Conversational shopping guide method and system based on large model, and storage medium

Abstract

The invention discloses a dialogue type shopping guide method and system based on a large model and a storage medium, and belongs to the technical field of artificial intelligence and electronic commerce. According to the method, structured query and unstructured retrieval are executed in parallel on received user query, on one hand, the query is converted into SQL statement query commodity structured data tables by utilizing a natural language to SQL model, and on the other hand, the query is converted into semantic vector retrieval commodity unstructured document knowledge bases through a text representation model; integrating the two paths of retrieval results and query statements to construct cue words containing conflict processing rules, and inputting the cue words into a large language model subjected to field fine tuning to generate structured text output; finally, recommended commodity information is extracted through a preset analysis rule, and rich text recommendation content and interactive marketing cards are generated. According to the invention, high-accuracy and high-interactivity intelligent shopping guide service is realized, and the system is especially suitable for purchasing scenes of high-unit-price commodities such as automobiles, household appliances and the like.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and e-commerce technology, and in particular to a conversational shopping guide method, system and storage medium based on a large model. Background Technology

[0002] With the rapid development of large language model technology, intelligent shopping guide systems based on natural language interaction are gradually replacing traditional human customer service and search-based recommendations. However, existing technical solutions still have significant limitations.

[0003] The first type is the pure large-model generation solution. This type of solution relies entirely on the model's parameterized knowledge for recommendations and cannot connect to a real product database. This often leads to illusion problems such as products not existing or incorrect attribute information in the recommendation results. At the same time, this solution cannot handle users' hard constraints on precise attributes such as price range and inventory status, making it difficult to establish user trust.

[0004] The second category is retrieval-enhanced generation solutions. While these solutions can retrieve relevant information from document knowledge bases, they are limited to unstructured text processing and cannot utilize structured attribute fields in product databases. When users request "price below a certain threshold" or "specific brand," the system cannot accurately filter the data, resulting in low accuracy in responding to complex requests.

[0005] The third type is structured query solutions, which use natural language to SQL technology to query relational databases. While these solutions can achieve precise attribute matching, they rely on a large number of manually constructed features and rules, resulting in high implementation and maintenance costs. More importantly, they cannot understand users' implicit needs and contextual descriptions, and struggle to capture the semantic relationships between natural language expressions such as "fuel-efficient" and "silent" and specific parameters.

[0006] The common drawback of the aforementioned solutions lies in their reliance on a single-path retrieval mechanism, failing to simultaneously achieve precise numerical constraints and deep semantic understanding within the same architecture. When users present complex needs with multiple constraints, the system often struggles to address all aspects, resulting in recommendation recall and conversion rates far lower than those of human guides. Furthermore, existing technologies primarily output plain text or simple links, lacking visual guidance and multimodal interaction capabilities, making it difficult to effectively guide users in making purchase decisions. Therefore, there is an urgent need for a conversational shopping guide technology solution that can integrate multi-source knowledge, possess conflict resolution capabilities, and support rich interactive formats. Summary of the Invention

[0007] The purpose of this invention is to provide a dialogic shopping guide method, system, and storage medium based on a large model, thereby solving the aforementioned problems existing in the prior art.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A conversational shopping guide method based on a large model includes the following steps:

[0010] Receive natural language query statements input by the user;

[0011] Parallel execution of structured and unstructured queries on query statements: The query statement is converted into an SQL query statement and executed through a natural language to SQL model to obtain structured query results containing product attributes; at the same time, the query statement is converted into a semantic vector through a text representation model to retrieve the semantic index of the product unstructured document knowledge base and obtain semantically related unstructured text fragments.

[0012] The query statement, structured query results and unstructured text fragments are integrated according to a preset template to construct prompt words. The prompt words contain conflict handling rules, which are used to remove duplicate product information and align fields in the two results.

[0013] Input the prompt words into a large language model fine-tuned with domain data to generate structured text output containing recommended product identifiers and reasons for recommendation;

[0014] The structured text output is parsed using preset parsing rules to extract recommended product information and generate rich text recommendation content as well as interactive marketing cards. The marketing cards contain hyperlinks that redirect to the product details page.

[0015] Furthermore, parallel execution of structured queries and unstructured retrieval specifically includes:

[0016] The natural language to SQL model is trained using a low-rank adaptive fine-tuning method based on historical query-SQL labeled corpus pairs, and the SQL execution accuracy is used as the evaluation metric.

[0017] The text representation model employs a dual-tower dense vector encoder, and the semantic index uses the FAISS library to construct an HNSW index and returns the top K text segments with the highest cosine similarity.

[0018] Furthermore, the conflict resolution rules include:

[0019] When structured query results and unstructured text fragments contain the same product, the product attribute data in the structured query results shall be used as the authoritative data.

[0020] When the field values ​​of the same product are inconsistent in the two results, the structured data shall prevail, and the differences in the unstructured text shall be added to the prompt words as supplementary descriptions.

[0021] Furthermore, the preset parsing rules extract product identifiers and recommendation reason texts using regular expressions; the interactive marketing cards include main product cards and related marketing cards, where the main product card contains the main product image, name, price, and first hyperlink, and the related marketing cards contain accessory or service information and second hyperlink.

[0022] Furthermore, the structured product data table is a car model data table, and the unstructured product document knowledge base is a car model description document library. The method is applied to the car purchase scenario, and the marketing cards display specific car models that meet the user's configuration needs and their associated financial and insurance service options.

[0023] A large-model-based conversational shopping guide system includes:

[0024] The query receiving module is used to receive natural language query statements input by the user;

[0025] The dual-path parallel retrieval module includes a natural language to SQL model unit and an SQL execution unit, a text representation model unit and a semantic index query unit, which are used to execute structured queries and unstructured retrievals in parallel.

[0026] The information fusion module is used to integrate query statements, structured query results, and unstructured text fragments according to preset templates to construct prompt words and execute conflict handling rules;

[0027] The large language model inference module is used to process prompt words and generate structured text output based on a large language model fine-tuned by domain data;

[0028] The output parsing and display module is used to extract recommended product information and generate rich text recommendation content and interactive marketing cards through preset parsing rules.

[0029] Furthermore, the large language model inference module is deployed on a cloud server, while the dual-path parallel retrieval module is deployed on edge computing nodes. The two communicate asynchronously through message queues to balance the differences in retrieval latency. The system also includes a model training and index building module, which is used to dynamically update conflict handling rules and preset parsing rules.

[0030] Furthermore, the information fusion module uses a reinforcement learning mechanism to dynamically optimize conflict handling rules, and automatically adjusts the weight distribution ratio between structured query results and unstructured text fragments based on A / B test data feedback.

[0031] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, implement the aforementioned large-model-based conversational shopping guide method.

[0032] A computer program product, comprising a computer program, which, when executed by a processor, is used to implement the aforementioned large-model-based conversational shopping guide method.

[0033] The beneficial effects of this invention are:

[0034] This invention significantly improves the overall performance of conversational shopping guide systems through an innovative dual-path parallel retrieval and intelligent fusion architecture, with the following specific benefits:

[0035] Firstly, it achieves a dual satisfaction of precise constraints and semantic understanding. This invention uniquely executes structured queries and unstructured searches in parallel. It precisely matches hard constraints such as price, brand, and specifications through an NL2SQL model, while simultaneously capturing soft requirements such as functional descriptions and usage scenarios through semantic vector retrieval. The two results are organically integrated through conflict handling rules, enabling the system to understand both the colloquial expression "boiling water on both sides" and the verification of the existence of corresponding functional fields in the database, fundamentally solving the problem of information partiality in single-path solutions.

[0036] Secondly, an intelligent conflict resolution mechanism based on multi-source knowledge has been established. The authoritative data priority rule and differential information supplementation strategy proposed in this invention can automatically handle information conflicts and missing fields between structured data and unstructured documents. When the description of the same product is inconsistent in two results, the system automatically prioritizes the structured data while using the supplementary descriptions in the documents as the basis for reasoning. This effectively avoids recommendation errors caused by data inconsistency and significantly improves the accuracy and reliability of the recommendation results.

[0037] Thirdly, it provides highly interactive rich text output capabilities. This invention breaks through the limitations of traditional plain text output, converting model output into interactive marketing cards through a protocol-based parsing mechanism. The main product card supports direct jumps to the details page, and associated cards can recommend value-added services such as accessories and financial solutions, forming a complete marketing loop. This visual and operable interactive format significantly improves user experience and purchase conversion efficiency.

[0038] Fourth, it reduces system implementation and maintenance costs. This invention can directly reuse existing enterprise product databases and documentation, eliminating the need to build large-scale QA knowledge bases or manually labeled intent systems. Different product categories can be adapted through domain-specific fine-tuning, significantly shortening the implementation cycle. Simultaneously, the system supports incremental updates; when product information changes, only the database and document indexes need to be updated, without retraining the core model, significantly reducing long-term maintenance costs.

[0039] Fifth, it expands the application boundaries of intelligent shopping guides. The architecture of this invention has good versatility and scalability, making it suitable not only for standardized products such as home appliances and 3C digital products, but also demonstrating unique advantages in complex scenarios with high unit prices and long decision-making cycles, such as automobiles and real estate. By introducing a reinforcement learning dynamic optimization mechanism, the system can automatically adjust the weight of the retrieval strategy based on user feedback, continuously improving the personalized recommendation effect and providing a reusable technical paradigm for the intelligent transformation of vertical industries. Attached Figure Description

[0040] Figure 1 This is a flowchart of the dialogic shopping guide method based on a large model according to the present invention;

[0041] Figure 2 This is the system architecture diagram of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] Reference Figure 1 The illustrated dialogic shopping guide method based on a large model includes the following steps:

[0044] S1. Receive the natural language query statement input by the user;

[0045] This step involves the system capturing user-initiated product inquiry requests in real time through multiple interactive channels. Users can type text descriptions into the text input box on their smart terminals or use the voice capture module to convert spoken requests into a text sequence. The system also supports image text recognition, allowing users to upload images containing their needs and automatically extracting the text content. Regardless of the input method, the system obtains the original query content as a string. This content is typically colloquial and non-standardized natural language, and may include multi-dimensional information such as product category, brand preference, functional characteristics, price range, and usage scenarios. It often includes linguistic phenomena such as omissions, pronoun references, and inversions.

[0046] The server-side API service continuously listens for HTTP requests from the front-end application. Upon receiving a valid query data packet, it immediately performs a preliminary validity check. This check includes verifying if the string length exceeds the system's processing capacity, filtering illegal input containing malicious code or abnormal characters, and identifying and removing obvious meaningless duplicate content. For queries that pass the validity check, the system stores them in a memory buffer and assigns a unique session identifier, which is used to associate multiple consecutive rounds of conversation with the same user. If the current query belongs to a subsequent round in the conversation context, the system automatically extracts historical conversation records, adding previously mentioned product features, excluded options, supplementary explanations, and other contextual information to the current query to form a complete description of the user's needs.

[0047] During the receiving phase, the system also performs basic semantic integrity checks on the query. When it detects that the user input is too brief or vague, it triggers a clarification mechanism to guide the user to provide more specific descriptive information. For valid queries, after completing all preprocessing, the system forwards them to the subsequent dual-path parallel retrieval module, while recording the receiving timestamp and source channel information for subsequent performance monitoring and log analysis. The entire receiving process adopts an asynchronous non-blocking design, ensuring millisecond-level response latency even in high-concurrency scenarios, providing users with a smooth interactive experience.

[0048] S2. Parallel execution of structured and unstructured queries on the query statement: The query statement is converted into an SQL query statement and executed through a natural language to SQL model to obtain structured query results containing product attributes; at the same time, the query statement is converted into a semantic vector through a text representation model to retrieve the semantic index of the product unstructured document knowledge base and obtain semantically related unstructured text fragments.

[0049] This step specifically involves the following: This system adopts a dual-path parallel architecture to process the same query statement synchronously, obtaining complementary information from different knowledge sources to provide diversified data support for subsequent fusion decision-making.

[0050] In the structured query path, the system first invokes a natural language to SQL model that has been deeply fine-tuned using domain data. This model, based on an open-source pre-trained language model, is trained on a massive corpus of annotated historical queries and corresponding SQL statements using low-rank adaptive fine-tuning techniques. This allows it to accurately map colloquial descriptions of product needs into standardized SQL query statements that conform to database syntax. When a user's query enters the model, the model identifies key constraints such as product category, brand restrictions, price range, and specifications through semantic parsing, automatically constructing a standardized query statement that includes selection fields, filtering conditions, sorting rules, and return quantity limits. The generated SQL statement is then submitted to the backend relational database execution engine for efficient retrieval in a pre-defined structured product data table. This data table covers basic product attribute information, including but not limited to fields such as unique identifier, product name, brand, price, technical specifications, and inventory status. After execution, the system obtains a structured result set that perfectly matches the query constraints, where each record represents a candidate product, and all field values ​​are precise and verifiable discrete data.

[0051] Meanwhile, in the unstructured retrieval path, the system initiates a semantic vector retrieval process. This path first invokes a text representation model, which employs a dual-tower dense vector encoding architecture capable of mapping natural language text of arbitrary length into a dense vector representation in a high-dimensional semantic vector space. The user's query statement, after being encoded by the model, is converted into a numerical vector containing deep semantic information. This vector is then used for approximate nearest neighbor search in a semantic index built on an efficient vector retrieval library. This index pre-vectorizes and spatially partitions all text fragments in the product's unstructured document knowledge base, supporting fast cosine similarity calculation. Based on the similarity score, the system returns multiple text fragments most semantically relevant to the query. These fragments originate from unstructured documents such as product promotional materials, user reviews, and technical specifications, containing rich information that cannot be directly reflected through structured fields, including feature descriptions, usage scenarios, and experience evaluations.

[0052] The two paths are completely decoupled at the physical level and deployed in parallel within independent computing units. The structured query path relies on the database engine's index optimization mechanism to complete precise matching within milliseconds; the unstructured retrieval path relies on the approximate algorithm of the vector retrieval library, also returning semantically relevant results within sub-seconds. The two processing steps are concurrently controlled through an asynchronous programming model, allowing them to start simultaneously without waiting for each other to complete. Once both paths return results, the system temporarily pairs and stores the structured result set corresponding to the same query with the unstructured text fragment in a memory buffer according to a preset session identifier, preparing for information fusion in subsequent steps. This parallel design significantly shortens the overall response time, avoids the accumulated latency caused by serial execution, and maximizes the breadth and depth of information acquisition. This allows the system to grasp both precise attribute constraints and understand implicit semantic requirements, laying a data foundation for generating high-quality recommendation results.

[0053] S3. Integrate the query statement, structured query results and unstructured text fragments according to the preset template to construct prompt words. The prompt words contain conflict handling rules, which are used to remove duplicate product information and align fields in the two results.

[0054] Specifically, this step involves the core task of integrating the multi-source heterogeneous information obtained in the first two stages in a structured manner to construct a complete, logically clear, and self-correcting prompt word, providing a high-quality input foundation for the reasoning and decision-making of the large language model.

[0055] The system first extracts the original user query as the primary component of the prompts, fully preserving the original semantics and expression of the user's needs. Then, it formats the result set returned by the structured query path, converting each candidate product record into a descriptive statement consisting of field names and values. These statements are generated strictly based on the actual stored content in the database, ensuring the accuracy of product attributes such as brand, price, technical specifications, and inventory status. All candidate products are arranged in descending order of matching degree, forming a clearly structured and hierarchical list of product information.

[0056] Meanwhile, the system filters and removes redundancy from the text fragments returned by the unstructured search path. These fragments originate from product promotional materials, user reviews, technical specifications, and other documents, and are rich in linguistic expressions such as feature descriptions, user experiences, and scenario recommendations. The system sorts the fragments according to semantic similarity scores, removes text with duplicate content or low relevance, retains the most representative descriptive content, and transforms it into fluent natural language paragraphs.

[0057] During information integration, the system loads a pre-defined prompt template. This template consists of multiple functional blocks, including a user intent reproduction area, a structured data display area, an unstructured text supplement area, and a task instruction area. The template uses specific placeholders to mark the insertion position of each information block, ensuring that different types of information are clearly distinguished in the prompts and avoiding semantic confusion.

[0058] Most importantly, the template incorporates conflict resolution rules to intelligently handle duplicate product information and inconsistencies in fields between the two results. When the same product appears in both the structured and unstructured text fragments, the rule automatically triggers a deduplication mechanism. The system uses unique product identifiers for matching and identification, retaining authoritative attribute data provided by the structured path as the final value to ensure the accuracy of key information such as price and model. For descriptive information that appears only in unstructured text but not in the structured results, the rule allows it to be appended as supplementary content to the product information, enriching the model's reasoning basis. If the same product has different values ​​for the same field in the two results, the rule prioritizes the structured data, while simultaneously alerting the model to the difference and explaining it in the generated recommendation reason if necessary.

[0059] Furthermore, the field alignment mechanism handles inconsistencies in field names through semantic mapping. For example, "rated power" in a structured field and "power consumption" in unstructured text are automatically identified as the same attribute, ensuring the accuracy of information classification. For missing fields, the system uses whitespace annotations to indicate to the model that the information was not obtained, preventing the model from making assumptions.

[0060] S4. Input the prompt words into the large language model that has been fine-tuned with domain data to generate structured text output containing recommended product identifiers and reasons for recommendation;

[0061] Specifically, this step involves feeding the aforementioned constructed prompts into the core reasoning engine. Through the deep semantic understanding and generation capabilities of the large language model, structured recommendation content that conforms to the preset protocol is generated.

[0062] The system employs a large language model based on an open-source foundational model architecture, which underwent initial pre-training on a massive general-purpose text corpus, demonstrating powerful language understanding and generation capabilities. Building upon this foundation, the model was further fine-tuned using domain-specific data. This fine-tuning process utilized real-world dialogue data from large-scale product recommendation scenarios, encompassing complete process samples of user questions, search results, recommendation reasons, and final decision-making. Through low-rank adaptive fine-tuning techniques, the model retains its general-purpose language capabilities while deeply learning specialized semantic patterns, expression habits, and output norms specific to the product recommendation domain. This enables it to accurately understand the meaning of database fields, the technical terminology of product descriptions, and the implicit intent of user needs.

[0063] Once the prompt words are input into the model, the model first performs semantic encoding on the prompt words as a whole. The deep transformer structure within the model uses a multi-layered self-attention mechanism to capture the complex relationships between the user query, structured product data, and unstructured text fragments within the prompt words. Specifically, the model can identify the core constraints in the user query and find precisely matching product instances in the structured data; simultaneously, the model can semantically align colloquial expressions in the user query with functional descriptions in the unstructured text, uncovering product characteristics that are not explicitly labeled but meet the user's potential needs. Based on a thorough understanding of multi-source information, the model initiates an autoregressive generation process, gradually constructing the output sequence.

[0064] The generated output strictly adheres to a pre-defined protocol format, which has been deeply internalized during model fine-tuning. The output consists of multiple functional fields. The primary field is the unique identifier of the recommended product, which directly corresponds to the product's primary key in the database or the entity number in the document library, ensuring the system can accurately locate the target product. Following this is the recommendation reason text. Here, the model integrates the results from both search engines to generate logically rigorous and naturally expressed descriptive content. The recommendation reason clearly states in which aspects the product meets the user's essential needs, such as price range matching and brand compatibility. It also explains, in conjunction with the functional descriptions in the unstructured text, how the product fits the user's usage scenarios and preferences.

[0065] The output format may also include auxiliary fields indicating the type of marketing card, recommendation identifiers for associated services, and other information. These fields are separated by clear markers to facilitate accurate extraction by the subsequent parsing module. The entire generation process is controlled within the model through a constraint decoding mechanism, ensuring that the content generated at each step conforms to the protocol specifications and avoids formatting errors or missing information. The final output structured text retains the fluency and readability of natural language while possessing machine-parsable structured features, providing a standardized data foundation for the subsequent implementation of rich text display and interactive functions.

[0066] S5. Parse the structured text output using preset parsing rules, extract recommended product information, and generate rich text recommendation content and interactive marketing cards. The marketing cards contain hyperlinks that redirect to the product details page.

[0067] Specifically, this step, as the final output of the system, undertakes the crucial task of converting the structured text generated by the large language model into visual interactive content that can be directly presented to users, thus realizing the transformation from machine-readable data to a human-friendly interface.

[0068] The system first loads a pre-defined set of parsing rules, which uses regular expressions and syntax tree recognition rules as its core to perform pattern matching on the protocolized format output by the large language model. The parsing engine scans the output text sequentially, identifying special symbols and delimiters used to mark different fields, and breaking down continuous character sequences into independent semantic units. During the scanning process, the rule set prioritizes locating and extracting the unique identifier of the recommended products. This identifier typically begins with a specific prefix followed by a coded string of numbers or letters. After capturing this identifier, the parsing engine immediately initiates a query request to the backend product information service system to retrieve the complete product profile corresponding to the identifier, obtaining multi-dimensional attribute data including the product's main image, detailed name, real-time price, and inventory status, providing rich material for subsequent rich text content construction.

[0069] Next, the parsing engine extracts the recommendation reason text field. This field, located after the identifier, consists of a natural language description generated by the model. The engine performs integrity checks on this text to ensure its fluency and consistency with the product attributes. If abnormally short text or garbled content is detected, an exception handling mechanism is triggered, reverting to a pre-set fallback recommendation script. The recommendation reason text that passes the check is marked as the main body of the rich text content, awaiting fusion and rendering with the product information.

[0070] After obtaining the core recommendation information, the parsing engine continues to identify marketing card-related fields in the output. These fields may contain additional information such as the main card type identifier and associated service identifiers. For the main product card, the engine queries product image resources based on the extracted identifiers, encodes the obtained image binary data into a format suitable for network transmission, and combines and encapsulates it with the product name and price text. For associated marketing cards, the engine queries the corresponding service resource pool based on the identified service type identifiers, which may include extended warranty services, recommended accessories, financial installment plans, etc. Each service includes its own icon, title, description, and redirection path information.

[0071] All extracted and retrieved information is integrated by the rendering engine into the final rich text recommendation content. The rendering engine formats the recommendation reason text as paragraph text, below which are arranged the main product card and related marketing cards. The main product card adopts a visually prominent card layout, with the main product image occupying the upper half of the card, and the product name and price information overlaid below the image. The entire card area is set as a clickable interactive hotspot. Related marketing cards are arranged horizontally below the main card in thumbnail form. Each small card contains a service icon and a short title, and is also clickable.

[0072] To enable redirection, the system embeds hyperlinks for each card during rendering. The hyperlink to the main product card points to a Uniform Locator (URL) for the product details page. This URL includes the product identifier as a parameter. Clicking this URL will open a new page or redirect to the details page route, directly displaying the product's complete information. Hyperlinks to related marketing cards point to different targets depending on the service type, such as an accessories list page, a financial services application page, or an after-sales service appointment page. All links undergo security checks to ensure the redirection target is legitimate and valid.

[0073] The final generated rich text recommendation content is returned to the front-end application in the form of a structured data package. The front-end then adapts its layout and rendering based on the device type and screen size. On desktop, the recommendation content may be presented in a left-right split format, with the text and main card on the left and a list of related services on the right. On mobile devices, the content uses a vertical scrolling layout to ensure good readability and ease of use even on small screens. The entire parsing and rendering process is completed within milliseconds, with virtually no perceptible delay for users. They receive rich, interactive, and visually appealing recommendation results, significantly improving information retrieval efficiency and the purchase conversion experience.

[0074] Furthermore, parallel execution of structured queries and unstructured retrieval specifically includes:

[0075] The natural language to SQL model is trained using a low-rank adaptive fine-tuning method based on historical query-SQL labeled corpus pairs, and the SQL execution accuracy is used as the evaluation metric.

[0076] The text representation model employs a dual-tower dense vector encoder, and the semantic index uses the FAISS library to construct an HNSW index and returns the top K text segments with the highest cosine similarity.

[0077] Furthermore, the conflict resolution rules include:

[0078] When structured query results and unstructured text fragments contain the same product, the product attribute data in the structured query results shall be used as the authoritative data.

[0079] When the field values ​​of the same product are inconsistent in the two results, the structured data shall prevail, and the differences in the unstructured text shall be added to the prompt words as supplementary descriptions.

[0080] Furthermore, the preset parsing rules extract product identifiers and recommendation reason texts using regular expressions; the interactive marketing cards include main product cards and related marketing cards, where the main product card contains the main product image, name, price, and first hyperlink, and the related marketing cards contain accessory or service information and second hyperlink.

[0081] Furthermore, the structured product data table is a car model data table, and the unstructured product document knowledge base is a car model description document library. The method is applied to the car purchase scenario, and the marketing cards display specific car models that meet the user's configuration needs and their associated financial and insurance service options.

[0082] Reference Figure 2 The illustrated dialogic shopping guide system based on a large model includes:

[0083] The query receiving module is used to receive natural language query statements input by the user;

[0084] The dual-path parallel retrieval module includes a natural language to SQL model unit and an SQL execution unit, a text representation model unit and a semantic index query unit, which are used to execute structured queries and unstructured retrievals in parallel.

[0085] The information fusion module is used to integrate query statements, structured query results, and unstructured text fragments according to preset templates to construct prompt words and execute conflict handling rules;

[0086] The large language model inference module is used to process prompt words and generate structured text output based on a large language model fine-tuned by domain data;

[0087] The output parsing and display module is used to extract recommended product information and generate rich text recommendation content and interactive marketing cards through preset parsing rules.

[0088] Furthermore, the large language model inference module is deployed on a cloud server, while the dual-path parallel retrieval module is deployed on edge computing nodes. The two communicate asynchronously through message queues to balance the differences in retrieval latency. The system also includes a model training and index building module, which is used to dynamically update conflict handling rules and preset parsing rules.

[0089] Furthermore, the information fusion module uses a reinforcement learning mechanism to dynamically optimize conflict handling rules, and automatically adjusts the weight distribution ratio between structured query results and unstructured text fragments based on A / B test data feedback.

[0090] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, implement the aforementioned large-model-based conversational shopping guide method.

[0091] A computer program product, comprising a computer program, which, when executed by a processor, is used to implement the aforementioned large-model-based conversational shopping guide method.

[0092] In this embodiment, the present invention constructs an end-to-end intelligent shopping guide system, which achieves fully automated processing from user query to final recommendation display through a modular architecture design. The system as a whole adopts a layered and decoupled design, and the modules communicate with each other through standardized interfaces to ensure high scalability and maintainability.

[0093] The query receiving module, serving as the system's sole entry point, continuously monitors user requests from multiple channels. This module supports three interaction methods: text input, speech recognition, and image text extraction, enabling it to capture user-expressed product needs in natural language. An integrated input validation subunit performs preliminary screening for string length, character validity, and semantic completeness, filtering out invalid or malicious requests. For valid queries that pass validation, the module assigns a globally unique session identifier and queries the user's historical dialogue context, integrating product preferences, excluded options, and supplementary information from previous rounds with the current query to form a complete description of the user's needs. The processed query statement is then pushed to the downstream dual-path parallel retrieval module.

[0094] The dual-path parallel retrieval module is the core data acquisition unit of the system, containing two functionally independent but synchronously executing sub-units. The Natural Language to SQL model unit handles structured query paths. This unit incorporates a lightweight language model fine-tuned with domain data, capable of mapping colloquial query expressions into standardized query statements conforming to database syntax. During the training phase, the model learns the mapping relationship between massive amounts of historical queries and corresponding structured statements, mastering the ability to extract key constraints such as product categories, brand limitations, price ranges, and specifications from natural language. The generated query statements are sent to the SQL execution unit, which establishes a persistent connection with the backend relational database. Utilizing the database engine's index optimization mechanism, it quickly retrieves the product structured data table, obtaining a set of candidate products that precisely match the hard constraints.

[0095] Meanwhile, the text representation model unit drives the unstructured retrieval path. This unit employs a dual-tower dense vector encoding architecture to convert user queries into high-dimensional semantic vectors. The semantic index query unit, based on an approximate nearest neighbor index built from an efficient vector retrieval library, performs fast similarity searches in the product's unstructured document knowledge base, returning text fragments highly relevant to the query semantics. These fragments originate from product promotional materials, user reviews, and technical specifications, containing rich functional descriptions and usage scenario information. The two sub-units are physically decoupled and deployed, achieving parallel execution through an asynchronous concurrency control mechanism. Retrieval results are temporarily stored in a memory buffer paired according to session identifiers.

[0096] The information fusion module is responsible for integrating heterogeneous multi-source information into unified prompts. The module has pre-set standardized templates, divided into user intent reproduction, structured data display, unstructured text supplementation, and task instruction areas. After loading the template, the module fills in the original query statement, structured candidate product data, and semantic search text fragments according to these areas. Conflict handling rules embedded in the template are activated at this stage, automatically identifying duplicate product information in the two results and deduplicating it through unique identifier matching. For differences in fields for the same product from different sources, the rules prioritize the authoritative attribute values ​​from the structured data while retaining supplementary descriptions from the unstructured text as value-added information. A field alignment mechanism handles naming inconsistencies, such as identifying "rated power" and "power consumption" as the same attribute. The fused prompts are then passed to the large language model inference module after integrity verification.

[0097] The large language model inference module carries the core inference capabilities and is deployed on a high-performance cloud server. The module incorporates a base model finely tuned from large-scale product recommendation dialogue data, deeply learning domain knowledge patterns while retaining general language capabilities. After inputting prompts, the model performs semantic encoding through a deep transformer structure, capturing the complex relationships between the query and multi-source information. The decoding phase employs a constraint generation strategy to ensure that the output strictly adheres to a preset protocol format. The generated structured text includes unique identifiers for recommended products, natural language recommendation reasons, and marketing card type tags. The inference process utilizes a model service framework to achieve batch processing and concurrency control, ensuring a stable sub-second response latency.

[0098] The output parsing and display module transforms the model output into visual interactive content. The module first loads a preset parsing rule set and extracts product identifiers, recommendation reason text, and card tags through pattern matching. The identifier triggers a query in the backend product information service to obtain a complete product profile, including the main image, name, real-time price, and inventory status. The recommendation reason text, after validation, is marked as rich text. For card tags, the module distinguishes between main product cards and associated marketing cards. The main card integrates the product's main image and price information, while associated cards query service resources such as accessories and financial solutions. The rendering engine integrates all elements into a rich text layout adapted to different terminals. The main card is set as a clickable hotspot and embeds a hyperlink to the product details page, while associated cards embed links to service pages. Finally, the data package is returned to the frontend application, completing adaptive rendering and forming a visually appealing and directly interactive recommendation result.

[0099] Further system optimization

[0100] In the optimized deployment architecture, the dual-parallel retrieval module has been lightweighted and deployed on edge computing nodes. These edge nodes are located in CDN data centers close to users, integrating a lightweight NL2SQL model and a text vectorization model to process user queries locally, reducing the average latency to less than 500 milliseconds. The large language model inference module remains in the cloud center, utilizing cluster computing power to handle complex inference. Asynchronous communication between the edge and cloud is achieved through a high-throughput message queue. Edge nodes package and push the two retrieval results to the cloud queue. The cloud consumes the messages, executes model inference, and returns the results, keeping the overall end-to-end latency below 800 milliseconds. The system also includes a model training and index building module. This module continuously collects real online interaction data, periodically triggers incremental fine-tuning of the NL2SQL model and text representation model, and dynamically updates the vector index based on changes in product information to ensure the timeliness of the knowledge base.

[0101] A deeper optimization mechanism

[0102] The information fusion module incorporates a reinforcement learning mechanism to dynamically optimize conflict resolution rules. This mechanism uses the weighting ratio of structured query results to unstructured text fragments as adjustable policy parameters, user click-through rate, conversion rate, and dialogue rounds as environmental states, and collects user feedback data through online A / B testing as reward signals. The reinforcement learning agent continuously interacts with the environment, constantly adjusting the weighting parameters to maximize long-term cumulative rewards. After thousands of rounds of iterative training, the weighting strategy converges to the optimal solution, enabling the system to adapt to different product categories and user demand patterns, further improving recommendation accuracy.

[0103] Computer-readable storage media and the product of this program

[0104] This invention also covers computer-readable storage media that carry system implementation logic, including but not limited to portable hard disks, solid-state memories, read-only memories, erasable programmable read-only memories, or flash memory. After the executable instructions stored on the medium are loaded by the processor, a system runtime environment is constructed in memory, sequentially executing the entire process of query reception, dual-path retrieval, information fusion, model reasoning, and output parsing, ensuring the complete reproducibility of the above method steps.

[0105] Furthermore, this invention is embodied as a computer program product, provided in the form of a compressed package, image file, or online distribution, containing computer program code that implements the functions of all modules. After being deployed on a server, the program creates instances of each functional module during processor execution, enabling communication between modules through memory sharing, message buses, or network interfaces. It provides a unified shopping guide service interface externally, supporting multi-tenant concurrent calls. The program product form covers both locally deployed software licensing models and cloud-hosted SaaS subscription models, ensuring that the invention rights can be claimed and protected under various business models.

[0106] Example 1: System Overall Architecture and Deployment

[0107] The conversational shopping guide system provided by this invention adopts a layered and decoupled microservice architecture, with each functional module deployed as an independent service in a distributed computing environment. The system as a whole is divided into an access layer, a logic processing layer, a data storage layer, and a model service layer, with communication between layers through standardized API interfaces.

[0108] The access layer handles external requests and is deployed behind a load balancer, automatically scaling up and down based on traffic fluctuations. This layer integrates multiple protocol adapters, supports mainstream communication protocols such as HTTP and WebSocket, and performs preliminary authentication, permission verification, and quota management on requests. Valid requests that have undergone preprocessing are encapsulated into standard message formats and asynchronously transmitted to the logic processing layer via a message bus.

[0109] The logic processing layer is the core of the entire system, comprising five major modules: query reception, dual-path retrieval, information fusion, model inference, and output parsing. These modules employ a heterogeneous deployment strategy based on varying resource requirements. The dual-path parallel retrieval module, sensitive to latency, is deployed on edge computing nodes close to users, typically within CDN data centers. These edge nodes are equipped with lightweight model instances, utilizing GPU acceleration cards to achieve millisecond-level response times. The large language model inference module, due to its computational intensity, is deployed on a high-performance cloud computing cluster employing distributed tensor parallelism technology, supporting batch inference and dynamic batch processing to maximize hardware utilization. The information fusion module resides between the edge and the cloud, achieving asynchronous decoupling through message queues. Upon arrival of results from both paths, this module automatically triggers the fusion process, constructs prompts, and pushes them to the cloud model service.

[0110] The data storage layer encompasses two types of knowledge bases. The structured product data utilizes a relational database management system, with tables vertically sharded according to product categories. Primary key indexes and foreign key constraints ensure data consistency. Daily incremental updates of product information are synchronized to the database via an ETL pipeline, a process that includes data cleaning, format validation, and anomaly detection. The unstructured document knowledge base is built on a distributed file system. Original documents are deduplicated, segmented, and vectorized before being stored in a vector retrieval system. The vector index is constructed using a hierarchical navigation small-world algorithm, supporting efficient approximate retrieval of hundreds of millions of vectors. Index updates employ an incremental merging strategy, with vectors from new documents added to the index in batches during off-peak hours at night to avoid impacting online query performance.

[0111] The model service layer operates independently, comprising the training pipeline and inference service. The training pipeline periodically pulls online interaction logs and filters them to form a high-quality training set. The training process employs a low-rank adaptive fine-tuning method, updating only a small number of model parameters while injecting domain knowledge and preserving general capabilities. The trained model undergoes automated evaluation, and once the performance metrics meet the requirements, it is deployed seamlessly via blue-green deployment. The inference service provides a RESTful interface externally and internally implements request queuing, load balancing, and circuit breaker mechanisms to ensure high availability.

[0112] Example 2: Concurrency Control of Dual-Path Parallel Retrieval

[0113] To achieve true parallelism between structured queries and unstructured retrievals, the system employs a coroutine-based asynchronous programming model. Once the query receiving module pushes a user request, the main coroutine immediately creates two sub-coroutines, responsible for executing the structured and unstructured paths respectively.

[0114] The structured query sub-coroutine first feeds the user query into the natural language to SQL model. This model is a lightweight distillation version with parameters controlled at the billion level, deployed on the GPU memory of edge nodes. After receiving the text, the model identifies entities and constraints in the query through lexical analysis, uses an attention mechanism to build a mapping relationship between the query and the database table structure, and decodes to generate a grammatically correct and standardized query statement. The generated statement is submitted for execution through a pre-established database connection pool. The connection pool maintains long-lived connections and limits the number of concurrent connections to prevent database overload. The query results are returned in a streaming manner; each record retrieved is placed in a shared memory buffer, and the main coroutine is notified that the data has arrived.

[0115] Unstructured retrieval sub-coroutines start in parallel, feeding user queries into the text representation model. This model employs a dual-tower architecture, with the query tower and document tower sharing parameters. The query is encoded by multiple layers of transformers, outputting a high-dimensional vector. This vector is then fed into the FAISS retrieval engine, which performs a hierarchical navigation search within a pre-built HNSW index, traversing downwards from the top-level entry point to quickly locate the nearest neighbor vector set. The retrieval process uses multi-threaded concurrent scanning, with each thread responsible for a different partition of the index. Finally, the results from each partition are merged and reordered. The returned document fragments contain the original text and metadata, truncated to a preset length to avoid information overload, and also stored in a shared memory buffer.

[0116] The main coroutine tracks the status of the two sub-coroutines through an event listening mechanism. When it detects that both results have arrived or that either sub-coroutine has timed out, the main coroutine retrieves data from the buffer, encapsulates it into a result object containing a timestamp, source identifier, and confidence score, and passes it to the information fusion module. The timeout mechanism sets a flexible threshold; if the structured path is delayed due to slow database queries, the results of the unstructured path can be temporarily stored in the cache, awaiting synchronization or triggering a degradation strategy.

[0117] Example 3: Implementation Details of Conflict Resolution Rules

[0118] In the prompt word template loaded by the information fusion module, conflict handling rules are embedded in natural language as constraints for model inference. The core implementation logic of the rules revolves around three stages: duplicate product identification, authoritative data source determination, and discrepancy information processing.

[0119] After the structured result set and unstructured text fragments are fed into the fusion module, the deduplication unit iterates through the product identifiers of both results. For structured results, product identifiers are typically database primary keys or standardized SKU codes; for unstructured text, identifiers are extracted from the text using a named entity recognition model. The deduplication unit employs a combination of exact string matching and semantic similarity matching. The former directly compares encoding consistency, while the latter calculates the vector similarity of product names; if the similarity exceeds a threshold, the product is considered the same. Only one record is retained for each identified duplicate product. The attribute data provided by the structured path serves as the main data, while the text description from the unstructured path is appended as a supplementary field to the end of the prompt.

[0120] The authoritative data judgment mechanism prioritizes numerical fields from structured data. For quantifiable attributes such as price, size, and weight, the values ​​in the database are considered the only reliable source. If the price mentioned in unstructured text differs from the database, the rules explicitly require the model to ignore the price information in the text and rely on the database value, stating "based on official data" in the recommendation reason. For descriptive fields such as features and usage scenarios, if structured data is missing but unstructured text exists, the rules allow the model to adopt the text description, but must add limiting statements such as "based on product introduction" during generation to avoid absolute assertions.

[0121] Field alignment is achieved through a semantic mapping dictionary. The dictionary maintains the correspondence between structured field names and synonyms in unstructured text, such as mapping "screen size" to expressions like "display size" and "screen diagonal length." During fusion, the system aggregates all text descriptions mapped to the same structured field, merging them into a comma-separated string, which is then used as supplementary information for that field in the prompt word template. If the aggregated text is too long, the summarization model extracts the core information to ensure the prompt word length does not exceed the model's context window limit.

[0122] Example 4: Protocol-based generation of reasoning for large language models

[0123] The large language model inference module is deployed on a cloud-based Kubernetes cluster. Each Pod runs one model instance, and instances share a distributed cache to accelerate lexical lookup. After receiving a prompt word, the model first performs lexical segmentation, converting Chinese characters into a sequence of sub-word units and adding special markers to distinguish between user queries, structured data, and unstructured text.

[0124] The model employs a grouped query attention mechanism to reduce computational complexity while maintaining long-context modeling capabilities. Within the attention layer, query, key, and value vectors are divided into multiple groups, each with its own independently calculated attention weights, which are then concatenated for output. This design allows the model to focus on different information sources in parallel: some attention heads focus on the core needs of the user query, some model the field relationships in structured data, and the remaining heads capture the semantic details of unstructured text. After multiple layers are stacked, the model integrates multi-source information in a deep representation space to form a unified decision representation.

[0125] The generation phase employs a tree-based constraint decoding algorithm. The decoder maintains a valid lexical prefix tree, where tree nodes represent field markers in the protocol format. During each generation step, the decoder is only allowed to sample lexical terms from child nodes reachable from the current node, ensuring the output strictly adheres to the preset structure. For example, generation must begin with a specific field marker, followed by the recommendation reason text, then the product identifier marker and code, and finally the card type marker. If the model attempts to generate an invalid lexical term, the decoder automatically sets its probability to zero and resamples. This constraint mechanism avoids formatting errors, enabling the parsing module to reliably extract information.

[0126] The inference service implements dynamic batch processing optimization. Individual user requests may arrive separately; the service places them in a buffer queue, waits for fifty milliseconds, or accumulates to eight requests before merging them into a batch and feeding them into the model. Within each batch, padding and masking techniques are used to process variable-length sequences, and matrix operations are highly parallelized along the batch dimension, significantly improving hardware utilization. After inference is complete, the results are split according to the request identifier and returned to the corresponding session thread.

[0127] Example 5: Rendering and Interaction of Rich Text Marketing Cards

[0128] The output parsing and display module is located at the very front of the system, adjacent to the user interface. After receiving the structured text generated by the large language model, the module first performs pattern matching through a regular expression engine. The engine pre-compiles multiple regular expression patterns, which are used to capture product identifiers, recommendation reason text, main card tags, and associated card tags. The identifier pattern matches a combination of sixteen digits and letters, the recommendation reason pattern matches Chinese characters of any length, and the card tag pattern matches symbols with specific HTML tag styles.

[0129] Upon successful matching, the module initiates a remote call to the product information service. This service queries a relational database based on the product identifier, returning the product's main image's Uniform Resource Locator (URI), full product name, current selling price, and promotional tags. For price information, the service additionally calls the promotion subsystem to check for any applicable discounts, coupons, or member discounts; if found, it calculates the final price. The main image resource is obtained from the object storage service, and the image is compressed and converted before being embedded into the returned data packet.

[0130] The rendering of rich text content employs a component-based design. The recommendation text is encapsulated as a text component, supporting style configuration for font size, color, and line height. The main product card is an image component overlaid with a text component. The image component uses lazy loading, downloading only when the card enters the browser viewport, reducing initial loading time. Below the image are the product name and price text, with the price value highlighted in bold and red. The entire card component is bound to a click event handler, which reads the product identifier and constructs a link to the details page. The link includes channel source parameters for performance tracking.

[0131] Related marketing cards are rendered in a horizontal list format. Each card contains an icon component and a text component. The icon is dynamically selected from a pre-built icon library based on the service type, and the text displays the service name and a brief selling point. Each card is also linked to a click event, linking to a service landing page, such as a financial solutions page, an accessories purchase page, or an online customer service page. If more than three services are associated, the list can be scrolled horizontally, allowing users to browse more options using gestures.

[0132] In the mobile application, the rendering engine is implemented using native components. The main card uses an ImageView to load the product image, a TextView to display the text, and a CardView to wrap the entire card, providing rounded corners and shadow effects. Click events trigger native route navigation, opening either the built-in browser or the native details page. Related services use a RecyclerView with a horizontal layout, supporting smooth scrolling and click feedback to enhance the user experience.

[0133] Example 6: Continuous Optimization of Model Training and Index Construction

[0134] The model training and index building module serves as the backend support system, triggering a batch processing task every day at midnight. This task retrieves the previous day's full user dialogue logs, filtering out negative samples with conversations shorter than ten seconds or where users explicitly expressed dissatisfaction, retaining high-quality interactions as training data. The data undergoes anonymization, removing sensitive information such as user names and phone numbers, and product identifiers are uniformly replaced with internal codes.

[0135] The training data for the NL2SQL model consists of query-SQL pairs. For each log entry, the system replays the successfully executed SQL statement, using the original user question as input and the executed SQL as labels to construct supervised learning samples. Training employs a low-rank adaptive fine-tuning method, freezing the model's backbone parameters and training only the injected low-rank matrices. Each weight matrix adds only a few million trainable parameters, allowing the fine-tuning process to be completed within hours on a single GPU. The training objective is to minimize the lexical differences between the generated SQL and standard SQL, while using contrastive learning loss to narrow the representational distance between similar queries.

[0136] The text representation model is trained using a contrastive learning paradigm. Positive sample pairs are different expressions of the same query or multiple descriptive documents of the same product, while negative sample pairs are semantically unrelated query-document combinations. The model optimizes by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs, making the vector space discriminative. The training data is enhanced using a hard negative sample mining strategy, selecting documents that are highly similar to the query but irrelevant in content from a massive pool of candidates as hard negative examples, thus improving the model's robustness.

[0137] The vector index is built using an incremental update strategy. Newly added product documents each day first undergo text cleaning, removing HTML tags and advertising copy while retaining the core product description. The cleaned text is then segmented into fixed-length fragments, each encoded as a vector using a text representation model. New vectors are not immediately added to the online index but are written to a buffer. When the buffer accumulates to 10,000 vectors or reaches a preset time window, an index merging operation is triggered. A hierarchical, regressive nearest neighbor graph algorithm efficiently integrates the new vectors into the existing index. The merging process is performed on a replica index, and an atomic switchover is completed to ensure uninterrupted service.

[0138] Example 7: In-depth application of car buying guide scenario

[0139] This invention demonstrates significant advantages in the automotive vertical industry. Automotive products are characterized by complex parameters, high prices, and long decision-making cycles, and user needs typically involve a comprehensive trade-off across dozens of dimensions. After deploying this system, automotive e-commerce platforms achieve an intelligent upgrade in their sales guidance system.

[0140] A user asked, "Recommend an SUV under 200,000 yuan that is fuel-efficient, spacious, and suitable for family trips." After receiving the query, the system uses a structured path to retrieve data from a car model table. This table contains over two hundred fields, covering all dimensions of parameters including price, fuel consumption, wheelbase, number of seats, safety features, and smart cockpit. An NL2SQL model accurately generates the query, filtering out a candidate set of compact SUVs that meet the price criteria and have fuel consumption below 7 liters per 100 kilometers. An unstructured path is used to retrieve user reviews, professional media test drive reports, and official promotional materials to obtain semantic descriptions of "spaciousness" and "convenience for family trips." The fusion module associates the wheelbase data of a specific model with the description of "ample rear legroom" from reviews to construct prompts. After comprehensive analysis by the large language model, a recommendation is output. The recommendation reasons include precise data such as "fuel consumption of 6.8 liters per 100 kilometers, meeting fuel-saving needs" and scenario-based descriptions such as "three adults can sit comfortably in the back row, and the trunk can accommodate a stroller and camping equipment". The generated marketing card links to the vehicle details page and is associated with "zero down payment financial plan" and "three years of free maintenance" service cards.

[0141] Real-world application data shows that the system's accuracy in understanding complex needs in this scenario is industry-leading, the average number of user dialogue rounds has decreased from five to two, and the click-through rate of financial services on marketing cards has reached a significant level, driving the growth of revenue from value-added financial services. This verifies the commercial value and technological advancement of this invention in guiding high-priced goods.

[0142] By adopting the above-disclosed technical solution of this invention, the following beneficial effects are obtained:

[0143] Dual-path retrieval improves accuracy and cost-effectiveness.

[0144] This invention, through its innovative dual-path parallel retrieval and intelligent fusion architecture, fundamentally solves the core pain points of traditional shopping guide systems, such as one-sided information from a single path and difficulty in coordinating multi-source knowledge, bringing significant and beneficial effects in all aspects and dimensions.

[0145] First, this invention achieves a dual breakthrough in precise attribute constraints and deep semantic understanding. Traditional solutions often fall into a dilemma: pure structured queries cannot understand the implicit intentions in users' colloquial descriptions, while pure semantic retrieval cannot verify the authenticity of hard conditions such as price and inventory. The invention's unique parallel mechanism allows both capabilities to work simultaneously, enabling the system to accurately match objective constraints such as user budget ranges and brand requirements, while also deeply understanding subjective demands such as usage scenarios and functional preferences. This fusion is not a simple addition, but rather an organic complementarity achieved through conflict handling rules, allowing the two types of information to mutually verify and supplement each other, significantly improving the accuracy of recommendations under complex needs.

[0146] Secondly, this invention establishes a reliable multi-source knowledge arbitration mechanism. When inconsistencies arise between specifications in the database and product description documents, the system can intelligently identify the conflict and automatically adjudicate based on authoritative data sources, effectively avoiding recommendation errors caused by information contradictions. This built-in conflict resolution capability greatly enhances the system's robustness and reliability, enabling it to maintain stable performance in e-commerce environments where product information changes frequently, reducing the frequency and cost of manual maintenance intervention.

[0147] Third, this invention reshapes the user interaction experience and business conversion path. Traditional text recommendation formats are monotonous, requiring users to jump to a details page to obtain complete information, resulting in a long decision-making chain and high drop-off rates. The interactive marketing cards generated by this invention integrate core product information, reasons for recommendation, visual elements, and jump entry points, allowing users to complete the closed loop from awareness to interest to action directly on the recommendation page. The main card intuitively displays the product image, while related cards extend recommendations for complementary services. This rich media format not only improves information delivery efficiency but also significantly improves marketing effectiveness by shortening the conversion path.

[0148] Fourth, this invention significantly reduces the technical implementation threshold and long-term maintenance costs. Existing solutions rely heavily on manual construction of knowledge bases and intent annotation systems, requiring substantial upfront investment and proving difficult to continuously update. This invention directly reuses existing enterprise product databases and documentation, enabling rapid deployment through automated fine-tuning and index building. When product information changes, only the data source needs to be updated, triggering incremental index updates; there is no need to retrain the core model, achieving adaptive evolution with zero manual intervention. This is particularly suitable for the fast-moving consumer goods (FMCG) sector, characterized by short product lifecycles and high update frequencies.

[0149] Finally, this invention possesses outstanding cross-domain migration capabilities and commercial expansion potential. From a technical architecture perspective, the dual-path retrieval, conflict handling, and protocol-based output design are universal, allowing for rapid adaptation to different categories such as automobiles, real estate, digital products, and home furnishings simply by changing the domain and fine-tuning the data. From a business model perspective, the system not only improves the accuracy of product recommendations but also creates incremental value entry points such as finance, insurance, and accessories through associated service cards, opening up diversified monetization avenues for the platform and forming a virtuous cycle of technology-driven business growth. This reusable and scalable architecture design makes it an ideal infrastructure for enterprise intelligent transformation.

[0150] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A conversational shopping guide method based on a large model, characterized in that, Includes the following steps: Receive natural language query statements input by the user; The query statement is processed in parallel with structured query and unstructured retrieval: the query statement is converted into an SQL query statement through a natural language to SQL model and executed to obtain structured query results containing product attributes; at the same time, the query statement is converted into a semantic vector through a text representation model, and the semantic index of the product unstructured document knowledge base is retrieved to obtain semantically related unstructured text fragments. The query statement, structured query results, and unstructured text fragments are integrated according to a preset template to construct prompt words. The prompt words include conflict handling rules, which are used to remove duplicate product information and align fields in the two results. The prompt words are input into a large language model fine-tuned by domain data to generate structured text output containing recommended product identifiers and reasons for recommendation; The structured text output is parsed using preset parsing rules to extract recommended product information and generate rich text recommendation content and interactive marketing cards. The marketing cards contain hyperlinks that redirect to the product details page.

2. The method according to claim 1, characterized in that, The parallel execution of structured queries and unstructured retrieval specifically includes: The natural language to SQL model is trained using a low-rank adaptive fine-tuning method based on historical query-SQL labeled corpus pairs, and the SQL execution accuracy is used as the evaluation metric. The text representation model employs a dual-tower dense vector encoder, and the semantic index uses the FAISS library to construct an HNSW index and returns the top K text segments with the highest cosine similarity.

3. The method according to claim 1, characterized in that, The conflict handling rules include: When structured query results and unstructured text fragments contain the same product, the product attribute data in the structured query results shall be used as the authoritative data. When the field values ​​of the same product are inconsistent in the two results, the structured data shall prevail, and the differences in the unstructured text shall be added to the prompt words as supplementary descriptions.

4. The method according to claim 1, characterized in that, The preset parsing rules extract product identifiers and recommendation reason texts using regular expressions; the interactive marketing cards include main product cards and related marketing cards, wherein the main product card contains the main product image, name, price and a first hyperlink, and the related marketing cards contain accessory or service information and a second hyperlink.

5. The method according to claim 4, characterized in that, The structured data table for the products is a car model data table, and the unstructured document knowledge base for the products is a car model description document library. The method is applied to the car purchase scenario, and the marketing card displays specific car models that meet the user's configuration needs and their associated financial and insurance service options.

6. A conversational shopping guide system based on a large model, characterized in that, include: The query receiving module is used to receive natural language query statements input by the user; The dual-path parallel retrieval module includes a natural language to SQL model unit and an SQL execution unit, a text representation model unit and a semantic index query unit, which are used to execute structured queries and unstructured retrievals in parallel. The information fusion module is used to integrate query statements, structured query results, and unstructured text fragments according to preset templates to construct prompt words and execute conflict handling rules; The large language model inference module is used to process prompt words and generate structured text output based on a large language model fine-tuned by domain data; The output parsing and display module is used to extract recommended product information and generate rich text recommendation content and interactive marketing cards through preset parsing rules.

7. The system according to claim 6, characterized in that, The large language model inference module is deployed on a cloud server, and the dual-path parallel retrieval module is deployed on an edge computing node. The two communicate asynchronously through a message queue to balance the difference in retrieval latency. The system also includes a model training and index building module, which is used to dynamically update the conflict handling rules and preset parsing rules.

8. The system according to claim 6, characterized in that, The information fusion module uses a reinforcement learning mechanism to dynamically optimize the conflict handling rules and automatically adjusts the weight distribution ratio between structured query results and unstructured text fragments based on A / B test data feedback.

9. A computer-readable storage medium having executable instructions stored thereon, characterized in that, When the instructions are executed by the processor, they implement the dialogic shopping guide method based on a large model as described in any one of claims 1-5.

10. A computer program product, the computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it is used to implement the dialogic shopping guide method based on a large model as described in any one of claims 1-5.

Images