Retrieval enhancement dialogue generation system and method applied to digital human shopping guide

Abstract

The invention relates to the technical field of natural language generation, in particular to a retrieval enhancement dialogue generation system and method applied to digital human shopping guide, and the system comprises an input module, a semantic analysis module, a multi-strategy memory module, a context management module, a retrieval module, a generation module and a privacy and security control module. The method has the beneficial effects that through the multi-strategy memory module and dynamic user portrait construction, personalized demands and preferences of users can be deeply understood and captured, and accurate product recommendation and customized purchase suggestions are realized. The shopping satisfaction degree of the user is improved, and the stickiness and loyalty of the user are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language generation, and specifically provides a retrieval augmented generation dialogue system and method for digital human shopping guides. Background Art

[0002] With the rapid development of e-commerce and online retail, the digital shopping experience has become an important factor influencing users' purchase decisions. Most existing online shopping guide systems rely on simple retrieval mechanisms and rule-based Q&A models, making it difficult to meet the increasingly diverse and personalized shopping needs of users. Such systems usually rely on keyword matching or preset response templates, lacking in-depth semantic understanding of the natural language input by users, resulting in recommended results that are often lacking in pertinence and flexibility and unable to effectively handle complex user inquiries and changing shopping scenarios.

[0003] In addition, with the rapid update of product information (such as new product launches, price adjustments, inventory changes) and the dynamic evolution of user behaviors and preferences, the response speed and accuracy of traditional shopping guide systems in information update and personalized recommendation are severely restricted. This not only affects the shopping experience of users but also limits the sales conversion rate of the platform. At the same time, the privacy and security protection of user data are also a major problem in the existing technology. Many systems have deficiencies in permission control and data desensitization, unable to fully protect the personal information security of users, thus affecting user trust and willingness to use.

[0004] In recent years, technologies such as natural language processing (NLP), vector retrieval, and deep learning have made significant progress, providing a technical foundation for building more intelligent and efficient dialogue systems. However, how to deeply integrate these advanced technologies to build a dialogue engine that can dynamically understand and respond to user needs, achieve context awareness and personalized recommendation, remains an urgent challenge to be solved. Especially in the specific application scenario of digital human shopping guides, there is an urgent need for a retrieval augmented generation (RAG) dialogue engine that can integrate multi-source data, implement multi-strategy memory management, and dynamic context adjustment to provide more intelligent, efficient, and secure shopping assistance and decision support.

[0005] Therefore, developing a RAG dialogue engine based on underlying APIs, which can effectively integrate user portraits, real-time update product information, dynamically adjust retrieval strategies, and achieve highly personalized recommendations and responses on the premise of ensuring data privacy and security, is of great significance for improving the user shopping experience and promoting the business growth of e-commerce platforms. Summary of the Invention

[0006] The purpose of the present invention is to provide a retrieval augmented generation dialogue system and method for digital human shopping guides to solve the problems raised in the above background art.

[0007] To achieve the above object, the present invention provides the following technical solutions: A retrieval-enhanced generation dialogue system applied to digital human shopping guides, the system includes:

[0008] An input module, configured to receive purchase consultation requests submitted by users through multiple channels and preprocess the input data;

[0009] A semantic parsing module, connected to the input module, configured to perform word segmentation, entity recognition, and intent analysis on the preprocessed user input, and extract key semantic features;

[0010] A multi-strategy memory module, connected to the semantic parsing module, includes a short-term memory unit and a long-term memory unit, configured to store and manage the user's historical interaction information and user portrait data, and combine multiple retrieval strategies to achieve precise response to user needs;

[0011] A context management module, connected to the multi-strategy memory module, configured to maintain and optimize the dialogue context information to ensure the coherence and relevance of multi-round conversations;

[0012] A retrieval module, connected to the multi-strategy memory module and the context management module, configured to efficiently retrieve relevant content from the product knowledge base, user evaluation database, and real-time inventory information based on user queries and context information;

[0013] A generation module, connected to the retrieval module, configured to integrate the retrieval results and use a deep language model to generate personalized purchase recommendations and natural language responses;

[0014] An interpretability interface, connected to the generation module, configured to display the decision-making process and recommendation basis of the system to improve the transparency of the system;

[0015] A privacy and security control module, connected to the generation module, configured to ensure the security and privacy of user information through data desensitization, encryption, and fine-grained permission management.

[0016] Preferably, the input module further includes: a speech recognition sub-module, configured to convert the user's speech input into text; an image recognition sub-module, configured to extract key information or labels from the images uploaded by the user;

[0017] The semantic parsing module further includes: a dependency syntax analysis sub-module, configured to parse the sentence structure of the user input and understand the relationships between entities; an emotion analysis sub-module, configured to identify the user's emotional tendency to adjust the tone and content of the response.

[0018] Preferably, the long-term memory unit in the multi-strategy memory module uses a vector database for efficient vectorized storage and retrieval, and integrates knowledge graph technology to achieve fast retrieval and association analysis of complex relationships;

[0019] The context management module divides context information into short-term context, medium-term context, and long-term context through a hierarchical management mechanism to ensure context consistency in multi-turn conversations.

[0020] Preferably, the retrieval module adopts a weighted fusion mechanism to comprehensively sort the results of different retrieval strategies to generate a final recommendation list.

[0021] The generation module dynamically generates prompts applicable to the current conversation context based on the user profile and retrieval results through dynamic Prompt construction technology, guiding the deep language model to generate personalized responses.

[0022] The interpretability interface includes visualization tools for displaying retrieval sources, recommendation reasons, and strategy selection processes, supporting developers and users in understanding and optimizing the system's decision-making process.

[0023] Preferably, the privacy and security control module defines access permissions for different roles by integrating the role-based access control (RBAC) mechanism to ensure that only authorized users and modules can access specific data. The privacy and security control module uses the Transport Layer Security (TLS) protocol and the Advanced Encryption Standard to encrypt and store user data during transmission, preventing data from being illegally obtained and tampered with.

[0024] A retrieval-enhanced generation dialogue method applied to digital human shopping guides is implemented using a retrieval-enhanced generation dialogue system for digital human shopping guides. The method includes the following steps:

[0025] Data input is used to receive purchase consultation requests submitted by users through multiple channels and preprocess the input data.

[0026] Semantic parsing is used to perform word segmentation, entity recognition, and intent analysis on the preprocessed user input to extract key semantic features.

[0027] Multi-strategy memory, including a short-term memory unit and a long-term memory unit, is used to store and manage the user's historical interaction information and user profile data, and combines multiple retrieval strategies to achieve an accurate response to the user's needs.

[0028] Context management is used to maintain and optimize dialogue context information to ensure the coherence and relevance of multi-turn conversations.

[0029] Retrieval is used to efficiently retrieve relevant content from the product knowledge base, user evaluation database, and real-time inventory information based on the user's query and context information.

[0030] Generation is used to integrate retrieval results and use a deep language model to generate personalized purchase suggestions and natural language responses.

[0031] An interpretability interface, used to display the system's decision-making process and recommendation basis, enhancing system transparency;

[0032] Privacy and security controls, used to ensure the security and privacy of user information through data desensitization, encryption, and fine-grained permission management.

[0033] Preferably, the data input further includes: a speech recognition sub-module for converting the user's speech input into text; an image recognition sub-module for extracting key information or tags from the images uploaded by the user;

[0034] Semantic parsing further includes: a dependency syntactic analysis sub-module for parsing the sentence structure of the user input and understanding the relationships between entities; a sentiment analysis sub-module for identifying the user's sentiment tendency to adjust the tone and content of the response.

[0035] Preferably, the long-term memory unit in the multi-strategy memory uses a vector database for efficient vectorized storage and retrieval, and integrates knowledge graph technology to achieve fast retrieval and association analysis of complex relationships;

[0036] Context management divides context information into short-term context, medium-term context, and long-term context through a hierarchical management mechanism to ensure context consistency in multi-turn conversations.

[0037] Preferably, the retrieval adopts a weighted fusion mechanism to comprehensively sort the results of different retrieval strategies to generate the final recommendation list;

[0038] Generation uses dynamic Prompt construction technology to dynamically generate prompts applicable to the current conversation context based on the user profile and retrieval results, guiding the deep language model to generate personalized responses;

[0039] The interpretability interface includes visualization tools for displaying retrieval sources, recommendation reasons, and strategy selection processes, supporting developers and users in understanding and optimizing the system's decision-making process.

[0040] Preferably, privacy and security controls integrate the role-based access control RBAC mechanism to define the access permissions of different roles, ensuring that only authorized users and modules can access specific data; the privacy and security control module uses the Transport Layer Security TLS protocol and the Advanced Encryption Standard to encrypt and store and transmit user data to prevent data from being illegally obtained and tampered with.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] The retrieval-enhanced generation dialogue system and method proposed by the present invention for digital human shopping guides, through the multi-strategy memory module and dynamic user profile construction, can deeply understand and capture users' personalized needs and preferences, and achieve accurate product recommendations and customized purchase suggestions. This not only improves users' shopping satisfaction, but also enhances users' stickiness and loyalty.

[0043] Combined with various retrieval strategies such as keyword retrieval, vector similarity retrieval, and knowledge graph query, the system can quickly locate the most relevant products in the vast amount of product information, ensuring the accuracy and timeliness of the recommendation results. This effectively solves the deficiencies of traditional shopping guide systems in information matching and recommendation accuracy.

[0044] Through hierarchical context management and dynamic summarization technology, the system can maintain the consistency and coherence of the context in multi-round conversations, and improve the ability to understand complex user needs. This enables the dialogue engine to still provide accurate and relevant responses when dealing with changing shopping scenarios and complex user inquiries.

[0045] The interpretability interface visualizes the decision-making process and recommendation basis of the system, enabling users and developers to clearly understand the working principle and recommendation logic of the system, increasing the transparency of the system and users' trust. At the same time, the integration of the user feedback mechanism enables the system to continuously optimize and improve the recommendation strategy according to users' actual feedback.

[0046] The privacy and security control module ensures the security and privacy of users' personal information and sensitive data during storage and transmission through data de-sensitization, encryption, and fine-grained permission management, meeting the requirements of international data privacy regulations such as GDPR and CCPA. This not only protects users' data security, but also enhances users' trust and willingness to use the system.

[0047] The modular system design enables each core module to be independently developed, optimized, and extended, facilitating flexible adjustment and upgrade according to different business requirements and technological developments. The support of the plug-in mechanism and open API interfaces enables the system to easily integrate new data sources and functional modules, maintaining the forward-looking and adaptability of the system.

[0048] Through accurate personalized recommendations and efficient information retrieval, the present invention can significantly improve users' purchase conversion rate and the platform's sales performance. At the same time, the improvement of intelligence and security enables the e-commerce platform to have stronger competitiveness and user attraction in the fierce market competition. Brief Description of the Drawings

[0049] Figure 1 It is the system block diagram of the present invention. Detailed Embodiment

[0050] In order to clearly and completely describe the objectives, technical solutions of the present invention and make its advantages more clearly understood, the following further elaborates on the embodiments of the present invention in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present invention, rather than all of them, and are only used to explain the embodiments of the present invention, not to limit the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0051] Embodiment 1. Please refer to Figure 1 , the present invention provides a technical solution: a retrieval-enhanced generation dialogue system applied to digital human shopping guides. The system includes:

[0052] 1) Input module

[0053] The input module is the external interaction entry of the system, used to receive input information from users in multiple channels and modalities, and preprocess it to provide a high-quality data basis for subsequent semantic parsing and information retrieval. This module can be compatible with various input forms such as text, voice, and images, providing support for the complex and diverse user interaction needs in the digital human shopping guide scenario. Through the API gateway, the system can be seamlessly docked with Web platforms, mobile applications, smart speakers, and AR / VR devices to achieve a cross-platform and cross-device shopping guide interaction experience.

[0054] In terms of text input, the input module will first perform standardization and cleaning operations on the original text data, such as removing illegal characters, punctuation normalization, and stop word filtering. For voice input, by integrating an online or offline speech recognition engine (such as Wit.ai or Vosk), the audio data is converted into text in real time, and appropriate adaptive tuning is performed according to background noise, speech rate, and accent to improve the recognition accuracy. For the product pictures or QR code information uploaded by users, this module will call an image recognition model (such as a CNN-based object detection model) to extract key information therein to assist subsequent product retrieval and recommendation.

[0055] In terms of input verification and security, the input module has built-in basic anomaly detection and filtering mechanisms. When malicious code injection, input containing sensitive words or bad information is detected, the input will be intercepted or marked to ensure the overall security and compliance of the system. In addition, this module also provides necessary input data access logs and metadata for the privacy and security control module, enabling fine-grained control in subsequent data desensitization and permission management. Through the comprehensive application of the above technical means, the input module not only provides a high-quality and diverse data input source for the system, but also lays a solid foundation for the efficient and secure operation of subsequent modules.

[0056] 2) Semantic parsing module

[0057] The semantic parsing module is a crucial natural language processing component in the system. It is responsible for deeply understanding and analyzing the information input by users, extracting key entities, intents, and semantic features to support the subsequent information retrieval and generation processes. This module adopts advanced natural language processing (NLP) technologies, including word segmentation, part-of-speech tagging, entity recognition, intent recognition, and dependency parsing, to ensure the accurate capture and understanding of user needs. By integrating multiple pre-trained language models (such as BERT, RoBERTa), the semantic parsing module can effectively process complex natural language inputs and identify the specific purchase intents and preferences of users.

[0058] In terms of entity recognition (Named Entity Recognition, NER), the semantic parsing module uses deep learning-based entity recognition algorithms to automatically identify key entities in the user's query, such as product names, brands, price ranges, functional requirements, etc. For example, when the user asks "Recommend a lightweight laptop suitable for students", the module can identify "students" as the user group, "lightweight" as the product feature, and "laptop" as the product category, etc. as key entities. At the same time, the intent recognition mechanism determines the purpose of the user's query through classification algorithms (such as Transformer-based classifiers), such as seeking recommendations, comparing products, asking about inventory, etc., thereby providing a basis for the subsequent selection of retrieval strategies.

[0059] In addition, the semantic parsing module also integrates dependency parsing and sentiment analysis functions to further enhance the depth of understanding of user inputs. Dependency parsing helps the module analyze the grammatical structure of sentences, understand the relationships and context dependencies between entities, and ensure the accurate parsing of complex sentences by the system. For example, distinguish the different focuses in "Recommend a lightweight laptop suitable for students" and "Recommend a lightweight and suitable laptop for students". Sentiment analysis is used to identify the user's sentiment tendency, such as satisfaction, anxiety, confusion, etc., to help the system adjust the tone and content of the response, and improve the personalization and humanization of the user experience. In addition, the module also has adaptive learning capabilities, and can continuously optimize and improve the accuracy and efficiency of semantic parsing through continuous user interaction and feedback, ensuring that the system maintains a high level of understanding ability in the face of changing user needs and language expressions.

[0060] 3) Multi-strategy Memory Module

[0061] The multi-strategy memory module is the core component in the system responsible for storing, managing, and invoking the user's historical interaction information and user profile. By integrating short-term memory (STM) and long-term memory (LTM) units and combining multiple retrieval strategies, this module achieves a deep understanding of user needs and precise responses. The multi-strategy memory module can not only efficiently manage the user's real-time interaction data but also dynamically select the most suitable retrieval method according to different retrieval requirements, enhancing the personalization and relevance of the recommendation system.

[0062] 1. Short-Term Memory Unit (STM)

[0063] The short-term memory unit (STM) is mainly used to store the latest user needs and temporary topic information in the current session. STM adopts a circular buffer structure, which can efficiently manage the most recent conversation turns within a fixed storage capacity, ensuring that the system can quickly access and update the latest user interaction data. Whenever a user makes a new interaction, STM will automatically overwrite the earliest conversation record to maintain the efficiency of memory usage. In addition, STM also integrates a caching mechanism to pre-cache frequently accessed key information, accelerating subsequent retrieval processes and ensuring that the system always maintains a high response speed in multi-turn conversations.

[0064] 2. Long-Term Memory Unit (LTM)

[0065] The long-term memory unit (LTM) is used to store the user's long-term preferences, historical purchase records, and other persistent user profile information. LTM uses a vector database (such as FAISS) for efficient vectorized storage and retrieval, supporting deep semantic matching based on vector similarity. In addition, LTM also integrates knowledge graph technology to structurally store information such as the user's behavior data, preference tags, and product attributes, forming a highly associated knowledge network. Through knowledge graph queries, the system can achieve rapid retrieval and correlation analysis of complex relationships, further enhancing the accuracy and relevance of recommendations. To maintain the timeliness and accuracy of LTM data, the module designs a regular update mechanism to automatically clean up outdated information and integrate the latest user behavior data, ensuring that the long-term memory library always reflects the user's latest preferences and demand changes.

[0066] 3. Multi-Strategy Invocation Mechanism

[0067] The multi-strategy invocation mechanism is the core function of the multi-strategy memory module, responsible for dynamically selecting the most suitable retrieval strategy according to the current user query and context information. This mechanism supports the integration of multiple retrieval methods, including keyword retrieval, vector similarity retrieval, and knowledge graph queries. To achieve this goal, the system introduces a strategy scheduler (StrategyScheduler), and its workflow is as follows:

[0068] 4. Strategy Fitness Scoring

[0069] The policy scheduler first calculates a fitness score for each retrieval strategy based on the characteristics of the user's current query (such as the keywords included, semantic vector representation, user profile features, and context information). The scoring takes into account the performance of each strategy in aspects such as keyword matching, vector similarity, and knowledge graph retrieval, as well as context information such as the user's historical preferences, geographical location, and time factors. These scores reflect the applicability and effectiveness of each retrieval strategy in the current situation.

[0070] 5. Policy Fusion and Result Ranking

[0071] Based on the fitness scores of each strategy, the policy scheduler determines the priority of each strategy and performs weighted fusion on the retrieval results of different strategies. The system comprehensively considers the scores of each strategy and integrates the results of different strategies into a unified recommendation list through a weighting mechanism. This process ensures that the retrieval results cover a wide range and can accurately match the user's personalized needs.

[0072] 6. Adaptive Learning and Optimization

[0073] To further improve the effectiveness of the multi-strategy invocation mechanism, the system introduces an adaptive learning and optimization algorithm. By continuously monitoring the user's interaction feedback and behavior data, the system dynamically adjusts the weight parameters of each retrieval strategy to adapt to user needs and market changes. The adaptive learning mechanism enables the multi-strategy invocation mechanism to continuously optimize strategy selection and improve the accuracy of retrieval results and user satisfaction.

[0074] 4) Context Management Module

[0075] The context management module is a key component in the system responsible for maintaining and optimizing the dialogue context information, aiming to ensure the coherence and relevance during the multi-round dialogue process. This module effectively processes and organizes the user's dialogue history and current interaction content through hierarchical management and dynamic summarization techniques, thus supporting the system to provide consistent and accurate responses in complex and changing shopping scenarios.

[0076] First, the context management module adopts a hierarchical structure to divide the dialogue context information into three levels: short-term context, medium-term context, and long-term context. The short-term context mainly covers the latest user needs and the content of the last few rounds of dialogue in the current session, ensuring that the system can immediately respond to the user's immediate needs. The medium-term context covers multiple historical rounds in this dialogue, extracting key information and theme changes to help the system understand the user's continuous intentions and the evolution of needs. The long-term context stores the user profile and historical behavior data across sessions, including the user's long-term preferences, purchase records, and historical interaction information, supporting the system to maintain consistent personalized services between different sessions.

[0077] Secondly, to improve the efficiency and accuracy of context management, the module introduces dynamic summarization and information compression techniques. When there are many dialogue turns, the system automatically summarizes and compresses redundant dialogue content and repetitive requirements, extracts high-level semantic information and keyword tags, and stores them in long-term memory. This not only reduces the occupation of storage space but also improves the retrieval and response speed. In addition, the context management module integrates a retrieval filtering mechanism based on metadata to refine the filtering and sorting of retrieval results through metadata such as timestamps, topic tags, and user preferences, ensuring that the system can prioritize matching the most relevant and user-demand-compliant content.

[0078] Finally, the context management module has the ability to migrate context and integrate knowledge, and can flexibly migrate relevant context information between different dialogue scenarios and domains to ensure the adaptability and consistency of the system in multi-domain applications. Through close cooperation with the multi-strategy memory module, the context management module can not only effectively organize and manage dialogue information but also dynamically adjust the context processing strategy according to the user's real-time behavior and feedback, continuously optimizing the coherence and relevance of the dialogue.

[0079] 5) Retrieval Module

[0080] The retrieval module is the core component of the system responsible for efficiently obtaining relevant information from multi-source data, aiming to provide accurate product recommendations and relevant content support based on the user's purchase consultation and context information. This module integrates a variety of advanced retrieval techniques, including keyword retrieval, vector similarity retrieval, and knowledge graph query, to ensure that it can comprehensively cover the user's needs and provide highly relevant retrieval results. Through close integration with the product knowledge base, user evaluation database, and real-time inventory information, the retrieval module can respond to user queries in real time and provide the latest product information and dynamic data support.

[0081] In terms of keyword retrieval, the retrieval module uses a retrieval engine based on inverted index (such as ElasticSearch) to quickly match the keywords in the user input with text data such as product descriptions and specifications. This method is suitable for handling clear query requirements, such as when a user directly searches for a product of a specific brand or model. In addition, the system also supports Boolean logic queries and fuzzy matching to handle spelling mistakes or fuzzy requirements in the user input, improving the fault tolerance and flexibility of the retrieval.

[0082] To achieve deeper semantic understanding and matching, vector similarity retrieval technology is widely used. By converting user queries and product information into high-dimensional vector representations (such as using Sentence-BERT or other pre-trained language models), the system can calculate the similarity between the query vector and the product vector, thus discovering semantically relevant products. This method is particularly suitable for processing complex natural language queries and implicit user needs, capable of capturing semantic associations that are difficult to identify by keyword retrieval, and improving the accuracy and relevance of recommendations.

[0083] In addition, knowledge graph query, as a structured retrieval method, is used to understand and mine complex relationships and context information in user queries. By constructing a knowledge graph that includes product attributes, user behaviors, and preferences, the system can execute complex graph queries to discover deep associations between user needs and product features. For example, when a user asks for "a waterproof watch suitable for outdoor activities", the knowledge graph query can combine the relationships between "outdoor activities", "waterproof", and "watch" to accurately recommend products that meet these features.

[0084] To ensure the timeliness and accuracy of retrieval results, the retrieval module also integrates a real-time data synchronization mechanism, which interfaces with the inventory and price systems of e-commerce platforms through APIs to ensure that the product information, inventory status, and price dynamics obtained are up-to-date. In addition, the system adopts a caching mechanism and index optimization technology to cache and pre-index frequently queried and popular products, significantly improving the retrieval speed and response efficiency, and ensuring that users can still obtain fast and reliable recommendation results during high-concurrency access.

[0085] 6) Generation Module

[0086] The generation module is the core component of the system responsible for integrating the retrieved information and converting it into a natural language response. This module uses advanced deep language models (such as the GPT series based on the Transformer architecture) for text generation, ensuring that the generated answers are not only grammatically correct, fluent and natural, but also highly conform to the user's personalized needs and context. The design of the generation module aims to provide accurate, relevant and attractive purchase suggestions through intelligent content integration and customized response strategies, significantly enhancing the user's shopping experience and satisfaction.

[0087] First, the generation module uses dynamic Prompt construction technology to dynamically generate a prompt suitable for the current conversation context based on the user's historical interaction data and real-time retrieval results. This process includes organically embedding the user's historical preferences, current query intent, and retrieved product information into the Prompt to guide the deep language model to generate responses that meet the user's needs. For example, when the user asks "Recommend a waterproof watch suitable for outdoor activities", the generation module will integrate key features such as "outdoor activities", "waterproof", and "watch" into the Prompt to guide the model to generate detailed and targeted recommendation content.

[0088] Secondly, the generation module adopts a content fusion and optimization strategy to organically integrate the results from different retrieval strategies (keyword retrieval, vector similarity retrieval, knowledge graph query). Through the fusion of multi-source information, the generation module can comprehensively consider the multi-dimensional attributes of products (such as price, brand, function), the user's personalized preferences, and real-time promotion information to generate comprehensive and personalized recommendation texts. In addition, the module also introduces a controllable generation mechanism to control the creativity and accuracy of the generated content by adjusting the parameters of the generation model (such as temperature, generation length, etc.), ensuring that the response is both innovative and accurate.

[0089] Finally, the generation module integrates multi-turn conversation support and sentiment regulation functions. In multi-turn conversation scenarios, the generation module can maintain the coherence and consistency of the conversation, ensuring that each response is related to the previous text, forming a natural and smooth conversation experience. At the same time, the module uses sentiment analysis technology to identify the user's sentiment tendency (such as excitement, confusion, satisfaction, etc.) and adjusts the tone and content of the response according to the sentiment state. For example, when the user expresses dissatisfaction with a certain product, the generation module will generate a more soothing and understanding answer to improve the user's satisfaction and trust.

[0090] 7) Interpretability Interface

[0091] The interpretability interface is an important component in the system for enhancing transparency and user trust, responsible for displaying the decision-making path and relevant basis relied on by the system during the recommendation process. This interface not only provides developers with a visualization tool for the internal workings of the system, facilitating debugging and optimization, but also provides end-users with clear recommendation reasons and reference information, enhancing the user's understanding and acceptance of the system's recommendation results.

[0092] First, the interpretability interface, through the decision-making process recording function, details the retrieval strategies used, the memory units called, and the input and output of the generation module during each recommendation process. These records include, but are not limited to, the keyword matching results of user queries, vector similarity scores, associated information of knowledge graph queries, etc. Through structured data logs, developers can comprehensively understand the recommendation logic and decision-making basis of the system, facilitating the identification and resolution of potential problems and optimizing recommendation algorithms and strategies.

[0093] Secondly, the interface provides a visualization tool to graphically present each link in the recommendation process. For example, next to the recommended products received by the user in the digital human shopping guide, it will show on what keywords, similarity scores, or knowledge graph relationships the product is recommended. This intuitive display method helps users understand the logic behind the recommendation, improving the transparency and credibility of the system. At the same time, developers can monitor the running state of the system through the visualization tool, analyze the effects of different retrieval strategies, and make data-driven optimization adjustments.

[0094] In addition, the interpretability interface also integrates a user feedback mechanism, allowing users to rate, comment on, or make suggestions for the recommendation results. By collecting and analyzing this feedback information, the system can not only adjust and optimize the recommendation strategy in real time but also further refine the interpretation content according to the specific needs and preferences of users. For example, when a user gives a high rating to a recommended product, the system can give priority to showing the recommendation basis related to that product; when a user is not satisfied with a certain recommendation, the system can provide corresponding improvement suggestions or adjust the recommendation strategy.

[0095] 8) Privacy and Security Control Module

[0096] The privacy and security control module is an important part of the system responsible for ensuring the security and privacy of user data, ensuring the confidentiality, integrity, and availability of all user information during storage, transmission, and processing. Through multi-level security measures and strict permission management mechanisms, this module complies with the requirements of international data privacy regulations such as GDPR and CCPA, protecting users' personal information from unauthorized access or leakage.

[0097] First, the privacy and security control module uses data desensitization and anonymization techniques to process users' sensitive information, ensuring that personal privacy is not exposed during data storage and transmission. For example, sensitive fields such as users' names, addresses, and contact information are encrypted and desensitized before storage to generate irreversible anonymous data, preventing abuse after data leakage. In addition, the system also introduces Differential Privacy technology, adding noise during data analysis and model training to further enhance data privacy protection, ensuring that users' personal information remains unexposed even if the data is accessed or leaked.

[0098] Second, the module implements multi-level access control and permission management to ensure that only authorized users and system components can access specific data and functions. Through the Role-Based Access Control (RBAC) mechanism, the system assigns different permissions to different roles, strictly restricting the access scope to sensitive data. For example, ordinary users can only access their own purchase records and personal preference data, while administrators and developers have higher permissions for system maintenance and optimization. At the same time, the module also supports fine-grained permission settings, allowing dynamic adjustment of access permissions according to the sensitivity of the data and the usage scenario, further enhancing the security of the system.

[0099] In addition, the privacy and security control module integrates data encryption and secure transmission protocols, using advanced encryption algorithms (such as AES-256) to encrypt users' data stored in the database to ensure the security of data in a static state. During data transmission, the system uses the Transport Layer Security (TLS) protocol to ensure that data communication between users and the server is not eavesdropped on or tampered with. To address potential security threats, the module also deploys an Intrusion Detection System (IDS) and a firewall to monitor and defend against malicious attacks in real-time, further strengthening the overall security protection ability of the system.

[0100] Finally, the privacy and security control module has comprehensive auditing and monitoring functions, recording all data access and operation logs, supporting security audits and issue tracking. By regularly reviewing and analyzing these logs, the system can promptly detect and respond to abnormal behaviors or security vulnerabilities, ensuring the effective implementation of data security policies. At the same time, the module also provides a user data management interface, allowing users to independently manage their personal data, including viewing, modifying, and deleting personal information, enhancing users' control over data privacy and trust.

[0101] Embodiment 2, based on Embodiment 1, proposes a retrieval-enhanced generation dialogue method applied to digital human shopping guides, which is implemented using a retrieval-enhanced generation dialogue system applied to digital human shopping guides. The method includes the following steps:

[0102] Data Input, which is used to receive purchase consultation requests submitted by users through various channels and preprocess the input data; further including: a speech recognition sub-module, which is used to convert the user's speech input into text; an image recognition sub-module, which is used to extract key information or labels from the images uploaded by users.

[0103] Semantic Parsing, which is used to perform word segmentation, entity recognition, and intent analysis on the preprocessed user input, and extract key semantic features; further including: a dependency syntax analysis sub-module, which is used to parse the sentence structure of the user input and understand the relationships between entities; an emotion analysis sub-module, which is used to identify the user's emotional tendency to adjust the tone and content of the response.

[0104] Multi-Strategy Memory, including a short-term memory unit and a long-term memory unit, which are used to store and manage the user's historical interaction information and user profile data, and combine multiple retrieval strategies to achieve an accurate response to the user's needs; the long-term memory unit in the multi-strategy memory uses a vector database for efficient vectorized storage and retrieval, and integrates knowledge graph technology to achieve fast retrieval and correlation analysis of complex relationships.

[0105] Context Management, which is used to maintain and optimize the dialogue context information to ensure the coherence and relevance of multi-turn conversations; through a hierarchical management mechanism, the context information is divided into short-term context, medium-term context, and long-term context to ensure context consistency in multi-turn conversations.

[0106] Retrieval, which is used to efficiently retrieve relevant content from the product knowledge base, user evaluation database, and real-time inventory information based on the user's query and context information; adopts a weighted fusion mechanism to comprehensively sort the results of different retrieval strategies to generate a final recommendation list.

[0107] Generation, which is used to integrate the retrieval results and use a deep language model to generate personalized purchase suggestions and natural language responses; through dynamic Prompt construction technology, according to the user profile and retrieval results, dynamically generate prompts applicable to the current dialogue context to guide the deep language model to generate personalized responses.

[0108] Interpretability Interface, which is used to display the system's decision-making process and recommendation basis to improve system transparency; includes visualization tools, which are used to display the retrieval sources, recommendation reasons, and strategy selection processes, and support developers and users to understand and optimize the system's decision-making process.

[0109] Privacy and security controls are used to ensure the security and privacy of user information through data masking, encryption, and fine-grained permission management; by integrating the role-based access control (RBAC) mechanism, the access permissions of different roles are defined to ensure that only authorized users and modules can access specific data; the privacy and security control module uses the Transport Layer Security (TLS) protocol and the Advanced Encryption Standard to encrypt and store user data during transmission, preventing the data from being illegally obtained and tampered with.

[0110] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A search-enhanced generation dialogue system for digital human shopping guide, characterized by: The system comprises: An input module is used to receive purchase consultation requests submitted by users through various channels and pre-process the input data; The semantic parsing module is connected to the input module and is used to perform word segmentation, entity recognition and intent analysis on the pre-processed user input to extract key semantic features; A multi-strategy memory module, connected to the semantic parsing module, including a short-term memory unit and a long-term memory unit, for storing and managing historical interaction information and user portrait data of users, and combining multiple retrieval strategies to achieve accurate response to user needs; The context management module is connected to the multi-strategy memory module and is used to maintain and optimize the conversation context information to ensure the coherence and relevance of multiple rounds of conversations; A retrieval module, connected to the multi-strategy memory module and the context management module, for efficiently retrieving relevant content from the product knowledge base, user evaluation database, and real-time inventory information based on user queries and context information; The generation module is connected to the retrieval module and is used to integrate the retrieval results and generate personalized purchase recommendations and natural language responses using the deep language model; The explainability interface is connected to the generation module to display the system's decision-making process and recommendation basis, thereby improving system transparency; The privacy and security control module is connected to the generation module and is used to ensure the security and privacy of user information through data desensitization, encryption and fine-grained permission management.

2. According to claim 1, a search enhancement generation dialogue system for digital human shopping guide, characterized in that: The input module further includes: a speech recognition submodule for converting the user's speech input into text; an image recognition submodule for extracting key information or tags from the image uploaded by the user; The semantic parsing module further includes: a dependency syntactic analysis submodule for parsing the sentence structure of the user input and understanding the relationship between entities; and a sentiment analysis submodule for identifying the user's emotional tendencies to adjust the tone and content of the response.

3. The search enhancement generation dialogue system for digital human shopping guide according to claim 1 is characterized by: The long-term memory unit in the multi-strategy memory module uses a vector database for efficient vectorized storage and retrieval, and integrates knowledge graph technology to achieve fast retrieval and association analysis of complex relationships; The context management module divides context information into short-term context, medium-term context and long-term context through a hierarchical management mechanism to ensure context consistency in multiple rounds of dialogue.

4. The search enhancement generation dialogue system for digital human shopping guide according to claim 1, characterized in that: The retrieval module adopts a weighted fusion mechanism to comprehensively sort the results of different retrieval strategies to generate the final recommendation list; The generation module uses dynamic prompt construction technology to dynamically generate prompts suitable for the current conversation context based on user portraits and search results, guiding the deep language model to generate personalized responses; The explainable interface includes visualization tools to display retrieval sources, recommendation reasons, and strategy selection process, supporting developers and users to understand and optimize the system decision-making process.

5. The search enhancement generation dialogue system for digital human shopping guide according to claim 1 is characterized by: The privacy and security control module integrates the role-based access control (RBAC) mechanism to define access rights for different roles, ensuring that only authorized users and modules access specific data. The privacy and security control module uses the transport layer security (TLS) protocol and advanced encryption standards to encrypt, store and transmit user data to prevent data from being illegally acquired and tampered with.

6. A search-enhanced dialogue generation method for digital human shopping guide, implemented by using a search-enhanced dialogue generation system for digital human shopping guide as described in any one of claims 1 to 5, characterized in that: The method comprises the following steps: Data input, used to receive purchase consultation requests submitted by users through various channels and pre-process the input data; Semantic parsing, which is used to perform word segmentation, entity recognition, and intent analysis on preprocessed user input to extract key semantic features; Multi-strategy memory, including short-term memory units and long-term memory units, is used to store and manage users' historical interaction information and user profile data, and combines multiple retrieval strategies to achieve accurate response to user needs; Context management, which is used to maintain and optimize conversation context information to ensure the coherence and relevance of multiple rounds of conversations; Retrieval, which is used to efficiently retrieve relevant content from product knowledge bases, user evaluation databases, and real-time inventory information based on user queries and context information; Generation, which is used to integrate search results and generate personalized purchase recommendations and natural language responses using deep language models; Explainable interface, used to display the system's decision-making process and recommendation basis, improving system transparency; Privacy and security controls are used to protect the security and privacy of user information through data desensitization, encryption, and fine-grained permission management.

7. The method for generating dialogues for search enhancement and digital human shopping guide according to claim 6, characterized in that: The data input further includes: a speech recognition submodule for converting the user's speech input into text; an image recognition submodule for extracting key information or tags from the images uploaded by the user; Semantic parsing further includes: a dependency syntactic analysis submodule, which is used to parse the sentence structure of the user input and understand the relationship between entities; a sentiment analysis submodule, which is used to identify the user's emotional tendencies in order to adjust the tone and content of the response.

8. The method for generating dialogues for search enhancement and digital human shopping guide according to claim 6, characterized in that: The long-term memory unit in the multi-strategy memory uses a vector database for efficient vectorized storage and retrieval, and integrates knowledge graph technology to achieve fast retrieval and association analysis of complex relationships; Context management uses a hierarchical management mechanism to divide context information into short-term context, medium-term context, and long-term context to ensure context consistency in multiple rounds of conversations.

9. The method for generating dialogues for search enhancement and digital human shopping guide according to claim 6, characterized in that: The retrieval adopts a weighted fusion mechanism to comprehensively sort the results of different retrieval strategies to generate the final recommendation list; Generate prompts that are suitable for the current conversation context dynamically based on user profiles and search results through dynamic prompt building technology, guiding the deep language model to generate personalized responses; The explainable interface includes visualization tools to display retrieval sources, recommendation reasons, and strategy selection process, supporting developers and users to understand and optimize the system decision-making process.

10. The method for generating dialogues for search enhancement applied to digital human shopping guide according to claim 6, characterized in that: Privacy and security control integrates the role-based access control (RBAC) mechanism to define access rights for different roles, ensuring that only authorized users and modules access specific data. The privacy and security control module uses the transport layer security (TLS) protocol and advanced encryption standards to encrypt, store and transmit user data to prevent data from being illegally acquired and tampered with.

Images