Article recommendation method and device, computer equipment and storage medium

Abstract

The invention relates to an article recommendation method and device, computer equipment and a storage medium. The method comprises the following steps: obtaining article structural features of a sample article input by a business party, performing feature extraction on article detail information of the sample article through a multi-modal large model, determining unstructured features of the sample article, and determining the unstructured features of the sample article; and correspondingly storing the article titles of the sample articles, the article identifiers of the sample articles, the article structural features and the non-structural features in a dynamic article feature database, obtaining an article recommendation request sent by a user, and determining article recommendation information through an article recommendation model according to the dynamic article feature database and the article recommendation request. According to the scheme, the accuracy and diversity of the article recommendation result generated by the intelligent shopping guide system can be improved, the article recommendation request can be deeply interpreted in combination with the dynamic article feature database, and the intelligent degree of the intelligent shopping guide system is improved.

Description

Technical Field

[0001] This application relates to the field of smart device technology, and in particular to a method, apparatus, computer device, and storage medium for recommending items. Background Technology

[0002] With the deep integration of artificial intelligence and the retail industry, intelligent shopping guide systems have become a core tool for enhancing the consumer experience and optimizing retail efficiency. Current intelligent shopping guide systems primarily rely on three major technical modules: intent recognition, item recommendation, and interactive services. For example, existing intelligent shopping guide systems can respond to user requests through basic intent classification and recommend items using collaborative filtering or vector recall. However, existing technologies still have significant bottlenecks: item recommendations often rely on a single recall strategy, making it difficult to cover structured needs such as item brands, attributes, and keywords, resulting in low relevance and insufficient diversity in the recall results. Furthermore, existing solutions often simply list items without in-depth analysis based on item feature databases, failing to provide personalized recommendation reasons like a human shopping guide, resulting in a low level of intelligence. Therefore, how to improve the intelligence level of intelligent shopping guide systems while simultaneously increasing the accuracy and diversity of the item recommendation results generated by these systems is a problem that needs to be solved. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, apparatus, computer equipment, and storage medium for recommending items that can improve the intelligence level of intelligent shopping guide systems and enhance the accuracy and diversity of item recommendation results generated by these systems, in order to address the aforementioned technical problems.

[0004] Firstly, this application provides a method for recommending items, the method comprising:

[0005] Obtain the structural features of the sample items entered by the business party; the structural features of the items include the item brand information, item series information, item material information, target audience of the item, core functions of the item, item style and main technical features of the item;

[0006] Using a multimodal large model, features are extracted from the item details information of the sample items to determine their unstructured features; the item details information includes item images and item description text.

[0007] The item title, item identifier, structured features, and unstructured features of the sample item are stored in a dynamic item feature database.

[0008] The system obtains item recommendation requests from users and determines item recommendation information based on the dynamic item feature database and the item recommendation request using an item recommendation model. The item recommendation model is a pre-trained large-scale artificial intelligence model.

[0009] In one embodiment, the above-mentioned item recommendation method further includes:

[0010] The AI ​​model is trained using data specific to the item recommendation scenario to determine the pre-trained model.

[0011] The model parameters of the pre-trained model are fine-tuned using a low-rank adapter to determine the item recommendation model.

[0012] In one embodiment, a multimodal large model is used to extract features from the item details information of the sample item to determine the unstructured features of the sample item, including:

[0013] Using the multimodal large model, the item color and main usage scenarios are extracted from the item images in the item details information;

[0014] Using the multimodal large model, text summaries and item features are extracted from the item description text of the item details information, and the item features are vectorized to determine the feature vectors of the item features;

[0015] The color of the item, the main usage scenarios of the item, the text summary of the item description text, and the feature vector of the item features are used as unstructured features of the sample item.

[0016] In one embodiment, a user-submitted item recommendation request is obtained, and item recommendation information is determined using an item recommendation model based on the dynamic item feature database and the item recommendation request, including:

[0017] Obtain the item recommendation request sent by the user, and extract keywords from the item recommendation request using the item recommendation model to determine the request keywords;

[0018] The query vector is determined by encoding the request keywords using a vector model.

[0019] Based on the query vector and the preset recall quantity, an approximate nearest neighbor search is performed on the vector database to determine candidate vectors from the dense vectors stored in the vector database, and the item identifier of the sample item corresponding to the candidate vector is determined as the first candidate identifier; the vector database stores the item identifier and sample vector of the sample item.

[0020] Based on the item recommendation request, and using the request keywords, the recall quantity, and the pre-built target index, a matching query is performed on the dynamic item feature database to determine a second candidate identifier from the item identifiers of the sample items in the dynamic item feature database.

[0021] Based on the requested keywords, determine the item brand dimension data, item attribute dimension data, and item keyword dimension data, and construct a structured query vector based on the item brand dimension data, the item attribute dimension data, and the item keyword dimension data;

[0022] The dynamic item feature database is searched based on the recall quantity and structured query vector, and a third candidate identifier is determined from the item identifiers of the sample items in the dynamic item feature database.

[0023] The first candidate identifier, the second candidate identifier, and the third candidate identifier are integrated to determine the identifier to be evaluated, and the comprehensive score of the item corresponding to the identifier to be evaluated is determined based on the recall path weight.

[0024] Based on the comprehensive score of the items, a target identifier is determined from the identifiers to be evaluated, and item recommendation information is generated based on the target identifier and the dynamic item feature database through the item recommendation model.

[0025] In one embodiment, a target identifier is determined from the identifiers to be evaluated based on the overall score of the items, and item recommendation information is generated based on the target identifier and the dynamic item feature database using an item recommendation model, including:

[0026] Based on the overall score of the items, the target identifier is determined from the identifiers to be evaluated;

[0027] The item metadata information is determined based on the target identifier, and the structured features and unstructured features corresponding to the item identifier to be selected are extracted from the dynamic item feature database based on the target identifier. The item metadata information includes the target identifier, the item title corresponding to the target identifier, and the item comprehensive score corresponding to the target identifier.

[0028] Based on the prompt word generation template, the item recommendation prompt words are generated according to the item metadata information corresponding to the target identifier, the structured features corresponding to the target identifier, and the unstructured features corresponding to the item identifier to be selected.

[0029] The item recommendation model determines item recommendation information based on the item recommendation prompts.

[0030] In one embodiment, determining item recommendation information based on the item recommendation prompts using an item recommendation model includes:

[0031] Obtain historical dialogue information between users and the intelligent shopping guide agent;

[0032] The item recommendation model determines item recommendation information based on the item recommendation prompts and the historical dialogue information.

[0033] In one embodiment, the above-mentioned item recommendation method further includes:

[0034] The item description text of the sample items is encoded into dense vectors using a vector model;

[0035] The dense vectors and the item identifiers of the sample items are stored in the vector database.

[0036] Secondly, this application also provides an item recommendation device, the device comprising:

[0037] The item feature acquisition module is used to acquire the item structure features of sample items entered by the business party. The item structure features include item brand information, item series information, item material information, target audience, core functions, style, and main technical features.

[0038] The feature extraction module is used to extract features from the item details information of the sample items using a multimodal large model, and to determine the unstructured features of the sample items; the item details information includes item images and item description text;

[0039] The data storage module is used to store the item title, item identifier, structured features, and unstructured features of the sample item in a dynamic item feature database.

[0040] The item recommendation information determination module is used to obtain the item recommendation request issued by the user, and determine the item recommendation information based on the dynamic item feature database and the item recommendation request through the item recommendation model; the item recommendation model is a pre-trained large artificial intelligence model.

[0041] Thirdly, this application also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0042] Obtain the structural features of the sample items entered by the business party; the structural features of the items include the item brand information, item series information, item material information, target audience of the item, core functions of the item, item style and main technical features of the item;

[0043] Using a multimodal large model, features are extracted from the item details information of the sample items to determine their unstructured features; the item details information includes item images and item description text.

[0044] The item title, item identifier, structured features, and unstructured features of the sample item are stored in a dynamic item feature database.

[0045] The system obtains item recommendation requests from users and determines item recommendation information based on the dynamic item feature database and the item recommendation request using an item recommendation model. The item recommendation model is a pre-trained large-scale artificial intelligence model.

[0046] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0047] Obtain the structural features of the sample items entered by the business party; the structural features of the items include the item brand information, item series information, item material information, target audience of the item, core functions of the item, item style and main technical features of the item;

[0048] Using a multimodal large model, features are extracted from the item details information of the sample items to determine their unstructured features; the item details information includes item images and item description text.

[0049] The item title, item identifier, structured features, and unstructured features of the sample item are stored in a dynamic item feature database.

[0050] The system obtains item recommendation requests from users and determines item recommendation information based on the dynamic item feature database and the item recommendation request using an item recommendation model. The item recommendation model is a pre-trained large-scale artificial intelligence model.

[0051] The aforementioned item recommendation method, apparatus, computer equipment, and storage medium acquire the structured features of sample items entered by the business party. Through a multimodal large model, feature extraction is performed on the item details information of the sample items to determine their unstructured features. The item title, item identifier, structured features, and unstructured features of the sample items are stored correspondingly in a dynamic item feature database. The system then acquires user-initiated item recommendation requests and, through the item recommendation model, determines item recommendation information based on the dynamic item feature database and the item recommendation request. This solution addresses the problem that current intelligent shopping guide systems, when recommending items to users based on user-input item recommendation requests, often rely on a single recall strategy, making it difficult to cover structured needs such as item brands, attributes, and keywords, resulting in low relevance and insufficient diversity in the recall results. It also solves the problem that existing solutions often simply list items, lacking in-depth interpretation combined with an item feature database, failing to provide personalized recommendation reasons like a human shopping guide, and exhibiting a low level of intelligence. By constructing structured features of sample items based on item information entered by business users, and constructing unstructured features of sample items through a multimodal large model, a dynamic item feature database is built based on the structured features, unstructured features, item identifiers, and item titles of the sample items. Through an item recommendation model, based on the dynamic item feature database, the item recommendation requests issued by users are analyzed to determine the item recommendation information. This can improve the accuracy and diversity of item recommendation results generated by the intelligent shopping guide system. At the same time, it can combine the dynamic item feature database to deeply interpret item recommendation requests, thereby improving the intelligence level of the intelligent shopping guide system. Attached Figure Description

[0052] Figure 1 This is a diagram illustrating the application environment of the item recommendation method in one embodiment;

[0053] Figure 2 This is a flowchart illustrating an item recommendation method in one embodiment;

[0054] Figure 3 This is a flowchart illustrating the item recommendation method in another embodiment;

[0055] Figure 4 This is a flowchart illustrating the item recommendation method in another embodiment;

[0056] Figure 5 This is a structural block diagram of an item recommendation device in one embodiment;

[0057] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] The item recommendation method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. Server 104 acquires the structured features of sample items entered by the business party; the structured features include item brand information, item series information, item material information, target audience, core functions, style, and main technical characteristics; through a multimodal large model, it extracts features from the item details information of the sample items to determine the unstructured features of the sample items; the item details information includes item images and item description text; the structured features and unstructured features are stored correspondingly in a dynamic item feature database; it acquires item recommendation requests from users, and through an item recommendation model, determines item recommendation information based on the dynamic item feature database and the item recommendation request; the item recommendation model is a pre-trained artificial intelligence large model. The item recommendation information is sent to terminal 102 via the communication network. The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle systems. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0060] In one embodiment, such as Figure 2 As shown, a product recommendation method is provided, which is applied to an intelligent shopping guide system. In this embodiment, the method includes the following steps:

[0061] S210. Obtain the item structure characteristics of the sample items entered by the business party.

[0062] The structural characteristics of an item include its brand information, series information, material information, target audience, core functions, style, and main technical features.

[0063] The main technical characteristics of an item include, but are not limited to: item technical parameters, item performance, and item specifications. Item technical parameters describe the item's technical specifications and performance indicators, such as dimensions, weight, and material composition. Item performance refers to the item's functional capabilities and effects, such as operational efficiency and stability. Item specifications refer to the item's design details and manufacturing standards, such as model, structure, and process requirements. Item series information is a standardized data set used to describe the item's classification, attributes, and transaction characteristics. For example, item series information may include standardized item units and inventory units. A standardized item unit is the smallest unit that aggregates information on similar items, containing the common characteristics of all configuration combinations of the item. The item's core function refers to the basic service provided by the item or the core needs it addresses. The item's target audience refers to the target group most suitable for using the item. The business party can be the merchant corresponding to the item. The sample item can be the commodity that the business party needs to transact with.

[0064] Specifically, business users can input the structural characteristics of sample items into the intelligent shopping guide system through terminal devices.

[0065] S220. Using a multimodal large model, feature extraction is performed on the item details information of the sample items to determine the unstructured features of the sample items.

[0066] Item details include item images and item description text.

[0067] The unstructured features of the sample items include, but are not limited to: the item color, the main usage scenarios of the item, the item description text, the text summary of the item description text, and the feature vector of the item features.

[0068] Specifically, the multimodal large model extracts image and text information from the sample items to determine the item color, main usage scenarios, description text, text summary of description text, and feature vectors of item features. The item color, main usage scenarios, description text, text summary of description text, and feature vectors of item features are then used as unstructured features of the sample items.

[0069] For example, using a multimodal large model, feature extraction is performed on the item details information of the sample items to determine the unstructured features of the sample items, including:

[0070] Using a multimodal large model, the color and main usage scenarios of the item are extracted from the item images in the item details information of the sample item. The text summary and item features are extracted from the item description text in the item details information, and the item features are vectorized to determine the feature vectors of the item features. The item color, main usage scenarios, text summary of item description text, and feature vectors of item features are used as unstructured features of the sample item.

[0071] Vectorization of item features is equivalent to encoding item features.

[0072] The above scheme, through a multimodal large model, determines the item color, main usage scenarios, text summary of item description text, and feature vectors of item features based on the item details information of the sample items, thereby improving the comprehensiveness of the unstructured features of the stored sample items.

[0073] S230. Store the item title, item identifier, structured features and unstructured features of the sample item in the dynamic item feature database.

[0074] The dynamic item feature database can be either MySQL (relational database) or HBase (distributed database). The item title of a sample item refers to its name, and the item identifier refers to its ID (Identity document).

[0075] S240. Obtain the item recommendation request sent by the user, and determine the item recommendation information based on the dynamic item feature database and the item recommendation request through the item recommendation model.

[0076] The item recommendation model is a pre-trained large-scale artificial intelligence model.

[0077] Specifically, the system obtains the item recommendation request sent by the user from the client, inputs the item recommendation request into the item recommendation model, determines the database retrieval information based on the item recommendation request, and, based on the database retrieval information, examines the dynamic item feature database to determine the structured and unstructured features of sample items that match the database retrieval information. Based on these structured and unstructured features, the system determines the item identifier of the sample items that match the database retrieval information. This item identifier is used as the recommendation identifier, and the system determines the corresponding item title, structured features, and unstructured features. The recommendation identifier, its corresponding item title, structured features, and unstructured features are then used as item recommendation information and fed back to the client.

[0078] For example, training methods for item recommendation models may include:

[0079] The AI ​​model is trained using specific data from the item recommendation scenario to determine the pre-trained model; the model parameters of the pre-trained model are then fine-tuned using a low-rank adapter to determine the item recommendation model.

[0080] For example, data specific to the item recommendation scenario includes: natural language data from user searches using the intelligent shopping guide system, multi-dimensional descriptions of the items recommended by the system, and item recommendation response text. The large-scale AI model can be the Qwen / Qwen2.5-14B-Instruct model. Low-rank adapter (LoRA) is a fine-tuning technique for large-scale pre-trained models. Its core idea is to reduce the number of training parameters and computational overhead by introducing low-rank matrix factorization while maintaining model performance.

[0081] For example, when fine-tuning the model parameters of a pre-trained model using a low-rank adapter, the rank r of the LoRA adaptation layer can be set to 8, and the scaling factor alpha to 32; the AdamW optimizer can be used, with the learning rate of LoRA set to 1e-4; the model can be iteratively trained on the training data for 5 epochs, where an epoch refers to the process of the entire training dataset passing through the neural network once; the objective function is the standard language model cross-entropy loss function. After the model parameters are fine-tuned, most of the parameters of the pre-trained model are frozen, retaining only the parameters of the LoRA adaptation layer and the newly added layer, thus obtaining the item recommendation model.

[0082] Understandably, employing LoRa fine-tuning technology to adapt the large-scale shopping guide model to the relevant domain enhances product reasoning and multi-turn dialogue capabilities while retaining general semantic capabilities. Compared to full parameter fine-tuning, training efficiency is improved, and model response latency is reduced, providing technical support for high-concurrency real-time shopping guide scenarios.

[0083] The above solution, while training the AI ​​model using data specific to the item recommendation scenario, uses a low-rank adapter to fine-tune the model parameters of the pre-trained model to determine the item recommendation model. This approach can enhance the adaptability of the AI ​​model to shopping guide tasks while maintaining its overall capabilities.

[0084] The above-mentioned item recommendation method obtains the structured features of sample items entered by the business party. Through a multimodal large model, features are extracted from the item details of the sample items to determine their unstructured features. The item title, item identifier, structured features, and unstructured features of the sample items are stored in a dynamic item feature database. The method then retrieves the item recommendation request from the user and, based on the dynamic item feature database and the request, determines the item recommendation information using the item recommendation model. This solution addresses the problem that current intelligent shopping guide systems, when recommending items to users based on their input requests, often rely on a single recall strategy, making it difficult to cover structured needs such as item brands, attributes, and keywords, resulting in low relevance and insufficient diversity in the recall results. It also solves the problem that existing solutions often simply list items, lacking in-depth interpretation combined with an item feature database, failing to provide personalized recommendation reasons like a human shopping guide, and exhibiting a low level of intelligence. By constructing structured features of sample items based on item information entered by business users, and constructing unstructured features of sample items through a multimodal large model, a dynamic item feature database is built based on the structured features, unstructured features, item identifiers, and item titles of the sample items. Through an item recommendation model, based on the dynamic item feature database, the item recommendation requests issued by users are analyzed to determine the item recommendation information. This can improve the accuracy and diversity of item recommendation results generated by the intelligent shopping guide system. At the same time, it can combine the dynamic item feature database to deeply interpret item recommendation requests, thereby improving the intelligence level of the intelligent shopping guide system.

[0085] In one embodiment, such as Figure 3 As shown, the system obtains a user's item recommendation request, and determines item recommendation information based on a dynamic item feature database and the item recommendation request using an item recommendation model. This includes:

[0086] S310. Obtain the item recommendation request sent by the user, extract keywords from the item recommendation request through the item recommendation model, and determine the request keywords.

[0087] In this context, a user's item recommendation request refers to a natural language request entered into the intelligent shopping guide system. This request can be either voice or text data. The request includes category tags, which can be used to represent the user's intent. For example, user intent may include: requesting item recommendations, requesting item descriptions, requesting a transaction, or requesting logistics information.

[0088] Specifically, after the intelligent shopping guide system initializes and runs, it acquires the item recommendation requests sent by users. For example, an item recommendation request could be, "I want to buy my father a pair of breathable and non-slip athletic shoes." The item recommendation model determines category tags based on the item recommendation request and then determines whether the user's intent is to request item recommendations based on the category tags. If so, the item recommendation model extracts keywords from the item recommendation request to determine the request keywords. For example, request keywords could include: "father," "a pair," "breathable," "non-slip," and "athletic shoes."

[0089] For example, if the user's intent is determined to be something else based on the category tags, the corresponding application interface is called based on the category tags, and the service corresponding to the category tags is provided to the user through the application interface. For instance, if the user's intent is determined to be to query logistics based on the category tags, the logistics query interface is called to provide the user with a logistics query service.

[0090] S320. Encode the request keywords using a vector model to determine the query vector.

[0091] Among them, the vector model is BGE (BAAI / bge-base-zh, vector model).

[0092] S330. Based on the query vector and the preset recall quantity, perform an approximate nearest neighbor search on the vector database, determine candidate vectors from the dense vectors stored in the vector database, and determine the item identifier of the sample item corresponding to the candidate vector as the first candidate identifier.

[0093] The vector database stores the item identifiers and sample vectors of the sample items.

[0094] Approximate Nearest Neighbor Search (ANNS) is a technique for quickly finding vectors that are similar to the query vector in high-dimensional data, sacrificing some precision to improve search efficiency. The recall count can be set according to actual needs; for example, the recall count could be 50.

[0095] Specifically, if the recall quantity is 50, then based on the query vector and the preset recall quantity, an approximate nearest neighbor search is performed on the vector database to determine the 50 dense vectors most similar to the query vector as candidate vectors, and the item identifier of the sample item corresponding to the candidate vector is determined as the first candidate identifier. The first candidate identifier can be recorded in the first identifier list.

[0096] For example, methods for constructing a vector database include:

[0097] Using a vector model, the item description text of the sample item is encoded into a dense vector; the dense vector and the item identifier of the sample item are then stored in a vector database.

[0098] Dense vectors are a numerical data representation widely used in machine learning and data analysis. They consist of arrays containing real numbers, where most or all elements are non-zero.

[0099] Specifically, using a vector model, the item description text of the sample item is encoded into a 768-dimensional dense vector, and the dense vector and the item identifier of the sample item are stored in the vector database.

[0100] The above scheme provides a way to construct a vector database, which facilitates subsequent item recommendation models in determining item recommendation information based on the vector database.

[0101] S340. Based on the item recommendation request, and using the request keywords, recall quantity, and pre-built target index, perform a matching query on the dynamic item feature database, and determine the second candidate identifier from the item identifiers of the sample items in the dynamic item feature database.

[0102] Match queries, also known as full-text search, are a database or search engine technology used for full-text retrieval. They process query keywords through word segmentation and match data containing specified phrases. For example, when the query keyword is "Hangzhou West Lake Scenic Area," the system will break it down into three words: "Hangzhou," "West Lake," and "scenic area," and then search the data for relevant records containing all three words.

[0103] It should be noted that structured information indexes can be created in Elasticsearch (a logical container). For example, structured indexes can include item brand indexes, item attribute indexes, and item keyword indexes. A Chinese word segmenter can be set in Elasticsearch to determine the target index.

[0104] Specifically, if the recall quantity is 50, based on the item recommendation request, and using the request keywords, recall quantity, and pre-built target index, a matching query is performed on the dynamic item feature database. From the dynamic item feature database, the 50 sample items that best match the target keywords and target index are identified as second candidate items. This process recalls 50 candidate items that meet the matching criteria, and the item identifier corresponding to each second candidate item is determined as the second candidate identifier. The second candidate identifier can be recorded in the second identifier list.

[0105] S350. Based on the requested keywords, determine the item brand dimension data, item attribute dimension data, and item keyword dimension data, and construct a structured query vector based on the item brand dimension data, item attribute dimension data, and item keyword dimension data.

[0106] For example, if the item recommendation request is "I want to buy my father a pair of breathable and non-slip sneakers," then the request keywords include: "father," "a pair," "breathable," "non-slip," and "sneakers." In this case, the item brand dimension data extraction result is null, indicating that no brand is specified in this query; the item attribute dimension data extraction result is: ["breathable," "non-slip," "father"]; and the item keyword dimension data extraction result is ["sneakers"]. The item brand dimension data, item attribute dimension data, and item keyword dimension data are concatenated to determine the structured query vector.

[0107] S360. Based on the recall quantity and structured query vector, the dynamic item feature database is searched, and a third candidate identifier is determined from the item identifiers of the sample items in the dynamic item feature database.

[0108] Specifically, if the recall quantity is 50, the dynamic item feature database is searched based on the recall quantity and the structured query vector. The item structured feature that best matches the structured query vector is selected from the dynamic item feature database as a candidate structured feature, and the item identifier of the sample item corresponding to the candidate structured feature is used as the third candidate identifier. The third candidate identifier can be recorded in the third identifier list.

[0109] S370. Integrate the first candidate identifier, the second candidate identifier, and the third candidate identifier to determine the identifier to be evaluated, and determine the comprehensive score of the item corresponding to the identifier to be evaluated based on the recall path weight.

[0110] The recall path weight refers to the weight of the determination method corresponding to the first candidate identifier, the second candidate identifier, and the third candidate identifier. For example, the weight of the determination method for the first candidate identifier can be the first-path recall weight, the weight of the determination method for the second candidate identifier can be the second-path recall weight, and the weight of the determination method for the third candidate identifier can be the third-path recall weight. The first-path recall weight 'a' can be 0.4, the second-path recall weight 'b' can be 0.3, and the third-path recall weight 'c' can be 0.3.

[0111] It should be noted that there may be overlapping item identifiers among the first, second, and third candidate identifiers. Therefore, the comprehensive item score can be calculated based on the recall path weight.

[0112] Specifically, the first, second, and third candidate identifiers are integrated to determine the identifier to be evaluated. The formula for calculating the comprehensive score of the item corresponding to the identifier to be evaluated is shown in formula (1):

[0113] (1)

[0114] Wherein, score_all is the overall score of the item corresponding to the identifier to be evaluated, score1 is the item score corresponding to the recall path of the first candidate identifier, score2 is the item score corresponding to the recall path of the second candidate identifier, and score3 is the item score corresponding to the recall path of the third candidate identifier.

[0115] For example, if a user's item recommendation request is "I need to buy milk," and each recall path retrieves four items, the vector recall data includes: Brand A milk: 0.91, Brand B milk: 0.88, Brand C milk: 0.73, Brand D milk: 0.63. Then 0.91, 0.88, 0.73, and 0.63 are score 1. The text matching recall data includes: Brand B milk: 0.93, Brand C milk: 0.75, Brand E milk candy: 0.52, Brand F milk: 0.45. Then 0.93, 0.75, 0.52, and 0.45 are score 2. The structured semantic recall data includes: Brand A milk: 0.93, Brand D milk: 0.89, Brand B milk: 0.82, Brand F milk: 0.75. Then 0.93, 0.89, 0.82, and 0.75 are score 3.

[0116] S380. Based on the comprehensive score of the items, determine the target identifier from the identifiers to be evaluated, and generate item recommendation information based on the target identifier and the dynamic item feature database through the item recommendation model.

[0117] For example, the ten highest scores in the overall score of an item can be identified as the target identifiers. Then, an item recommendation model can be used to generate item recommendation information based on these target identifiers and a dynamic item feature database. The target identifier is the recommendation identifier, and the item recommendation information includes the target identifier, the item title corresponding to the recommendation identifier, the structured features of the item corresponding to the recommendation identifier, and the unstructured features corresponding to the recommendation identifier.

[0118] The above solution provides a multi-level intent-driven item recall mechanism that can accurately classify user-issued item recommendation requests using a large model. It constructs an architecture for keyword structured segmentation and three-way collaborative item recall, integrating vector recall, ES text matching recall, and structured semantic recall methods. Furthermore, it optimizes the accuracy of recall results through weighted calculation and a BGE model, effectively solving the problems of single recall dimensions and semantic understanding bias in traditional intelligent shopping guide systems, and improving the accuracy of item recommendations.

[0119] For example, such as Figure 4 As shown, based on the comprehensive score of the items, a target identifier is determined from the identifiers to be evaluated, and item recommendation information is generated based on the target identifier and the dynamic item feature database using an item recommendation model, including:

[0120] S410. Based on the overall score of the items, determine the target identifier from the identifiers to be evaluated.

[0121] For example, the top 30 items with the highest overall scores can be identified as the target identifiers.

[0122] S420. Determine the item metadata information based on the target identifier, and extract the structured features and unstructured features corresponding to the item identifier to be selected from the dynamic item feature database based on the target identifier.

[0123] The item metadata information includes the target identifier, the item title corresponding to the target identifier, and the item's overall score corresponding to the target identifier.

[0124] Specifically, the item metadata information is determined based on the target identifier, and the target identifier and its corresponding metadata information are input into the re-ranking model, which is a BGE-reranker model. Based on the ranking result of the re-ranking model, the top K target identifiers are determined as optional identifiers, and a set of optional identifiers is output. K is an integer greater than or equal to 1, and K can be 10. When K=10, the set of optional identifiers can be: [Prod1_id, Prod2_id, ...,Prod10_id]. The structured features and unstructured features corresponding to the optional item identifiers are extracted from the dynamic item feature database.

[0125] S430. Generate template based on prompt words: Generate item recommendation prompt words based on the item metadata information corresponding to the target identifier, the structured features corresponding to the target identifier, and the unstructured features corresponding to the item identifier to be selected.

[0126] Specifically, the item metadata information corresponding to the target identifier, the structured features corresponding to the target identifier, and the unstructured features corresponding to the item identifier to be selected are filled into a preset prompt word generation template to generate item recommendation prompt words.

[0127] S440. Using an item recommendation model, determine item recommendation information based on item recommendation prompts.

[0128] Specifically, the item recommendation prompts are input into the item recommendation model, which then determines the item recommendation information based on the prompts.

[0129] For example, an item recommendation model determines item recommendation information based on item recommendation prompts, including:

[0130] Obtain historical dialogue information between the user and the shopping guide agent; determine item recommendation information based on item recommendation prompts and historical dialogue information using an item recommendation model.

[0131] Specifically, a context memory module is pre-built as a dialogue state recorder. Historical dialogue information between the user and the intelligent sales agent is recorded in the context memory module, and the stored data in the context memory module is continuously updated. When item recommendation information needs to be generated, the user's account information is determined based on the item recommendation request issued by the user. With the user's permission, the historical dialogue information between the user and the intelligent sales agent is extracted from the context memory module based on the user's account information. The item recommendation prompts and the historical dialogue information between the user and the intelligent sales agent are input into the item recommendation model, and the item recommendation model outputs item recommendation information.

[0132] The above solution addresses the problem that traditional intelligent shopping guide systems, lacking long-term memory, cannot capture the evolution of user needs during multi-round dialogues, leading to intent recognition deviating from the user's true needs. It achieves deep coupling between item recommendations and user scenarios, ensuring that item recommendation results have both feature completeness and dialogue continuity, thereby improving the accuracy of personalized item recommendations.

[0133] It should be noted that, based on the comprehensive score of the items, the evaluation identifiers are further filtered to determine the target identifiers. Then, based on the item metadata information corresponding to the target identifiers and the candidate identifiers, the structured features and unstructured features corresponding to the candidate item identifiers are extracted from the dynamic item feature database. Based on the item metadata information corresponding to the target identifier, the structured features corresponding to the target identifier, and the unstructured features corresponding to the candidate item identifiers, item recommendation prompts are generated. Through the item recommendation model, the item recommendation information is determined based on the item recommendation prompts, which can improve the item matching coverage of complex shopping requests and improve the item recommendation accuracy.

[0134] For example, based on the above embodiments, the item recommendation method includes:

[0135] The AI ​​model is trained using specific data for the item recommendation scenario to determine the pre-trained model. The low-rank adapter (LoRA) is then used to fine-tune the parameters of the pre-trained model to determine the item recommendation model. This specific data includes natural language data from user searches through an intelligent shopping guide system, multi-dimensional descriptions of recommended items, and item recommendation response text. The AI ​​model can be a Qwen / Qwen2.5-14B-Instruct model. When fine-tuning the pre-trained model's parameters using the low-rank adapter, the LoRA adaptation layer's rank (r) is set to 8, and the scaling factor (alpha) to 32. The AdamW optimizer is used, with a learning rate of 1e-4 for LoRA. The model is iteratively trained on the training data for 5 epochs (epochs refer to the entire training dataset passing through the neural network once). The objective function is the standard language model cross-entropy loss function. After fine-tuning the model parameters, most of the pre-trained model's parameters are frozen, retaining only the LoRA adaptation layer parameters and the newly added layers, thus obtaining the item recommendation model.

[0136] Business users can input the structured features of sample items into the intelligent shopping guide system via terminal devices. These structured features include brand information, series information, material information, target audience, core functions, style, and main technical characteristics. Using a multimodal large model, the system extracts the item color and main usage scenarios from the item images in the item details information. It also extracts text summaries and item features from the item description text in the details information, and vectorizes these features to determine their feature vectors. The item color, main usage scenarios, text summaries of the description text, and feature vectors are used as the unstructured features of the sample items. The item title, item identifier, structured features, and unstructured features are stored in a dynamic item feature database. This dynamic feature database can be MySQL or HBase.

[0137] After the intelligent shopping guide system initializes and starts running, it receives item recommendation requests from users. The item recommendation model determines category tags based on the request and then determines whether the user's intent is to request item recommendations. If so, the model extracts keywords from the request to determine the desired keywords. If the category tags indicate the user's intent is something else, the system calls the corresponding application interface to provide the user with the services associated with those category tags.

[0138] The query vector is determined by encoding the request keywords using a vector model. The item description text of the sample items is then encoded into a 768-dimensional dense vector using the vector model, and the dense vectors and item identifiers of the sample items are stored in a vector database. If the recall quantity is 50, an approximate nearest neighbor search is performed on the vector database based on the query vector and the preset recall quantity. The 50 dense vectors most similar to the query vector are selected as candidate vectors, and the item identifiers of the sample items corresponding to these candidate vectors are designated as the first candidate identifiers. These first candidate identifiers are recorded in a first identifier list. Based on the item recommendation request, the request keywords, recall quantity, and pre-built target index are used to perform a matching query on the dynamic item feature database. The 50 sample items most closely matching the target keywords and target index are selected as second candidate items, thus recalling 50 candidate items that meet the matching conditions. The item identifiers corresponding to these second candidate items are designated as second candidate identifiers. These second candidate identifiers are recorded in a second identifier list. Based on the request keywords, we determine the item brand dimension data, item attribute dimension data, and item keyword dimension data. Then, we construct a structured query vector based on these data. We search a dynamic item feature database based on the recall count and the structured query vector. From this database, we identify the item structured features that best match the structured query vector as candidate structured features, and the item identifiers of the sample items corresponding to these candidate features are used as third candidate identifiers. These third candidate identifiers are recorded in a third identifier list. We integrate the first, second, and third candidate identifiers to determine the identifiers to be evaluated. Based on the recall path weights, we determine the overall item score corresponding to each identifier to be evaluated. We identify the ten identifiers with the highest overall item scores as target identifiers, and generate item recommendation information based on these target identifiers and the dynamic item feature database using an item recommendation model. The target identifier is the recommendation identifier, and the item recommendation information includes the target identifier, the item title corresponding to the recommendation identifier, the item structured features corresponding to the recommendation identifier, and the unstructured features corresponding to the recommendation identifier.

[0139] The top 30 items with the highest overall scores are identified as target identifiers. Item metadata is determined based on these target identifiers, and the target identifiers and their corresponding metadata are input into a re-ranking model (BGE-reranker model). The top K target identifiers are selected as optional identifiers based on the re-ranking model's results, and a set of optional identifiers is output. The item metadata, structured features, and unstructured features of the candidate item identifiers are then filled into a pre-defined prompt generation template to generate item recommendation prompts. A context memory module is pre-built as a dialogue state recorder. Historical dialogue information between the user and the shopping guide agent is recorded in the context memory module, and the stored data is continuously updated. When item recommendation information needs to be generated, the user's account information is determined based on the user's item recommendation request. With the user's permission, historical dialogue information between the user and the shopping guide agent is extracted from the context memory module based on the user's account information. The item recommendation prompts and the historical dialogue information are then input into the item recommendation model, which outputs the item recommendation information.

[0140] The above solution acquires the structured features of sample items entered by the business party. Using a multimodal large model, it extracts features from the item details of the sample items to determine their unstructured features. The item title, item identifier, structured features, and unstructured features are stored in a dynamic item feature database. The solution then receives user-initiated item recommendation requests and, using an item recommendation model, determines recommended item information based on the dynamic feature database and the request. This solution addresses the problem that current intelligent shopping guide systems, when recommending items based on user-input requests, often rely on a single recall strategy, failing to cover structured needs such as brand, attributes, and keywords, resulting in low relevance and insufficient diversity in the recall results. It also solves the problem that existing solutions often simply list items, lacking in-depth analysis combined with an item feature database, and cannot provide personalized recommendation reasons like a human shopping guide, resulting in a low level of intelligence. By constructing structured features of sample items based on item information entered by business users, and constructing unstructured features of sample items through a multimodal large model, a dynamic item feature database is built based on the structured features, unstructured features, item identifiers, and item titles of the sample items. Through an item recommendation model, based on the dynamic item feature database, the item recommendation requests issued by users are analyzed to determine the item recommendation information. This can improve the accuracy and diversity of item recommendation results generated by the intelligent shopping guide system. At the same time, it can combine the dynamic item feature database to deeply interpret item recommendation requests, thereby improving the intelligence level of the intelligent shopping guide system.

[0141] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0142] Based on the same inventive concept, this application also provides an item recommendation device for implementing the item recommendation method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more item recommendation device embodiments provided below can be found in the limitations of the item recommendation method described above, and will not be repeated here.

[0143] In one embodiment, such as Figure 5 As shown, an item recommendation device is provided, including: an item feature acquisition module 501, a feature extraction module 502, a data storage module 503, and an item recommendation information determination module 504, wherein:

[0144] The item feature acquisition module 501 is used to acquire the item structure features of the sample items entered by the business party; the item structure features include item brand information, item series information, item material information, target audience of the item, item core functions, item style and main technical features of the item;

[0145] The feature extraction module 502 is used to extract features from the item details information of the sample item using a multimodal large model to determine the unstructured features of the sample item; the item details information includes item images and item description text;

[0146] The data storage module 503 is used to store the item title, item identifier, structured features and unstructured features of the sample item in a dynamic item feature database.

[0147] The item recommendation information determination module 504 is used to obtain the item recommendation request issued by the user, and determine the item recommendation information based on the dynamic item feature database and the item recommendation request through the item recommendation model; the item recommendation model is a pre-trained large artificial intelligence model.

[0148] For example, the item recommendation device also includes:

[0149] The model training module is used to train a large-scale artificial intelligence model using data specific to the item recommendation scenario, and to determine the pre-trained model; a low-rank adapter is used to fine-tune the model parameters of the pre-trained model to determine the item recommendation model.

[0150] For example, the feature extraction module 502 is specifically used for:

[0151] Using a multimodal large model, the color of the item and its main usage scenarios are extracted from the item images in the item details information of the sample items;

[0152] Using the multimodal large model, text summaries and item features are extracted from the item description text in the item details information, and the item features are vectorized to determine the feature vectors of the item features;

[0153] The color of the item, the main usage scenarios of the item, the text summary of the item description text, and the feature vector of the item features are used as unstructured features of the sample item.

[0154] For example, the item recommendation information determination module 504 is specifically used for:

[0155] Get the item recommendation request sent by the user, extract keywords from the item recommendation request through the item recommendation model, and determine the request keywords;

[0156] The query vector is determined by encoding the request keywords using a vector model.

[0157] Based on the query vector and the preset recall quantity, an approximate nearest neighbor search is performed on the vector database to determine candidate vectors from the dense vectors stored in the vector database, and the item identifier of the sample item corresponding to the candidate vector is determined as the first candidate identifier; the vector database stores the item identifier and sample vector of the sample item.

[0158] Based on the item recommendation request, and using the request keywords, the recall quantity, and the pre-built target index, a matching query is performed on the dynamic item feature database to determine a second candidate identifier from the item identifiers of the sample items in the dynamic item feature database.

[0159] Based on the requested keywords, determine the item brand dimension data, item attribute dimension data, and item keyword dimension data, and construct a structured query vector based on the item brand dimension data, the item attribute dimension data, and the item keyword dimension data;

[0160] The dynamic item feature database is searched based on the recall quantity and structured query vector, and a third candidate identifier is determined from the item identifiers of the sample items in the dynamic item feature database.

[0161] The first candidate identifier, the second candidate identifier, and the third candidate identifier are integrated to determine the identifier to be evaluated, and the comprehensive score of the item corresponding to the identifier to be evaluated is determined based on the recall path weight.

[0162] Based on the comprehensive score of the items, a target identifier is determined from the identifiers to be evaluated, and item recommendation information is generated based on the target identifier and the dynamic item feature database through the item recommendation model.

[0163] For example, the item recommendation information determination module 504 is also used for:

[0164] Based on the overall score of the items, the target labels are determined from the labels to be evaluated.

[0165] The item metadata information is determined based on the target identifier, and the structured features and unstructured features corresponding to the item identifier to be selected are extracted from the dynamic item feature database based on the target identifier. The item metadata information includes the target identifier, the item title corresponding to the target identifier, and the item comprehensive score corresponding to the target identifier.

[0166] Based on the prompt word generation template, the item recommendation prompt words are generated according to the item metadata information corresponding to the target identifier, the structured features corresponding to the target identifier, and the unstructured features corresponding to the item identifier to be selected.

[0167] The item recommendation model determines item recommendation information based on the item recommendation prompts.

[0168] For example, the item recommendation information determination module 504 is also used for:

[0169] Obtain historical dialogue information between users and the intelligent shopping guide agent;

[0170] The item recommendation model determines item recommendation information based on the item recommendation prompts and the historical dialogue information.

[0171] For example, the item recommendation information determination module 504 is also used for:

[0172] The item description text of the sample items is encoded into dense vectors using a vector model;

[0173] The dense vectors and the item identifiers of the sample items are stored in the vector database.

[0174] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements an item recommendation method. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0175] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0176] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0177] Step 1: Obtain the structural features of the sample items entered by the business party; the structural features of the items include the item brand information, item series information, item material information, target audience of the item, core functions of the item, item style and main technical features of the item;

[0178] Step 2: Using a multimodal large model, extract features from the item details information of the sample items to determine the unstructured features of the sample items; the item details information includes item images and item description text;

[0179] Step 3: Store the item title, item identifier, structured features, and unstructured features of the sample item in the dynamic item feature database.

[0180] Step 4: Obtain the item recommendation request sent by the user, and determine the item recommendation information based on the dynamic item feature database and the item recommendation request through the item recommendation model; the item recommendation model is a pre-trained large artificial intelligence model.

[0181] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0182] Step 1: Obtain the structural features of the sample items entered by the business party; the structural features of the items include the item brand information, item series information, item material information, target audience of the item, core functions of the item, item style and main technical features of the item;

[0183] Step 2: Using a multimodal large model, extract features from the item details information of the sample items to determine the unstructured features of the sample items; the item details information includes item images and item description text;

[0184] Step 3: Store the item title, item identifier, structured features, and unstructured features of the sample item in the dynamic item feature database.

[0185] Step 4: Obtain the item recommendation request sent by the user, and determine the item recommendation information based on the dynamic item feature database and the item recommendation request through the item recommendation model; the item recommendation model is a pre-trained large artificial intelligence model.

[0186] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0187] Step 1: Obtain the structural features of the sample items entered by the business party; the structural features of the items include the item brand information, item series information, item material information, target audience of the item, core functions of the item, item style and main technical features of the item;

[0188] Step 2: Using a multimodal large model, extract features from the item details information of the sample items to determine the unstructured features of the sample items; the item details information includes item images and item description text;

[0189] Step 3: Store the item title, item identifier, structured features, and unstructured features of the sample item in the dynamic item feature database.

[0190] Step 4: Obtain the item recommendation request sent by the user, and determine the item recommendation information based on the dynamic item feature database and the item recommendation request through the item recommendation model; the item recommendation model is a pre-trained large artificial intelligence model.

[0191] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0192] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0193] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0194] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for recommending items, characterized in that, The item recommendation method is applied to an intelligent shopping guide system, and the item recommendation method includes: Obtain the structural features of the sample items entered by the business party; the structural features of the items include the item brand information, item series information, item material information, target audience of the item, core functions of the item, item style and main technical features of the item; Using a multimodal large model, features are extracted from the item details information of the sample items to determine their unstructured features; the item details information includes item images and item description text. The item title, item identifier, structured features, and unstructured features of the sample item are stored in a dynamic item feature database. The system obtains item recommendation requests from users and determines item recommendation information based on the dynamic item feature database and the item recommendation request using an item recommendation model. The item recommendation model is a pre-trained large-scale artificial intelligence model.

2. The method according to claim 1, characterized in that, Also includes: The AI ​​model is trained using data specific to the item recommendation scenario to determine the pre-trained model. The model parameters of the pre-trained model are fine-tuned using a low-rank adapter to determine the item recommendation model.

3. The method according to claim 1, characterized in that, Using a multimodal large model, feature extraction is performed on the item details information of the sample items to determine the unstructured features of the sample items, including: Using the multimodal large model, the item color and main usage scenarios are extracted from the item images in the item details information; Using the multimodal large model, text summaries and item features are extracted from the item description text of the item details information, and the item features are vectorized to determine the feature vectors of the item features; The color of the item, the main usage scenarios of the item, the text summary of the item description text, and the feature vector of the item features are used as unstructured features of the sample item.

4. The method according to claim 1, characterized in that, The process involves obtaining a user's item recommendation request, and then determining item recommendation information using an item recommendation model based on the dynamic item feature database and the item recommendation request. This includes: Obtain the item recommendation request sent by the user, and extract keywords from the item recommendation request using the item recommendation model to determine the request keywords; The query vector is determined by encoding the request keywords using a vector model. Based on the query vector and the preset recall quantity, an approximate nearest neighbor search is performed on the vector database to determine candidate vectors from the dense vectors stored in the vector database, and the item identifier of the sample item corresponding to the candidate vector is determined as the first candidate identifier; the vector database stores the item identifier and sample vector of the sample item. Based on the item recommendation request, and using the request keywords, the recall quantity, and the pre-built target index, a matching query is performed on the dynamic item feature database to determine a second candidate identifier from the item identifiers of the sample items in the dynamic item feature database. Based on the requested keywords, determine the item brand dimension data, item attribute dimension data, and item keyword dimension data, and construct a structured query vector based on the item brand dimension data, the item attribute dimension data, and the item keyword dimension data; The dynamic item feature database is searched based on the recall quantity and structured query vector, and a third candidate identifier is determined from the item identifiers of the sample items in the dynamic item feature database. The first candidate identifier, the second candidate identifier, and the third candidate identifier are integrated to determine the identifier to be evaluated, and the comprehensive score of the item corresponding to the identifier to be evaluated is determined based on the recall path weight. Based on the comprehensive score of the items, target identifiers are determined from the identifiers to be evaluated, and item recommendation information is generated based on the target identifiers and the dynamic item feature database using an item recommendation model.

5. The method according to claim 4, characterized in that, Based on the overall score of the items, a target identifier is determined from the identifiers to be evaluated, and item recommendation information is generated using an item recommendation model based on the target identifier and the dynamic item feature database, including: Based on the overall score of the items, the target identifier is determined from the identifiers to be evaluated; The item metadata information is determined based on the target identifier, and the structured features and unstructured features corresponding to the item identifier to be selected are extracted from the dynamic item feature database based on the target identifier. The item metadata information includes the target identifier, the item title corresponding to the target identifier, and the item comprehensive score corresponding to the target identifier. Based on the prompt word generation template, the item recommendation prompt words are generated according to the item metadata information corresponding to the target identifier, the structured features corresponding to the target identifier, and the unstructured features corresponding to the item identifier to be selected. The item recommendation model determines item recommendation information based on the item recommendation prompts.

6. The method according to claim 5, characterized in that, The item recommendation model determines item recommendation information based on the item recommendation prompts, including: Obtain historical dialogue information between users and the intelligent shopping guide agent; The item recommendation model determines item recommendation information based on the item recommendation prompts and the historical dialogue information.

7. The method according to claim 4, characterized in that, Also includes: The item description text of the sample items is encoded into dense vectors using a vector model; The dense vectors and the item identifiers of the sample items are stored in the vector database.

8. An item recommendation device, characterized in that, The item recommendation device includes: The item feature acquisition module is used to acquire the item structure features of sample items entered by the business party. The item structure features include item brand information, item series information, item material information, target audience, core functions, style, and main technical features. The feature extraction module is used to extract features from the item details information of the sample items using a multimodal large model, and to determine the unstructured features of the sample items; the item details information includes item images and item description text; The data storage module is used to store the item title, item identifier, structured features, and unstructured features of the sample item in a dynamic item feature database. The item recommendation information determination module is used to obtain the item recommendation request issued by the user, and determine the item recommendation information based on the dynamic item feature database and the item recommendation request through the item recommendation model; the item recommendation model is a pre-trained large artificial intelligence model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images