate, modification rate), continuously optimize the distribution logic, and ensure that the strategy matches the sales staff's capabilities.

[0097] This improved solution overcomes the limitations of traditional static strategies by constructing a "full-link closed-loop" iterative system for sales guide strategies. Through continuous collection and analysis of service performance data, it achieves dynamic optimization of strategies, addressing the core pain points of "rigid strategies and poor adaptability." A multi-dimensional performance evaluation model is proposed, comprehensively assessing strategy value from three dimensions: conversion, interaction, and adaptability. This avoids the bias of relying on a single conversion metric. Simultaneously, data collection strictly protects user privacy, focusing solely on the service itself while balancing effectiveness and compliance. Personalized strategy distribution is achieved, adapting to sales guides with varying abilities and scenario needs. This solves the industry problem of "unified strategies failing to adapt to all sales guides," improving strategy execution effectiveness and sales guide acceptance.

[0098] Example 3: To more clearly reveal the operational details of the merchandise sales system involved in this invention, see [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the interface interaction for generating shopping guide recommendation strategy data on the shopping guide side of the product sales system provided in Embodiment 3 of the present invention.

[0099] Embodiment 3 of the present invention continues to use the embodiment of Ms. Zhang's commodity sales system structure from Embodiment 2, and further see [link to embodiment 2]. Figure 3The document provides an example of another preferred interface diagram for Ms. Zhang's shopping guide in Embodiment 2. The generated shopping guide recommendation strategy data, known as the icebreaker script, is generated during the customer's first offline shopping trip. It is based on the customer's tags, product recommendations, customer usage data, and the customer's browsing history on the smart shopping guide device during offline shopping, as well as preference analysis and feedback after trying on items. Different shopping guide strategy scripts are available, offering general advice, providing value, and reducing pressure, facilitating precise sales guidance for sales staff. Preferably, the further shopping guide recommendation strategy data may also include: marketing strategy data, which provides sales strategies for the sales staff to complete the sales activity. For example, it analyzes customer tags, product recommendations, customer usage data, and the customer's browsing history on the smart shopping guide device during offline shopping, as well as preference analysis and feedback after trying on items, to indicate that the customer is a rational consumer, values ​​sensory experience, and prefers practical, empirically-based trying on items, avoiding anything that might offend the customer's aesthetic sense. Preferably, when generating the shopping guide recommendation strategy data, information in the scenario is further considered: for example, bringing a pet, bringing a child, or bringing a close friend. The scenario information is further used when the third shopping guide agent generates the shopping guide recommendation strategy data.

[0100] The system device embodiments described in the above examples are merely illustrative. The units described as separate components may or may not be physically separate. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for processing shopping guide information, running in a merchandise sales system, comprising: The intelligent shopping guide device in the product sales system obtains the body data confirmed and provided by the customer, and generates the body analysis data of the customer; The first recommendation agent generates the customer's first product recommendation information based on the body analysis data and the product information accessed from the product sales system; The second effect generation agent generates the first usage effect generation data of the recommended product based on the shape analysis data and the first product recommendation information. Based on the customer's user tag data, the first product recommendation information, the first usage effect generation data, and the customer's interaction feedback information in the product sales system, the first shopping guide recommendation strategy data is generated on the shopping guide side.

2. The method according to claim 1, characterized in that, The shape data includes: The customer's body shape dimensions, body proportions and shape, facial features, or at least one of the following:

3. The method according to claim 1, characterized in that, The first recommendation agent generates the customer's first product recommendation information based on the body shape analysis data and the product information accessed from the product sales system, including: The first recommendation agent, based on the customer's body shape tags and / or style tags in the body shape analysis data, generates a first product recommendation list from the product information accessed from the product sales system through a person-product matching rule; and / or The first recommendation agent, based on a clothing formula library and product information accessed from the product sales system, uses the aesthetic agent rating results from the virtual model's try-on based on the body shape analysis data to generate a first product recommendation list; and / or The first recommendation agent combines product information accessed from the product sales system, and based on the body analysis data, after virtual models try on the products, retrieves collaborative scores from designer agents and / or stylist agents and / or fashion agents, and then generates a first product recommendation list.

4. The method according to claim 1, characterized in that, The second effect-generating agent generates the first usage effect generation data of the recommended product based on the shape analysis data and the first product recommendation information, including: Based on the body shape analysis data, a model is matched to wear the recommended product, and first image data of the model wearing the recommended product is generated. Face detection and alignment: A face detection and key point alignment model is used to detect facial regions in the first image data, and alignment is achieved by face masking type; Feature replacement: Facial feature transfer is performed using a face-swapping model to preserve the facial structure of the customer target image in the body analysis data; Detail optimization: Natural transitions in the facial region of the first image data are achieved through edge expansion and blurring, and the skin texture of the facial region in the first image data is restored by combining a face restoration model; Super-resolution enhancement: Based on the diffusion transformer model and variational autoencoder model, the output resolution of the facial region in the first image data is improved to ensure the skin color of the facial region in the first image data is consistent, and the output is generated as the first effect data.

5. The method according to claim 1, characterized in that, The second effect-generating agent generates the first usage effect generation data of the recommended product based on the shape analysis data and the first product recommendation information, including: Trigger a try-on rendering request on the smart shopping guide device side; The second effect generation intelligent agent sends a try-on command to the rendering engine on the shopping terminal side, directly calls the locally stored body analysis data, completes the first try-on rendering locally, and obtains the first basic image data; On the sales guide side, the first desensitized body feature is extracted from the body analysis data; The first desensitized body features, the first product recommendation information and the rendering requirement information are uploaded to the cloud-based detail rendering engine to perform the second try-on rendering and obtain the second detail enhancement map data. The first base map data and the second detail enhancement map data are fused to generate the first usage effect data of the recommended product.

6. The method according to claim 1, characterized in that, The step of generating first shopping guide recommendation strategy data on the shopping guide side based on the customer's user tag data, first product recommendation information, first usage effect generation data, and the customer's interaction feedback information in the product sales system includes: The first shopping guide state machine, in real time at each shopping guide stage, triggers the transition conditions of the shopping guide state machine based on the customer's user tag data, the first product recommendation information, the first usage effect generated data, and the customer's interactive feedback information at each shopping guide stage in the product sales system, and generates the corresponding first shopping guide recommendation strategy data according to the prompt word template.

7. The method according to claim 1, characterized in that, The step of generating first shopping guide recommendation strategy data on the shopping guide side based on the customer's user tag data, first product recommendation information, first usage effect generation data, and the customer's interaction feedback information in the product sales system includes: The system acquires real-time information on the status of the shopping guide stage, the customer's user tag data, the first product recommendation information, the first usage effect generation data, and the customer's interaction feedback information in the product sales system. Based on the third shopping guide intelligent agent, it outputs the shopping guide recommendation strategy data for the customer in the corresponding shopping guide stage.

8. The method according to claim 1, characterized in that, The product information accessed in the product sales system includes: product information entered into the product sales system, and product information obtained from at least one external enterprise resource planning data interface accessed by a brand owner.

9. The method according to claim 8, characterized in that, Further includes: The fourth product analysis intelligent agent analyzes the sales data of the product sales system and the customer's interaction feedback information in the product sales system, and feeds back the analysis results to the product analysis big model on the enterprise resource planning side accessed by the brand owner.

10. A product shopping guide system, comprising: The body shape data acquisition and analysis module is configured in the intelligent shopping guide device of the product sales system, and is configured to acquire body shape data provided by the customer and generate the customer's body shape analysis data. The first product recommendation module is configured to use a first recommendation agent to generate first product recommendation information for the customer based on the body analysis data and the product information accessed from the product sales system; The first effect generation module is configured to use a second effect generation agent to generate the first usage effect generation data of the recommended product based on the shape analysis data and the first product recommendation information; The shopping guide recommendation strategy data generation module is configured to generate first shopping guide recommendation strategy data for the shopping guide terminal based on the customer's user tag data, first product recommendation information, first usage effect generation data, and the customer's interaction feedback information in the product sales system.

Images