Automatic inquiry, retrieval and analysis method for materials in electronic shopping mall
By combining a large language model and an automatic price comparison module, the problem of parsing unstructured demands in the e-commerce material retrieval system was solved, achieving multi-dimensional adaptation and price monitoring, thereby improving the accuracy of material procurement and risk management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANENG ENERGY & COMM HLDG CO LTD
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-17
AI Technical Summary
Existing e-commerce material retrieval systems cannot accurately parse unstructured material demands in pure natural language form. The material recommendation mechanism lacks multi-dimensional semantic adaptation capabilities and lacks automated price comparison monitoring of external market prices and historical transaction data, resulting in blind spots in price control during the procurement process.
The system employs a large language model parsing module to perform structured semantic analysis on pure natural language text, combines an industry keyword dictionary to generate standardized search keywords, conducts multi-dimensional comprehensive analysis through a personalized recommendation module, and utilizes an automatic price comparison module to monitor price deviations and generate discrepancy alerts.
It enables accurate analysis and personalized recommendations of unstructured material needs, ensuring the accuracy and completeness of search results. It also monitors price risks through external market data to help users avoid price premiums.
Smart Images
Figure CN121880650A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of e-commerce material procurement technology, specifically to an automated query, retrieval, and analysis method for e-commerce material procurement. Background Technology
[0002] In the procurement scenarios of power, energy and large industrial enterprises, e-commerce platforms are the main channels for material inquiry and procurement. Traditional e-commerce platform searches mainly rely on preset keyword matching technology or category filtering logic to perform material retrieval tasks. When performing a search operation, users must enter the precise material name or standard specification model that conforms to the database specifications. When users enter unstructured demand text containing natural language descriptions, non-standard terms or fuzzy parameters, the keyword matching engine based on rigid rules cannot perform semantic understanding and intent recognition of unstructured demand text, resulting in empty search results or search results containing a large number of irrelevant entries.
[0003] Existing material recommendation mechanisms use statistical ranking algorithms based on sales volume, price, or click-through rate. These algorithms struggle to handle industrial material recommendation tasks with complex technical parameter constraints, fail to deeply analyze the compatibility between material technical specifications and application scenarios, and do not fully utilize information on manufacturer preferences and price sensitivity in users' historical purchase records. The lack of multi-dimensional semantic analysis capabilities in these recommendation mechanisms makes it difficult for the recommended lists to meet the procurement needs of highly specialized materials.
[0004] The price display function in the e-commerce platform only shows the current internal sales price. When making decisions, procurement personnel lack reference data on the market prices of similar products on external e-commerce platforms and the historical transaction price trends. Due to the lack of automated external price comparison and historical price analysis methods, it is impossible to automatically identify the premium risk of material quotations. Relying on manual cross-platform price comparison increases the time cost of procurement operations and makes it difficult to guarantee the real-time and coverage of price monitoring, resulting in price control blind spots in the material procurement process. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention is achieved through the following technical solution: The first aspect of this invention provides an automated query and retrieval analysis method for materials in e-commerce, which solves the problems of existing e-commerce retrieval systems being unable to accurately parse unstructured material demands in pure natural language form, lacking multi-dimensional semantic adaptation capabilities based on technical parameters and historical preferences in material recommendation mechanisms, and lacking an automated price comparison risk monitoring mechanism for external market prices and historical transaction data.
[0006] To achieve the above objectives, the first aspect of this invention provides an automated query, retrieval, and analysis method for materials in an e-commerce platform, comprising the following steps: S1. Natural language requirement input: The natural language interaction module loads and displays the e-commerce interactive interface on the front end of the e-commerce platform. The natural language interaction module receives the material procurement requirement information input by the user on the e-commerce interactive interface. The material procurement requirement information is unstructured text in pure natural language form. The natural language interaction module encapsulates the material procurement requirement information into a pure natural language text sequence and transmits the pure natural language text sequence to the intelligent agent scheduling module through the data interface. S2. Demand Analysis and Keyword Decomposition: The intelligent agent scheduling module starts and triggers the large language model parsing module. The large language model parsing module receives the pure natural language text sequence. The large language model parsing module performs structured semantic analysis on the pure natural language text sequence and extracts the core demand information set from the pure natural language text sequence. The large language model parsing module uses a preset industry keyword dictionary to map the core demand information set into a standardized search keyword set. S3. E-commerce material retrieval: The intelligent agent scheduling module inputs a standardized set of search keywords into the e-commerce retrieval module. The e-commerce retrieval module uses the search engine to perform a matching search operation in the material database and generates a preliminary search result set. S4. Comprehensive analysis and personalized recommendation using a large language model: The agent scheduling module transmits the preliminary search result set, core demand information set, and user historical procurement records to the personalized recommendation module. The personalized recommendation module calls the large language model to perform multi-dimensional comprehensive analysis on each material object in the preliminary search result set. The personalized recommendation module calculates the comprehensive recommendation score for each material object in the preliminary search result set. The personalized recommendation module generates a personalized recommendation list and recommendation reasons based on the comprehensive recommendation score. S5. Automatic price comparison and difference alert: The intelligent agent scheduling module identifies the target material object selected by the user from the personalized recommendation list. The intelligent agent scheduling module transmits the target material object to the automatic price comparison module. The automatic price comparison module calculates the price deviation index of the target material object. The automatic price comparison module compares the price deviation index with the preset price anomaly threshold. When the price deviation index is greater than the price anomaly threshold, a difference alert warning message is generated. S6. Information Push: The intelligent agent scheduling module converts the personalized recommendation list, recommendation reasons, and difference reminders and warnings into front-end visual instructions. The intelligent agent scheduling module sends the front-end visual instructions to the e-commerce interactive interface. The e-commerce interactive interface parses the front-end visual instructions and renders the list of material cards and price risk pop-ups.
[0007] Furthermore, in step S2, the large language model parsing module is configured with an integrity verification function. The large language model parsing module performs integrity verification on the core requirement information set. When the core requirement information set is missing key performance parameters or the budget is unclear, the agent scheduling module triggers a multi-round dialogue follow-up process. The agent scheduling module generates follow-up questions based on the missing information slots. The agent scheduling module receives user feedback on the follow-up questions through the e-commerce interactive interface. The large language model parsing module updates the core requirement information set using the feedback information. The large language model parsing module performs integrity verification again on the updated core requirement information set. The large language model parsing module repeats the multi-round dialogue follow-up process until the integrity verification passes.
[0008] Furthermore, in step S3, the e-commerce retrieval module internally constructs an inverted index database and a vector database. The e-commerce retrieval module performs an exact match retrieval operation based on keyword search queries in the inverted index database to obtain a list of keyword search results. The e-commerce retrieval module performs a semantic similarity retrieval operation based on vector search query vectors in the vector database to obtain a list of vector search results. The e-commerce retrieval module uses an inverted ranking fusion algorithm to perform a result fusion operation on the keyword search result list and the vector search result list, generating a preliminary retrieval result set based on the fusion ranking score.
[0009] Furthermore, in step S4, the personalized recommendation module calls the large language model to perform multi-dimensional comprehensive analysis on each material object in the preliminary search result set. The multi-dimensional comprehensive analysis includes matching degree analysis dimension, historical procurement preference analysis dimension, and comprehensive cost-effectiveness analysis dimension. The personalized recommendation module calculates the matching degree score based on the matching degree analysis dimension, the historical preference score based on the historical procurement preference analysis dimension, and the cost-effectiveness score based on the comprehensive cost-effectiveness analysis dimension. The personalized recommendation module calculates the comprehensive recommendation score by weighting the matching degree score, historical preference score, and cost-effectiveness score.
[0010] Further, in step S4, the personalized recommendation module sorts the preliminary search result set in descending order based on the comprehensive recommendation score, extracting the first-ranked material object at the top of the ranking list and the second-ranked material object at the second position. The personalized recommendation module calculates the score discrimination between the first-ranked and second-ranked material objects, and triggers an interactive preference clarification process when the score discrimination is less than a preset discrimination threshold. The personalized recommendation module calls the large language model to generate a comparative analysis summary, and the agent scheduling module receives the user's preference feedback information on the comparative analysis summary through the e-commerce interactive interface. The personalized recommendation module uses the preference feedback information to adjust the weight coefficient vector and recalculates the comprehensive recommendation score using the adjusted weight coefficient vector.
[0011] Furthermore, in step S5, the automatic price comparison module accesses a preset internet e-commerce platform data source through a web crawler to obtain multiple external original price quotes, which constitute an external original price quote set. The automatic price comparison module performs statistical analysis on the external original price quote set, calculates the arithmetic mean and standard deviation of the external original price quote set, and uses the arithmetic mean and standard deviation to construct outlier filtering conditions. It removes external original price quotes that do not meet the outlier filtering conditions from the external original price quote set, and retains the external original price quotes that meet the outlier filtering conditions as valid price quotes. The automatic price comparison module calculates the external market reference price based on the valid price quotes.
[0012] Furthermore, in step S5, the automatic price comparison module continuously monitors the transaction status of the target material object. Upon receiving a transaction completion signal, it obtains the transaction price of the target material object in this transaction and reads the previous historical average transaction price of the target material object from the historical transaction database. The automatic price comparison module obtains the current time and the last update time, calculates the time interval between the current time and the last update time, calculates the time decay weight coefficient using the time interval, and calculates the updated historical average transaction price using the time decay weight coefficient, the transaction price of this transaction, and the previous historical average transaction price. The updated historical average transaction price is then written into the historical transaction database.
[0013] Furthermore, in step S5, the automatic price comparison module reads the internal price quote, external market reference price, and historical average transaction price of the target material object from the running cache; the automatic price comparison module calculates the minimum value between the external market reference price and the historical average transaction price, sets the minimum value as the price comparison benchmark, and calculates the price deviation index based on the internal price quote of the e-commerce platform and the price comparison benchmark.
[0014] Furthermore, prior to step S1, the large language model selection process is pre-executed. A candidate model set is constructed through the large language model parsing module. Performance tests are performed on the candidate models in the candidate model set on a pre-defined e-commerce domain validation dataset to obtain the intent recognition accuracy and entity extraction F1 score of the candidate models. The large language model parsing module monitors the average inference latency and memory usage of the candidate models in a standard hardware environment. Based on the intent recognition accuracy, entity extraction F1 score, average inference latency, and memory usage, the large language model parsing module calculates the model fit score and loads the candidate model with the highest model fit score into the large language model parsing module.
[0015] The second aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any one of the automated query, retrieval, and analysis methods for e-commerce materials in the first aspect.
[0016] This invention provides an automated method for querying, retrieving, and analyzing goods in an e-commerce platform. It offers the following advantages: 1. This invention utilizes a large language model parsing module and prompt word engineering technology to perform structured semantic analysis on pure natural language text sequences, extracting a core demand information set. This core demand information set is then mapped to a standardized set of search keywords using an industry keyword dictionary. The intelligent agent scheduling module triggers a multi-round dialogue process when the core demand information set is incomplete. This solves the problem of low search matching rates caused by non-standardized user-input material procurement demand information. By combining semantic understanding with standardized mapping, the accuracy and completeness of the standardized search keyword set obtained by the e-commerce search module are ensured.
[0017] 2. This invention uses a personalized recommendation module to call a large language model to perform a multi-dimensional comprehensive analysis on the preliminary search result set, including matching degree analysis, historical procurement preference analysis, and comprehensive cost-effectiveness analysis. It calculates a comprehensive recommendation score and generates recommendation reasons in natural language form. When the score discrimination is insufficient, it triggers an interactive preference clarification process. It uses the logical reasoning ability of the large language model to transform quantitative data into interpretable decision-making basis. Through multi-dimensional weighted calculation, it achieves deep matching between material objects and users' historical procurement records and real-time needs.
[0018] 3. This invention utilizes an automatic price comparison module to obtain a set of external original price quotes through a web crawler and performs outlier filtering based on statistics. The automatic price comparison module combines a time decay weighting coefficient to calculate the updated historical average transaction price, constructing a price comparison benchmark composed of the external market reference price and the historical average transaction price. It calculates the price deviation index, realizing real-time risk monitoring of internal quotations in the e-commerce platform. By introducing external market data and historical transaction trend data to establish a dynamic price reference system, it generates a difference reminder warning when the price deviation index exceeds the price anomaly threshold, helping users avoid premium risks in the material procurement process. Attached Figure Description
[0019] Figure 1 This is a block diagram of the logical architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] See attached document Figure 1 This invention provides an automated query, retrieval, and analysis method for goods in an e-commerce platform, based on a logical architecture that includes a natural language interaction module, an intelligent agent scheduling module, a large language model parsing module, an e-commerce platform retrieval module, a personalized recommendation module, an automatic price comparison module, and an information push module.
[0022] The natural language interaction module provides an interactive interface, receiving material procurement demand information input by the user in pure natural language. The agent scheduling module is connected to the large language model parsing module. After receiving the material procurement demand information, the agent scheduling module triggers the large language model parsing module. The large language model parsing module uses prompt word engineering technology to perform semantic understanding of the material procurement demand information, extracting material type, core parameters, and constraints. It uses an industry keyword dictionary to decompose the material type, core parameters, and constraints into standardized search keywords. When incomplete material procurement demand information is detected, the agent scheduling module and the natural language interaction module initiate multi-round dialogue and follow-up questions.
[0023] The intelligent agent scheduling module is connected to the e-commerce search module. The intelligent agent scheduling module inputs standardized search keywords into the e-commerce search module. The e-commerce search module uses the search engine to match relevant materials and products in the materials database and generates a preliminary search result set that includes material name, model and specifications, manufacturer, price, inventory quantity, quality certification, user reviews and historical transaction records.
[0024] The intelligent agent scheduling module is connected to the personalized recommendation module. The intelligent agent scheduling module inputs the preliminary search result set and the user's historical purchase records into the personalized recommendation module. The personalized recommendation module calls the large language model and uses the large language model to perform multi-dimensional analysis by combining the matching degree between materials and user needs, the user's historical purchase records, and the overall cost-effectiveness of materials. The personalized recommendation module generates a personalized recommendation list sorted by matching degree and generates recommendation reasons.
[0025] The intelligent agent scheduling module is connected to the automatic price comparison module. The automatic price comparison module uses web crawlers or application programming interfaces to obtain external price information of similar products from external e-commerce platforms. The automatic price comparison module is equipped with a dynamic update mechanism and abnormal price removal logic. It re-obtains external platform price information every preset time (e.g., 2 hours) and calculates the difference between the e-commerce platform's price quote and the external price information.
[0026] The intelligent agent scheduling module is connected to the information push module. The information push module pushes a personalized recommendation list and the reasons for the recommendation to the interactive interface. When the difference exceeds a preset threshold (such as 5%), a price difference reminder message is generated and pushed to the interactive interface.
[0027] The natural language interaction module is used to generate the front-end interactive interface of the e-commerce platform. It receives text sequences input by users through the front-end interactive interface. The text sequences are descriptions of material procurement needs in pure natural language. The natural language interaction module then transmits the text sequences to the intelligent agent scheduling module.
[0028] As the core logic control unit, the agent scheduling module establishes data connections with the natural language interaction module, the large language model parsing module, the e-commerce retrieval module, the personalized recommendation module, the automatic price comparison module, and the information push module. The agent scheduling module coordinates the data flow of each functional module, triggers the large language model parsing module after receiving a text sequence, and generates follow-up instructions when receiving incomplete demand signals. The agent scheduling module is built using the LangChain framework or the Dify open-source framework.
[0029] The large language model parsing module is internally deployed with a pre-trained large language model and prompt word engineering templates. The large language model parsing module receives text sequences, uses prompt word engineering templates to perform structured semantic analysis on the text sequences, and maps the core requirement information set into standardized search keywords. The large language model adopts DeepSeek or Qwen open source general models, and is fine-tuned to adapt to the terminology of power generation industry materials.
[0030] The e-commerce search module includes a search engine and a materials database. The e-commerce search module receives standardized search keywords, the search engine performs matching operations in the materials database, and the e-commerce search module outputs a preliminary search result set. The preliminary search result set contains multiple materials objects, and the materials objects include attributes such as material name, model and specifications, manufacturer, e-commerce internal price, inventory quantity, quality certification, user reviews, and historical transaction records.
[0031] The personalized recommendation module is connected to a user behavior database, which stores users' historical purchase records. These records include data on historical purchased models, usage feedback, purchase frequency, and manufacturer preferences. The personalized recommendation module receives the preliminary search results set and the user's historical purchase records, uses a large language model to calculate a comprehensive recommendation score for each material object in the preliminary search results set, and generates a personalized recommendation list and recommendation reasons based on the comprehensive recommendation score.
[0032] The automatic price comparison module is equipped with a web crawler and an application programming interface (API). The automatic price comparison module receives a personalized recommendation list, extracts the material name, model specifications, material and functional parameters of the material objects in the personalized recommendation list, and obtains external price information from external e-commerce platforms through a web crawler (such as based on the Scrapy framework) or an API. When the price difference exceeds a preset threshold and the internal price of the e-commerce platform is higher than the external price information, a price difference reminder message is generated.
[0033] The information push module receives a personalized recommendation list, the reasons for the recommendation, and a price difference reminder, and then renders the personalized recommendation list, the reasons for the recommendation, and the price difference reminder to the front-end interactive interface.
[0034] See attached document Figure 2 This invention provides an automated query, retrieval, and analysis method for materials in an e-commerce platform, comprising the following steps: S1: Natural language input requirement. The natural language interaction module loads and displays the e-commerce interactive interface on the front end of the e-commerce platform. The e-commerce interactive interface is used for text input or speech-to-text input. It receives the material procurement requirement information input by the user. The material procurement requirement information is unstructured text in pure natural language form. The e-commerce interactive interface encapsulates the material procurement requirement information into a pure natural language text sequence.
[0035] The semantic content contained in the pure natural language text sequence covers material names, application scenarios, performance parameters, quantity requirements, budget range, and delivery time dimensions. The pure natural language text sequence is not required to follow specific e-commerce search syntax or keyword format. The pure natural language text sequence is transmitted to the intelligent agent scheduling module through the data interface. The intelligent agent scheduling module receives the pure natural language text sequence and prepares to execute the subsequent requirement parsing steps.
[0036] S2: Requirements analysis and keyword decomposition. The agent scheduling module starts and calls the large language model parsing module. The large language model parsing module receives the pure natural language text sequence. The large language model parsing module uses prompt word engineering technology to construct requirement analysis prompt words. The requirement analysis prompt words guide the large language model to perform structured semantic analysis on the pure natural language text sequence. The large language model parsing module extracts the core requirement information set.
[0037] The core requirement information set output by the large language model parsing module includes material type, core parameters, application scenarios, and constraints.
[0038] The large language model parsing module performs integrity checks. The large language model parsing module defines an integrity check function. When the integrity check function is equal to zero, the large language model parsing module determines that the core requirement information set is incomplete. When the intelligent agent scheduling module determines that the core requirement information set is incomplete, it initiates follow-up questions to the user through the e-commerce interactive interface. The intelligent agent scheduling module generates follow-up question statements.
[0039] The intelligent agent scheduling module receives user feedback on follow-up questions through the e-commerce interactive interface, and the large language model parsing module uses the feedback information to update the core requirement information set.
[0040] The large language model parsing module loads a pre-set industry keyword dictionary and uses it to map the core demand information set into a standardized search keyword set. The standardized search keyword set conforms to the search rules of the e-commerce search engine, and the intelligent agent scheduling module obtains the standardized search keyword set.
[0041] The large language model parsing module receives a sequence of pure natural language text, calls the internally stored prompt word engineering configuration library, loads the preset requirement parsing prompt word template from the prompt word engineering configuration library, and constructs requirement parsing prompt words based on the requirement parsing prompt word template.
[0042] The large language model parsing module inputs the pure natural language text sequence and requirement parsing prompt words into the large language model. Based on the instruction constraints of the requirement parsing prompt words, the large language model performs semantic analysis operations on the pure natural language text sequence, identifies key entities and attribute values from the pure natural language text sequence, and outputs a set of core requirement information.
[0043] The core requirement information set includes material type variables, core parameters and application scenario variables, and constraint variables. The large language model parsing module stores the core requirement information set in the runtime cache.
[0044] The large language model parsing module is equipped with an integrity check function. The integrity check function performs logical checks on the core requirement information set. When the core requirement information set is missing key performance parameters or the budget is unclear, the output logical check result is zero. A logical check result of zero indicates that the integrity check function is equal to zero.
[0045] When the integrity verification function equals zero, the agent scheduling module triggers a multi-round dialogue and follow-up questioning process, calls the large language model parsing module, identifies the missing information slots in the core requirement information set, and the agent scheduling module generates follow-up questions based on the missing information slots.
[0046] The agent scheduling module sends follow-up questions to the natural language interaction module, which displays the follow-up questions on the e-commerce interface. The natural language interaction module receives feedback information input by the user and transmits the feedback information to the large language model parsing module. The large language model parsing module uses the feedback information to update the core requirement information set.
[0047] The large language model parsing module executes the integrity verification function again on the updated core requirement information set. The large language model parsing module repeats the multi-round dialogue probing process until the integrity verification function outputs a logical verification result of 1.
[0048] The large language model parsing module internally stores a preset industry keyword dictionary, which contains the standard names, standard models, and standard parameter descriptions of power generation industry materials, and receives a set of core requirement information that has undergone integrity verification.
[0049] The large language model parsing module uses an industry keyword dictionary to perform standardized mapping processing on the core demand information set. The standardized mapping processing converts the non-standard natural language descriptions in the core demand information set into standardized vocabulary that conforms to the search rules of the e-commerce search engine, generating a standardized search keyword set. The intelligent agent scheduling module obtains the standardized search keyword set and transmits it to the e-commerce search module.
[0050] S3: E-commerce platform material retrieval. The e-commerce platform retrieval module has an inverted index database and a vector database built internally. The intelligent agent scheduling module inputs the standardized search keyword set into the e-commerce platform retrieval module. The e-commerce platform retrieval module integrates a search engine and a material database. The search engine receives the standardized search keyword set.
[0051] The search engine uses a standardized set of search keywords to perform matching searches in the materials database, extracting materials data that match the definitions of the standardized search keyword set. The e-commerce search module then generates a preliminary search result set. S4: Comprehensive analysis of large language models and personalized recommendations. The agent scheduling module transmits the preliminary search result set, core demand information set and user historical purchase records to the personalized recommendation module.
[0052] The personalized recommendation module calls a large language model and uses it to perform multi-dimensional comprehensive analysis on each material object in the preliminary search results set. The multi-dimensional comprehensive analysis includes matching degree analysis, historical procurement preference analysis, and comprehensive cost-effectiveness analysis.
[0053] The personalized recommendation module calculates the first result in the preliminary search results set. The overall recommendation score for each material item is denoted as [score]. The formula for calculating the overall recommendation score is as follows: ; In the formula, Indicates the first The overall recommendation score for each material item; Represents the weight coefficients for the dimensions of the matching analysis; The matching score is calculated based on the degree of matching between the model, specifications, and functional parameters of the material object and the core parameters and application scenarios in the core requirements information set. Indicates the weighting coefficients for the dimensions of historical purchasing preference analysis; This represents the historical preference score, which is calculated based on the correlation between the manufacturer and price of the material and the user's manufacturer preference and price sensitivity in their historical purchasing records. Indicates the weighting coefficients for the dimensions of the comprehensive cost-effectiveness analysis; The cost-effectiveness score is calculated based on a comprehensive analysis of the product's internal pricing data, quality certifications, and user reviews from the online marketplace.
[0054] The personalized recommendation module determines whether the comprehensive recommendation scores of the top-ranked material objects have insufficient differentiation. If the differentiation is insufficient, it generates a comparative analysis summary and initiates a clear preference dialogue through the agent scheduling module. Based on the final determined comprehensive recommendation scores, it sorts the preliminary search result set in descending order and generates a personalized recommendation list.
[0055] The personalized recommendation module uses a large language model to generate natural language recommendation reasons for the material objects in the personalized recommendation list. The recommendation reasons are generated based on the matching degree score, historical preference score, and cost-effectiveness score. The personalized recommendation list and recommendation reasons are then transmitted to the agent scheduling module.
[0056] The matching score is calculated by the personalized recommendation module using a large language model. The large language model calculates the first... The semantic similarity between the model, specifications, and functional parameters of a material object and the core parameters and application scenarios in the core requirement information set is mapped to a matching score.
[0057] Historical preference scores are calculated by the personalized recommendation module based on the user's historical purchasing records. The personalized recommendation module extracts the list of historical purchasing manufacturers and the historical transaction price range from the user's historical purchasing records to calculate the score. The personalized recommendation module maps the degree of fit between the manufacturer and internal price quote of each material item and the historical list of purchasing manufacturers and historical transaction price ranges to a historical preference score.
[0058] The cost-effectiveness score is determined by the personalized recommendation module based on the first... The cost-effectiveness score is calculated by comprehensively considering the internal price, quality certification level, and user rating of each material in the e-commerce platform. The personalized recommendation module then normalizes the internal price, quality certification level, and user rating data and performs a weighted sum to generate the cost-effectiveness score.
[0059] The personalized recommendation module sorts the material objects in the preliminary search result set in descending order based on the comprehensive recommendation score, and extracts the first-ranked material object at the top of the sort list and the second-ranked material object at the next position in the sort list.
[0060] The personalized recommendation module calculates the score differentiation between the first-ranked material object and the second-ranked material object, compares the score differentiation with a preset differentiation threshold, and determines that the differentiation is insufficient when the score differentiation is less than the preset differentiation threshold. When the differentiation is insufficient, the interactive preference clarification process is triggered.
[0061] The personalized recommendation module calls the large language model to generate a comparative analysis summary. The agent scheduling module displays the comparative analysis summary to the user through the e-commerce interface and receives the user's preference feedback information on the comparative analysis summary. The personalized recommendation module uses the preference feedback information to adjust the weight coefficient vector.
[0062] The personalized recommendation module recalculates the comprehensive recommendation score for all material objects in the preliminary search result set using the adjusted weight coefficient vector, and updates the personalized recommendation list based on the recalculated comprehensive recommendation score.
[0063] The personalized recommendation module sets a length threshold for the recommendation list. Based on the comprehensive recommendation score, it extracts the material objects that rank at the top of the recommendation list length threshold from the initial search results set. The extracted material objects are recorded as recommended material objects.
[0064] The personalized recommendation module calls the large language model to generate natural language recommendation reasons for each recommended product object. The recommended product object and the natural language recommendation reasons are combined to construct a personalized recommendation list. The personalized recommendation list is transmitted to the intelligent agent scheduling module, which controls the e-commerce interactive interface to display the personalized recommendation list to the user.
[0065] S5: Automatic price comparison and difference reminder. The intelligent agent scheduling module monitors the user's operation behavior on the e-commerce interactive interface, identifies the target material object selected by the user from the personalized recommendation list, and transmits the target material object to the automatic price comparison module.
[0066] The automatic price comparison module obtains the current sales price of the target material in the e-commerce platform, calls the web crawler program to access the external Internet e-commerce platform, collects the external market reference price that is consistent with the model and specifications of the target material, accesses the historical transaction database, and extracts the historical average transaction price of the target material within a preset time period.
[0067] The automatic price comparison module stores preset price anomaly thresholds, calculates the price deviation index of the target material object, compares the price deviation index with the price anomaly threshold, and determines that there is a premium risk when the price deviation index is greater than the price anomaly threshold. The automatic price comparison module then generates a difference reminder warning message.
[0068] The automatic price comparison module transmits the difference alert information to the intelligent agent scheduling module, which then controls the e-commerce interface to display the difference alert information in the details page area of the target material.
[0069] The automatic price comparison module is equipped with a web crawler program and a data cleaning algorithm. The web crawler program accesses a preset internet e-commerce platform data source and performs a price grabbing operation in the internet e-commerce platform data source based on the material name and model specifications of the target material object, thereby obtaining multiple external original price data, which constitute an external original price set.
[0070] The automatic price comparison module performs statistical analysis on the external original price quotation set, calculating the arithmetic mean and standard deviation of the external original price quotation set. Outlier filtering conditions are then constructed using these conditions. External original price quotation data that does not meet the outlier filtering conditions are removed from the external original price quotation set, and the external original price quotation data that meets the outlier filtering conditions is retained as valid price quotation data. The construction process of the valid price quotation set is expressed by the formula: ; In the formula, This represents the set of valid quotes after cleaning. This represents a single external raw quote data within a set of external raw quotes; Represents the set of external original quotes; Represents the arithmetic mean of the original set of external quotes; This represents the preset anomaly removal coefficient; This represents the standard deviation of the external original quote set.
[0071] The automatic price comparison module calculates the external market reference price based on the set of valid quotations and stores the external market reference price in the runtime cache for use in calculating the price deviation index.
[0072] The automatic price comparison module continuously monitors the transaction status of the target material object. When it receives the transaction completion signal, it obtains the transaction price of the target material object in this transaction and reads the previous historical average transaction price of the target material object from the historical transaction database.
[0073] The automatic price comparison module obtains the current time and the last update time, calculates the time interval between the current time and the last update time, and uses this time interval to calculate the time decay weight coefficient. The formula for calculating the time decay weight coefficient is as follows: ; In the formula, This represents the time decay weighting coefficient, which is used to measure the degree of influence of the transaction price on historical trends. The base of the natural logarithm; This represents the preset time decay constant, which is used to control the rate at which historical data is forgotten. Indicates a time interval.
[0074] The automatic price comparison module uses the time decay weighting factor, the current transaction price, and the previous historical average transaction price to calculate the updated historical average transaction price. The calculation process for the updated historical average transaction price is expressed by the formula: ; In the formula, This indicates the updated historical average transaction price; This represents the time decay weighting coefficient; This indicates the transaction price. This represents the historical average transaction price.
[0075] The automatic price comparison module writes the updated historical average transaction price into the historical transaction database, overwrites the previous historical average transaction price with the updated historical average transaction price, and records the current time as the new last update time.
[0076] The automatic price comparison module reads the internal e-commerce price, external market reference price, and historical average transaction price of the target material from the runtime cache. It calculates the minimum value between the external market reference price and the historical average transaction price, and sets this minimum value as the price comparison benchmark. Based on the internal e-commerce price and the price comparison benchmark, the automatic price comparison module calculates a price deviation index. The formula for calculating the price deviation index is as follows: ; In the formula, Indicator of price deviation; This indicates the internal pricing information of the online marketplace. Indicates the external market reference price; This represents the historical average transaction price; This represents the function for selecting the minimum value.
[0077] The automatic price comparison module obtains a preset price anomaly threshold, performs a numerical comparison operation between the price deviation index and the price anomaly threshold, and determines whether the price deviation index is greater than the price anomaly threshold. If the price deviation index is greater than the price anomaly threshold, it is determined that the target material object has a premium risk. When a premium risk is determined, an alarm trigger signal is sent to the intelligent agent scheduling module.
[0078] S6: Information push. The intelligent agent scheduling module receives a personalized recommendation list from the personalized recommendation module and a difference reminder and warning message from the automatic price comparison module. The intelligent agent scheduling module constructs the final display information set based on the price deviation index and the price anomaly threshold.
[0079] The intelligent agent scheduling module converts the final information set into front-end visual instructions, which are then sent to the e-commerce interactive interface. The e-commerce interactive interface parses the front-end visual instructions and renders a list of material cards and a price risk pop-up on the user's terminal screen.
[0080] The implementation process of this invention will be further explained in detail with reference to actual application process. This embodiment takes the purchase of a high-pressure shut-off valve for a 300MW unit boiler by a power generation company as an example to explain the implementation process of the invention in detail.
[0081] The user enters text through the e-commerce interactive interface: "We need to purchase a batch of high-pressure shut-off valves for a 300MW unit boiler. The requirements are: pressure resistance ≥10MPa, temperature ≤540℃, quantity 10 units, budget between 50,000 and 80,000 yuan." The natural language interaction module transmits this text sequence to the intelligent agent scheduling module.
[0082] The intelligent agent scheduling module triggers the large language model parsing module. The large language model parsing module performs semantic analysis on the user's request using preset prompts, extracting core information: the material type is a high-pressure shut-off valve, the application scenario is a 300MW unit boiler, the performance parameters are pressure resistance ≥10MPa, temperature ≤540℃, quantity is 10 units, and the budget is 50,000-80,000 yuan. This core information is broken down into standardized search keywords: 300MW unit boiler high-pressure shut-off valve, pressure resistance ≥10MPa, temperature ≤540℃. Because the user's request information is complete and the integrity check passes, no follow-up questions are needed.
[0083] The intelligent agent scheduling module inputs standardized keywords into the e-commerce search engine. The search engine matches relevant materials in the materials database and obtains a preliminary search result set, containing five matching material products. Each product's name, model, manufacturer (e.g., manufacturer A, manufacturer B), price (48,000-75,000 RMB), inventory, quality certification (e.g., ISO9001), user reviews, and historical transaction records are recorded. Among these, the user previously purchased similar valves from manufacturer A and gave them positive feedback.
[0084] The intelligent agent scheduling module inputs the preliminary search results set and the user's historical purchasing records into the personalized recommendation module for comprehensive analysis: (1) Matching degree analysis: All 5 products meet the performance parameters and delivery time requirements; (2) Historical purchasing preference analysis: The user has purchased products from manufacturer A and has given positive reviews, so manufacturer A is the preferred choice; (3) Cost-effectiveness analysis: Manufacturer A's price of 52,000 yuan is within the budget range, and user reviews are better than other manufacturers, making it the most cost-effective. No further questioning of users is needed during the analysis. A personalized recommendation list is finally generated, sorted by matching degree: High-pressure shut-off valve from manufacturer A (Recommended because: meets performance requirements, within budget, and has a good track record of purchases). Manufacturer C's high-pressure shut-off valve (recommended because: performance meets standards, price is relatively low); detailed information on each product is also included.
[0085] The intelligent agent scheduling module initiates automatic price comparison, extracting the core attributes of each material in the recommended list (name: high-pressure shut-off valve, model: compatible with 300MW units, pressure resistance ≥10MPa, temperature ≤540℃), and obtaining price information for similar materials from external e-commerce platforms via API or web crawler. After filtering, valid price information is obtained: Manufacturer A's product of the same model is priced at 50,000 yuan on external platforms and 52,000 yuan on e-commerce platforms, with a difference of 3.8%, which does not exceed the preset threshold of 5%. Manufacturer C's product of the same model is priced at 45,000 yuan on external platforms and 46,000 yuan on e-commerce platforms, with a price difference of 2.2%. At this time, there is no price difference exceeding the threshold, so no price difference warning information is generated.
[0086] The information push module pushes a personalized recommendation list to the user interface, where users can directly view the recommendation results and detailed information to complete subsequent purchasing operations.
[0087] The intelligent agent scheduling module monitors the e-commerce interface in real time, captures natural language demand descriptions, synchronously obtains the user identity of the current user, and records the processing timestamp of the demand input. Based on the natural language demand description, user identity, and processing timestamp, the intelligent agent scheduling module constructs an initial demand request vector, stores the initial demand request vector in the local data buffer, and transmits the initial demand request vector to the large language model parsing module to trigger subsequent data processing.
[0088] To ensure the performance of the large language model parsing module, before executing step S1, this invention pre-executes the large language model selection process, configures the large language model parsing module, constructs a candidate model set through the large language model parsing module, the candidate model set contains multiple pre-trained Transformer architecture basic models, and performs performance tests on the candidate models in the candidate model set on a preset e-commerce domain validation dataset to obtain the intent recognition accuracy and entity extraction F1 score of the candidate models.
[0089] The large language model parsing module deploys candidate models in a standard hardware environment to perform stress tests and monitors the average inference latency and memory usage of the candidate models.
[0090] The large language model parsing module calculates the model fit score based on intent recognition accuracy, entity extraction F1 score, average inference latency, and GPU memory usage. The formula for calculating the model fit score is as follows: ; In the formula, This represents the model fit score used to measure the overall effectiveness of candidate models; This represents the intent recognition accuracy of the candidate model in the intent recognition task; Weighting coefficients representing the accuracy of intent recognition; This represents the F1 score of the candidate model in entity extraction in the named entity recognition task. The weighting coefficients representing the F1 score of an entity; This represents the average inference latency of the candidate model when processing a text sequence of unit length. The penalty weighting coefficient represents the average inference delay; This indicates the amount of video memory required to run the candidate model. This represents the penalty weighting coefficient for video memory usage.
[0091] The large language model parsing module traverses the candidate model set, identifies the candidate model with the highest model fit score as the target large language model, and loads the target large language model into the large language model parsing module and the personalized recommendation module to support the natural language processing task.
[0092] Regarding the collaborative working mechanism of each functional module, this invention adopts a multi-agent collaboration mode that combines centralized scheduling and distributed execution. The agent scheduling module serves as the control core, while the large language model parsing module, e-commerce retrieval module, personalized recommendation module, and automatic price comparison module serve as functional execution units. The agent scheduling module maintains a global session state vector, which is used to record the processing progress of the current task and the intermediate data context.
[0093] The agent scheduling module calculates the next action based on the global session state vector, sends task instructions to specific functional execution units according to the next action, the functional execution units execute the task instructions and return the execution result data, and update the global session state vector using the execution result data.
[0094] The intelligent agent scheduling module cyclically performs state monitoring, action decision-making, and state update operations until the task termination conditions are met. It interacts with the large language model parsing module, e-commerce retrieval module, personalized recommendation module, and automatic price comparison module through a standardized message bus.
[0095] In the requirement parsing stage of step S2, the large language model parsing module reads the initial requirement request vector from the local data buffer, parses the initial requirement request vector and obtains the natural language requirement description, calls the large language model to perform semantic analysis and entity extraction on the natural language requirement description, and generates the core requirement information set.
[0096] The large language model parsing module uses the large language model to evaluate the completeness of the core requirement information set, calculates the requirement clarity index, and compares the requirement clarity index with a preset clarity threshold. When the requirement clarity index is greater than or equal to the clarity threshold, the core requirement information set is transmitted to the e-commerce retrieval module. When the requirement clarity index is less than the clarity threshold, the interactive follow-up questioning mechanism is activated. The interactive follow-up questioning mechanism generates query text for the missing information and feeds it back to the e-commerce interactive interface.
[0097] During the retrieval process in step S3, the e-commerce retrieval module receives a core demand information set, constructs keyword retrieval query statements and vector retrieval query vectors based on the core demand information set, performs an exact match retrieval operation based on the keyword retrieval query statements in the inverted index database to obtain a list of keyword retrieval results, and performs a semantic similarity retrieval operation based on the vector retrieval query vectors in the vector database to obtain a list of vector retrieval results.
[0098] The e-commerce search module uses a reciprocal ranking fusion algorithm to perform a result fusion operation on the keyword search result list and the vector search result list, generating a preliminary search result set. The fusion ranking score of each material object in the preliminary search result set is expressed by the formula: ; In the formula, This indicates the fusion ranking score of the material object during the fusion process; This represents the set of search results that participated in the fusion process. This represents a list of keyword search results. This represents a list of vector search results. This represents the preset smoothing constant; This indicates the ranking position of the material object in the corresponding search results list.
[0099] During the external price acquisition process in step S5, the automatic price comparison module maintains a set of target external data source addresses. This set contains Uniform Resource Locators (URLs) for product detail pages from multiple mainstream e-commerce platforms. The automatic price comparison module is configured with a web crawler, which uses a dynamic proxy IP pool to send Hypertext Transfer Protocol (HTTP) requests to the target external data source address set.
[0100] The automatic price comparison module receives raw webpage response data from a set of target external data source addresses, performs document object model parsing on the raw webpage response data, extracts price value nodes and product description text nodes based on a predefined page structure template, performs regularization cleaning on the price value nodes and product description text nodes to remove currency symbols, thousands separators and Hypertext Markup Language tags, and obtains standardized external market reference prices and external product description information.
[0101] The automatic price comparison module calculates the data quality score of the collected data. The formula for calculating the data quality score is as follows: ; In the formula, This represents a data quality score that measures the completeness and usability of the collected data; This represents the price effectiveness weighting coefficient; This represents the price validity indicator function, which takes a value of 1 when the external market reference price is a positive real number, and a value of 0 otherwise. This represents the weighting coefficient for the richness of descriptive information; This function represents the text length calculation function. This indicates external product description information; This represents the preset standard description length constant.
[0102] The automatic price comparison module compares the data quality score with the preset quality qualification threshold. When the data quality score is greater than the preset quality qualification threshold, the automatic price comparison module stores the external market reference price in the associated database for subsequent price comparison calculations.
[0103] The personalized recommendation module receives the preliminary search result set, calls the large language model to analyze the attribute information of the material objects in the preliminary search result set, calculates the comprehensive recommendation score for each material object, sorts the preliminary search result set in descending order according to the comprehensive recommendation score, calculates the score discrimination between the first and second ranked material objects, triggers the interactive preference clarification process when the score discrimination is less than a preset threshold, adjusts the weight coefficients and updates the comprehensive recommendation score using the interactive preference clarification process, and selects recommended material objects from the preliminary search result set based on the finally determined comprehensive recommendation score, generates a personalized recommendation list containing recommended material objects and recommendation reasons, and transmits the personalized recommendation list to the agent scheduling module.
[0104] The intelligent agent scheduling module transmits the target material object to the automatic price comparison module. The automatic price comparison module obtains the internal price of the target material object from the e-commerce platform. Based on the web crawler program and data cleaning algorithm, the automatic price comparison module calculates the external market reference price of the target material object and reads the historical average transaction price of the target material object from the historical transaction database.
[0105] The automatic price comparison module sets the smaller value between the external market reference price and the historical average transaction price as the price comparison benchmark, and calculates the price deviation index using the internal price of the e-commerce platform and the price comparison benchmark.
[0106] The automatic price comparison module calls the preset price anomaly threshold, compares the price deviation index with the price anomaly threshold, and determines that there is a premium risk when the price deviation index is greater than the price anomaly threshold. It then generates a difference reminder warning message and feeds the difference reminder warning message back to the intelligent agent scheduling module. The intelligent agent scheduling module displays the difference reminder warning message on the e-commerce interactive interface.
[0107] The intelligent agent scheduling module receives a personalized recommendation list from the personalized recommendation module and a difference alert from the automatic price comparison module. It also calls up the price deviation index and the price anomaly threshold. Based on the comparison results of the price deviation index and the price anomaly threshold, the intelligent agent scheduling module constructs the final display information set.
[0108] The intelligent agent scheduling module converts the final information set into front-end visual instructions and sends these instructions to the e-commerce interactive interface. The e-commerce interactive interface parses the instructions, performs rendering on the user's terminal screen, and displays a list of material cards. When the final information set contains discrepancy alerts, the e-commerce interactive interface simultaneously displays a price risk pop-up.
Claims
1. A method for automated querying, retrieval, and analysis of goods in an e-commerce platform, characterized in that, Includes the following steps: S1. Natural language requirement input: The natural language interaction module receives the material procurement requirement information input by the user, encapsulates the material procurement requirement information into a pure natural language text sequence, and transmits the pure natural language text sequence to the intelligent agent scheduling module through the data interface. S2. Demand analysis and keyword decomposition: The intelligent agent scheduling module triggers the large language model parsing module. The large language model parsing module performs structured semantic analysis on the pure natural language text sequence, extracts the core demand information set, and uses an industry keyword dictionary to map the core demand information set into a standardized search keyword set. S3, E-commerce Mall Material Retrieval: The intelligent agent scheduling module inputs a standardized set of search keywords into the e-commerce mall retrieval module, which then uses the search engine to generate a preliminary search result set. S4. Comprehensive analysis and personalized recommendation of large language model: The intelligent agent scheduling module transmits the preliminary search result set, core demand information set and user historical procurement records to the personalized recommendation module. The personalized recommendation module calculates the comprehensive recommendation score of each material object in the preliminary search result set and generates a personalized recommendation list and recommendation reasons based on the comprehensive recommendation score. S5. Automatic price comparison and difference alert: The intelligent agent scheduling module identifies the target material object selected by the user from the personalized recommendation list. The intelligent agent scheduling module transmits the target material object to the automatic price comparison module. The automatic price comparison module calculates the price deviation index of the target material object. When the price deviation index is greater than the price abnormality threshold, a difference alert warning message is generated. S6. Information push: The intelligent agent scheduling module converts personalized recommendation lists, recommendation reasons, and difference reminders and warnings into front-end visual commands. The e-commerce interactive interface parses the front-end visual commands and renders the material card list and price risk pop-up.
2. The automated query, retrieval, and analysis method for e-commerce goods according to claim 1, characterized in that, In step S2, the large language model parsing module performs a completeness check on the core requirement information set. When the core requirement information set is missing key performance parameters or the budget is unclear, the agent scheduling module triggers a multi-round dialogue and questioning process. The intelligent agent scheduling module generates follow-up questions based on the missing information slots. The intelligent agent scheduling module receives user feedback on the follow-up questions through the e-commerce interactive interface. The large language model parsing module updates the core requirement information set using the feedback information. The large language model parsing module performs integrity verification on the updated core requirement information set again until the integrity verification passes.
3. The automated query, retrieval, and analysis method for e-commerce goods according to claim 1, characterized in that, In step S3, the e-commerce retrieval module internally constructs an inverted index database and a vector database; The e-commerce search module performs an exact match search operation based on keyword search query statements in the inverted index database to obtain a list of keyword search results. The e-commerce search module also performs a semantic similarity search operation based on vector search query vectors in the vector database to obtain a list of vector search results. The e-commerce search module uses a reciprocal ranking fusion algorithm to perform a result fusion operation on the keyword search result list and the vector search result list to generate the preliminary search result set.
4. The automated query, retrieval, and analysis method for e-commerce goods according to claim 1, characterized in that, In step S4, the personalized recommendation module calls a large language model to perform a multi-dimensional comprehensive analysis on each material object in the preliminary search result set. The multi-dimensional comprehensive analysis includes a matching degree analysis dimension, a historical procurement preference analysis dimension, and a comprehensive cost-effectiveness analysis dimension. The personalized recommendation module calculates a matching score based on the matching score analysis dimension, a historical preference score based on the historical purchasing preference analysis dimension, and a cost-effectiveness score based on the comprehensive cost-effectiveness analysis dimension. The personalized recommendation module calculates the comprehensive recommendation score based on the matching score, the historical preference score, and the cost-effectiveness score.
5. The automated query, retrieval, and analysis method for e-commerce goods according to claim 4, characterized in that, In step S4, the personalized recommendation module sorts the preliminary search result set in descending order based on the comprehensive recommendation score, and extracts the first sorted material object at the top of the sort list and the second sorted material object at the second position in the sort list. The personalized recommendation module calculates the score differentiation between the first-ranked material object and the second-ranked material object, and triggers an interactive preference clarification process when the score differentiation is less than a preset differentiation threshold. The personalized recommendation module calls the large language model to generate a comparative analysis summary. The intelligent agent scheduling module receives the user's preference feedback information on the comparative analysis summary through the e-commerce interactive interface. The personalized recommendation module uses the preference feedback information to adjust the weight coefficient vector and update the comprehensive recommendation score.
6. The automated query, retrieval, and analysis method for e-commerce goods according to claim 1, characterized in that, In step S5, the automatic price comparison module accesses a preset internet e-commerce platform data source through a web crawler program to obtain multiple external original price data, which constitute an external original price set. The automatic price comparison module performs statistical analysis on the external original price quotation set, calculates the arithmetic mean and standard deviation of the external original price quotation set, constructs outlier filtering conditions using the arithmetic mean and standard deviation, removes external original price quotation data that does not meet the outlier filtering conditions from the external original price quotation set, retains external original price quotation data that meets the outlier filtering conditions as valid price quotation data, and calculates the external market reference price based on the valid price quotation data.
7. The automated query, retrieval, and analysis method for e-commerce goods according to claim 1, characterized in that, In step S5, the automatic price comparison module continuously monitors the transaction status of the target material object, obtains the transaction price of the target material object in this transaction when it receives the transaction completion signal, and reads the previous historical average transaction price of the target material object from the historical transaction database. The automatic price comparison module obtains the current time and the last update time, calculates the time interval between the current time and the last update time, and uses the time interval to calculate the time decay weight coefficient. The automatic price comparison module uses the time decay weight coefficient, the transaction price of this transaction and the previous historical average transaction price to calculate the updated historical average transaction price, and writes the updated historical average transaction price into the historical transaction database.
8. The automated query, retrieval, and analysis method for e-commerce goods according to claim 1, characterized in that, In step S5, the automatic price comparison module reads the internal price quote of the target material object in the e-commerce marketplace, the external market reference price, and the historical average transaction price from the running cache. The automatic price comparison module calculates the minimum value between the external market reference price and the historical average transaction price, and sets the minimum value as the price comparison benchmark. The automatic price comparison module calculates the price deviation index based on the internal price of the e-commerce platform and the price comparison benchmark.
9. The automated query, retrieval, and analysis method for e-commerce goods according to claim 1, characterized in that, Before step S1, the large language model selection process is performed in advance. A candidate model set is constructed through the large language model parsing module. The performance of the candidate models in the candidate model set is tested on a preset e-commerce domain validation dataset to obtain the intent recognition accuracy and entity extraction F1 score of the candidate models. The large language model parsing module monitors the average inference latency and memory usage of candidate models in a standard hardware environment. Based on the intent recognition accuracy, entity extraction F1 score, average inference latency, and memory usage, it calculates the model fit score and loads the candidate model with the highest model fit score into the large language model parsing module.
10. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the electronic mall material automated query, retrieval, and analysis method according to any one of claims 1 to 9.