Commodity object search method and apparatus, and electronic device and storage medium

Abstract

Provided in the present disclosure are a commodity object search method and apparatus, and an electronic device and a storage medium. According to the embodiments of the present disclosure, on the basis of a search request triggered on the basis of a client, search basis data for searching for a commodity object may be acquired first; then, on the basis of the search basis data, a first prompt word corresponding to a commodity object search scenario is generated, and interaction with a large language model is performed by means of the first prompt word, so as to obtain search intent information corresponding to the current commodity object search; and finally, on the basis of the search basis data and the search intent information, a second prompt word corresponding to the commodity object search scenario is generated, and interaction with the large language model is performed by means of the second prompt word, so as to obtain a search result corresponding to the current commodity object search. By using the solution, interaction can be performed with a large language model, so as to deeply analyze and understand a search intent of a client user in searching for a commodity object, and provide a more accurate commodity object recommendation scheme on the basis of the search intent.

Description

Search method and device for commodity object, electronic device and storage medium

[0001] The present disclosure claims priority to Chinese Patent Application No. 202410660175.9, filed on May 24, 2024, with the Chinese Patent Office, entitled "Search method and device for commodity object, electronic device and storage medium", the entire contents of which are incorporated herein by reference. TECHNICAL FIELD

[0002] The present disclosure relates to the technical field of computers, and in particular to a search method and device for commodity objects, an electronic device and a storage medium. BACKGROUND

[0003] With the development of wireless network technology in the computer field and mobile devices, users are increasingly inclined to use smart terminal devices or wireless devices such as smartphones and tablet computers for online shopping and information acquisition. In this process, due to the vast amount of online information, it is necessary to search for the required information in a large amount of data, which is time-consuming and labor-intensive, and therefore personalized recommendation is a key technology for improving user experience and enhancing user stickiness.

[0004] Taking B-end user procurement as an example, since B-end users are mostly commodity operators or sellers, when B-end users search for required commodities (source finding) through shopping websites, there are characteristics such as complex demand expression, simultaneous collection and main collection, and multiple decision dimensions, for example, a user operates a home store, since the types of commodities purchased by the home store are many, they may involve cabinets, bowls, tablecloths, sofas, and even home supplies such as aromatherapy and ornaments, and the source finding function of the shopping website used when purchasing these commodities is very complex (especially for customized commodities), the procurement process of each commodity may include screening, inquiring, comparing prices, and sampling professional operations, which all require the user to search and compare a large number of similar commodities, and the user is required to have a certain degree of proficiency in using the shopping website, which will generate a certain learning cost, and some users may feel that the shopping website is complex, or give up using the shopping website because they cannot find some functions in the shopping website, resulting in user loss; in addition, some users also lack effective comparison and decision basis for a large number of commodity supplies, and only rely on manual search and comparison of multiple commodities before making a decision, which not only consumes time and effort for the user, but also may result in decision errors due to the inability to obtain objective and comprehensive data, causing certain economic losses, and if professional persons are entrusted to conduct market research, additional costs are also required, which increases the user's operating costs. SUMMARY

[0005] The present disclosure provides a search method and device for commodity objects, an electronic device and a storage medium to solve one or more of the above technical problems.

[0006] In a first aspect, the embodiments of the present disclosure provide a search method for a commodity object, comprising: obtaining search basis data for searching the commodity object according to a search request triggered by a client; generating a first prompt word corresponding to a commodity object search scene based on the search basis data, and interacting with a large language model through the first prompt word to obtain search intent information corresponding to the current commodity object search; generating a second prompt word corresponding to the commodity object search scene based on the search basis data and the search intent information, and interacting with the large language model through the second prompt word to obtain a search result corresponding to the current commodity object search.

[0007] In a second aspect, the embodiments of the present disclosure provide a search method for a commodity object, comprising: obtaining a search result corresponding to the current commodity object search according to a search request triggered by a client; wherein the search result is obtained by interacting with a large language model through a second prompt word corresponding to the commodity object search scene after the second prompt word is generated based on search basis data and search intent information; the search intent information is obtained by interacting with the large language model through a first prompt word corresponding to the commodity object search scene after the first prompt word is generated based on the search basis data for searching the commodity object; and the search result is displayed based on the client.

[0008] In a third aspect, the embodiments of the present disclosure provide a search device for a commodity object, comprising: a search basis data obtaining module configured to obtain search basis data for searching the commodity object according to a search request triggered by a client; a search intent information obtaining module configured to generate a first prompt word corresponding to a commodity object search scene based on the search basis data, and obtain search intent information corresponding to the current commodity object search by interacting with a large language model through the first prompt word; and a search result obtaining module configured to generate a second prompt word corresponding to the commodity object search scene based on the search basis data and the search intent information, and obtain a search result corresponding to the current commodity object search by interacting with the large language model through the second prompt word.

[0009] In a fourth aspect, the embodiments of the present disclosure provide a search device for a commodity object, comprising: a search result obtaining module configured to obtain a search result corresponding to the current commodity object search according to a search request triggered by a client; wherein the search result is obtained by interacting with a large language model through a second prompt word corresponding to the commodity object search scene after the second prompt word is generated based on search basis data and search intent information; the search intent information is obtained by interacting with the large language model through a first prompt word corresponding to the commodity object search scene after the first prompt word is generated based on the search basis data for searching the commodity object; and a search result display module configured to display the search result based on the client.

[0010] In a fifth aspect, an electronic device is provided, which includes a memory, a processor, and a computer program stored in the memory, and the processor implements any of the above methods when executing the computer program.

[0011] In a sixth aspect, a computer-readable storage medium is provided, which stores a computer program, and the computer program, when executed by a processor, implements any of the above methods.

[0012] In a seventh aspect, a computer program product is provided, which includes a computer program, and the computer program, when executed by a processor, implements any of the above methods.

[0013] According to the embodiments of the present disclosure, the search basis data for searching the commodity object can be obtained according to the search request triggered by the client first; then the first prompt word corresponding to the commodity object search scene is generated based on the search basis data, and the large language model is interacted through the first prompt word to obtain the search intent information corresponding to the current commodity object search; finally, the second prompt word corresponding to the commodity object search scene is generated based on the search basis data and the search intent information, and the large language model is interacted through the second prompt word to obtain the search result corresponding to the current commodity object search. The above scheme can deeply analyze and understand the search intent of the client user searching for the commodity object through the interaction with the large language model, and generate more accurate commodity object recommendation scheme according to the search intent, and optimize the search result of the commodity object. Especially in the scene of B-end user purchasing commodities through a shopping website, if the search scheme of the commodity object provided by the present disclosure is adopted, the commodity procurement demand (search intent information) contained in the natural language expression of the user can be deeply understood through the multiple dialogue interactions between the large language model and the user, and then the assisted recommendation and deep search of the commodity object are carried out based on the commodity procurement demand of the user.

[0014] The above description is only a summary of the technical solutions of the present disclosure. In order to enable a clearer understanding of the technical means of the present disclosure, the contents of the specification can be implemented, and in order to enable the above and other purposes, features and advantages of the present disclosure to be more obvious and easy to understand, the specific implementation manner of the present disclosure is described below. BRIEF DESCRIPTION OF DRAWINGS

[0015] In the drawings, the same reference numbers in the several figures indicate corresponding or similar components or elements. The drawings are not necessarily to scale, and the emphasis is generally placed upon illustrating the principles of the present disclosure. It should be understood that the drawings only depict some embodiments of the present disclosure and should not be considered as limiting the scope of the present disclosure.

[0016] FIG. 1 shows a system flow diagram of a search scheme for a commodity object according to an embodiment of the present disclosure;

[0017] FIG. 2 shows an example of a client front-end page of a search scheme for a commodity object according to an embodiment of the present disclosure;

[0018] FIG. 3 shows an overall architecture diagram of a search scheme for a commodity object according to an embodiment of the present disclosure;

[0019] FIG. 4 shows an intent understanding flow diagram in a search scheme for a commodity object according to an embodiment of the present disclosure;

[0020] FIG. 5 shows a flow chart of a search method for a commodity object according to an embodiment of the present disclosure;

[0021] FIG. 6 shows a flow chart of another search method for a commodity object according to an embodiment of the present disclosure;

[0022] FIG. 7 shows a structural block diagram of a search device for a commodity object according to an embodiment of the present disclosure;

[0023] FIG. 8 shows a structural block diagram of another search device for a commodity object according to an embodiment of the present disclosure; and

[0024] FIG. 9 shows a block diagram of an electronic device for implementing an embodiment of the present disclosure. DETAILED DESCRIPTION

[0025] In the following, only certain exemplary embodiments are simply described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the concept or scope of the present disclosure. Therefore, the drawings and the description are considered to be exemplary in nature, rather than limiting.

[0026] To facilitate understanding of the technical solutions of the embodiments of the present disclosure, the related technologies of the embodiments of the present disclosure are described below. The following related technologies can be combined with the technical solutions of the embodiments of the present disclosure in any manner as optional schemes, and all of them belong to the protection scope of the embodiments of the present disclosure.

[0027] Due to the differences in identity, industry, etc. of different types of users of the client, the purchasing commodity logic is quite different. In some cases, the potential identity of the user is not inferred through the information and dialogue expression of the user, so that the commodity sourcing matching lacks the intention understanding of the identity of the user; only the keyword extraction of the information input by the user is performed without considering the associated purchasing and user mind preference appeal and the purchasing mode difference among different industries, so that the further sourcing and exploration space of the user for the commodity is relatively limited; in other cases, the search for the commodity object is biased towards the general content or knowledge query retrieval. Taking the B-end user purchasing commodity as an example, the general content or knowledge query retrieval lacks the understanding of the e-commerce industrialization and customization of purchasing and the professional knowledge, and it is difficult to provide accurate matching and filtering logic in the e-commerce field, which will lead to the search results of the commodity object that cannot accurately match the actual needs of the user.

[0028] The innovation of the present disclosure integrates the artificial intelligence (AIGC) capability on the search engine, focuses on the incremental sourcing value, and realizes the purpose of helping the user to completely express the sourcing appeal, deeply understanding the B-type spot or customized purchasing intention, and discovering more matching business opportunities based on the interaction with the large language model based on the natural language input. Based on the in-depth understanding of the expression needs of the user, the intelligent personalized B-end user purchasing suggestion and diversified undertaking are provided to optimize the matching efficiency of the business opportunity. Specifically, in the scheme provided by the embodiments of the present disclosure, after the user of the client inputs the dialogue information in the interface of the client, the large language model can perform intention recognition and reasoning according to the dialogue content, for example, recognizing the search intention (three types of sourcing intention, i.e. single category search, multi-category search or fuzzy search without specifying the category), and extracting the search intention keywords, price tendency, MOQ tendency (minimum order quantity) in the dialogue content, triggering the search recommendation and other sourcing service queries. When the search intention is recognized, the undertaking mode corresponding to the different types of search intention (single category search, multi-category search or fuzzy search without specifying the category) can be used, that is, the single category search corresponds to the single category search process, the multi-category search or the fuzzy search without specifying the category corresponds to the search process of multiple commodity object categories, for example, the single category search process can include providing a purchasing suggestion and centrally filtering the commodities meeting the search intention of the user; the multi-category search corresponds to the search process of multiple commodity object categories, which can include searching multiple commodity categories meeting the search intention in parallel; the fuzzy search without specifying the category corresponds to the search process of multiple commodity object categories, which can not only include searching multiple commodity categories meeting the search intention in parallel, but also give the suggestion of disassembling, purchasing or grouping according to the search intention.

[0029] Generally, the search only focuses on query understanding, focusing on the recall and screening of goods. In the present scheme, not only the keyword is focused on, but also the search basis data such as conversation with the user is considered, and the expression method, purchase intention and source seeking method of the user are understood, so as to better provide the user with accurate goods object search results and meet the user's needs. On the data link, the general search link is "query extraction-query recall-goods object screening"; and the search link of the scheme provided by the present disclosure can be summarized as "conversation-intention understanding-keyword (query), price, ability extraction (such as search engine, comparison tool, etc.)-ability distribution (such as returning search or comparison results of search engine, comparison tool)-combination of keyword (query) recall, price screening, comparison suggestion, etc.

[0030] FIG. 1 shows a system flow diagram of a goods object search scheme provided in an embodiment of the present disclosure. As shown in FIG. 1, the system flow of the goods object search scheme provided in the embodiment of the present disclosure can include the interaction flow among the client front end, the client backend (magellan), the intelligent search agent (SmartSearchAgent) and the large language model (LLM). Among them, the intelligent search agent (SmartSearchAgent) can be used as the execution subject in the embodiment of the present disclosure, and the intelligent search agent (SmartSearchAgent) is used to interact with the client backend and call the large language model or search engine, comparison tool (R-Lab Product Comparision Application) and the like, so as to realize the goods object search scheme provided by the present disclosure. Among them, the client backend can have the ability of session object (Session) maintenance, model reply cache, goods search result cache and the like. In addition, the large language model called in the embodiment of the present disclosure can be a large language model with a multi-layer Transformer network structure, and the training data thereof can include general large language model data or specific goods platform access behavior data, so that the large language model is more suitable for the specific goods platform, and then more accurate and personalized goods search suggestions can be provided for the user using the goods platform.

[0031] Firstly, search basis data for searching the commodity object can be acquired according to a search request triggered by a client. The search request triggered by the client can be a commodity search request (user natural language input) in the form of an image, voice and / or text input by the user through the client front end, and the search basis data for searching the commodity object can include the content input by the user to the client front end (which can be the content after a large language model is called to convert the commodity search request in the form of text, voice, picture, etc. into text form), and can also include an answer given by an artificial customer service to the content input by the user or an answer generated by an artificial intelligence (AI) (which can be realized by calling a large language model) according to the content input by the user. Taking the case of a buyer (user) purchasing through a shopping website as an example, the specific scenario can be that the user has a multi-round conversation with an artificial customer service or an AI, and a large language model can take the content of the acquired single-round conversation or multi-round conversation as the search basis data for searching the commodity object.

[0032] Secondly, a first prompt word corresponding to the commodity object search scene can be generated based on the search basis data, and the first prompt word is used to interact with the large language model to obtain search intent information corresponding to the current commodity object search. The first prompt word can be a prompt word with search basis data, and the first prompt word corresponding to the commodity object search scene can be generated by combining the first prompt word template for requesting to identify the search intent and the search basis data. As shown in FIG. 1, the large language model (LLM) can be called to perform intent understanding (IntentUnderstand) on the search basis data for searching the commodity object, and the first prompt word can be generated by combining the first prompt word template for requesting to identify the search intent and the search basis data (input, Input). Further, the first prompt word is used to interact with the large language model to obtain search intent information corresponding to the current commodity object search. The search intent information can include at least one of a search intent type, commodity object description information (which can be attribute description of the commodity object, and can be text or image, and the text can be input text or voice converted text), a commodity object category, a commodity object price, and a commodity object order quantity. The search intent type can include single category search, multi-category search, or fuzzy search without specifying a category. That is, the large language model can perform user intent identification (such as determining the search intent type), user intent disassembly (such as obtaining the commodity object price, the commodity object order quantity, etc.) according to the first prompt word, and further generate recommended commodities, give a recommendation reason, etc., and output the model results (which can be the above-mentioned search intent type, recommended commodities, recommendation reason, etc.) to the smart search agent (SmartSearchAgent).The smart search agent can recognize the model results of the large language model (Recognize), such as analyzing the model recognition results, identifying the search intent type returned by the large language model (such as multi-category search), prompting the large language model to use a multi-intent tool (multi-category search corresponds to a search process of multiple product object categories), for example, and returning diversified results in parallel search, and calling tools (such as calling parallel search capabilities in search engine tools, or calling capabilities in comparison tools, etc.). After retrieving the multi-intent parallel search results, the large language model (LLM) is called for observation (Obervation). The large language model can determine (End?) whether the final answer (Final Answer) is obtained from the perspective of whether it meets the user's intent. If yes (Yes), the user's intent can be further analyzed. If no (No), the input (Input) stage is returned to, and the first prompt word corresponding to the product object search scenario is regenerated based on the search basis data (which can be updated search basis data, such as data after several rounds of dialogue between the user and the artificial customer or intelligent AI), and the search intent information is obtained. That is, the search basis data can be updated in real time according to the user's input. For example, if the user's input or multiple rounds of dialogue do not completely express the user's search intent, the first prompt word generated based thereon cannot determine the user's intent as the final answer, so the search basis data can be reacquired and the first prompt word can be regenerated. This process can be repeated multiple times, and the present disclosure does not make any limitation on this. Optionally, in the above process, the generated model results or result recognition results can also be returned to the client front end for user confirmation.

[0033] Finally, based on the search basis data and the search intent information, a second prompt word corresponding to the product object search scenario can be generated, and the second prompt word can be used to interact with the large language model to obtain search results corresponding to the current product object search. As shown in FIG. 1, after the large language model determines the final answer (Yes), a second prompt word (not shown in FIG. 1) can be further generated to further analyze the user's intent.

[0034] The generation manner of the second prompt word can be based on the search user's commodity object access behavior (for example, according to the aforementioned first prompt word, it is judged that the user's intention is to find the source, and find the commodity object that meets the expectation) and / or commodity object search behavior (for example, according to the user's search history, or the expression manner when searching, it is judged that the user is concerned about the commodity object category, etc.), and the search basis data and the aforementioned search intention information identified according to the first prompt word, to generate the second prompt word corresponding to the commodity object search scene. Alternatively, the generation manner of the second prompt word can be to obtain user behavior feature information generated based on search user related user behavior information and / or user session information, wherein the user behavior feature information represents the user's preference for the commodity object and / or the transportation process. The user behavior feature can be used as an auxiliary parameter for post-search screening, and then based on the user behavior feature information, the search basis data and the search intention information, the second prompt word corresponding to the commodity object search scene is generated. The purpose of generating the second prompt word is to more accurately identify the user's intention according to more user and commodity related information, so as to give more accurate commodity object suggestions, etc.

[0035] The user intention can include a single category intention (single category search), a multi-category intention (multi-category search), and a wide category intention (fuzzy search without specifying the category). The single category search corresponds to a single category search process, and the multi-category search or the fuzzy search without specifying the category corresponds to a search process of multiple commodity object categories. For example, when it is judged as a single category intention, model procurement suggestions, platform screening conditions, etc. can be further given; when it is judged as a multi-category intention, multi-category search words can be further given; and when it is judged as a wide category intention, wide category search recommendations can be further given.

[0036] Further, the second prompt word is input into the large language model to enable the large language model to call a commodity search process corresponding to a search intent type indicated by the search intent information, obtain a search result corresponding to the current commodity object search, and return the search result to the intelligent search agent (the intelligent search agent obtains an intent understanding result, an IntentUnderstand result), and then return the result to the client backend (the large model answers and recommends content), and the client backend sends the result to the client front end for display (displaying the large model reply result + recommended search categories). The commodity object search process searches for commodity objects based on search basis data and search intent information, and the large language model can delete commodity objects that do not match the user behavior characteristic information from the search result. In addition, in the case where the search intent type indicated by the search intent information is a broad category intent (fuzzy search without specifying a category), the large language model can also output a plurality of commodity object categories corresponding to the fuzzy search without specifying a category and a purchase and combination purchase scheme of the plurality of commodity object categories. At the same time, the large language model can also output comparison data of the search result in at least one dimension, search intent information, and a commodity search process corresponding to the search intent type indicated by the search intent information.

[0037] For example, for different buyers, the user can input a search request through the corresponding entrance in the client page, and according to the search request input by the user, the large language model can be called to infer the user's identity, such as what type of buyer. The types of buyers can include group procurement buyers, special procurement buyers, new product buyers, and the like. The dialogue line of the group procurement buyer generally includes: broad expression initiation - issuing a multi-product procurement instruction - recommending a merchant to meet the multi-product demand, so the search intent type of the group procurement buyer is mostly fuzzy search without specifying a category. The dialogue line of the special procurement buyer generally includes: clear demand - AI provides indication and screening suggestion - multi-condition screening is applied, and then direct comparison is initiated for fine screening - comparison result, so the search intent type of the special procurement buyer is mostly multi-category search. The dialogue line of the new product buyer generally includes: looking for new - AI provides indication and screening suggestion, so the search intent type of the new product buyer is mostly single-category search, and part of it is multi-category search or fuzzy search without specifying a category. In some embodiments, the present disclosure can first infer the user's identity by calling the large language model to determine the user's preliminary search intent type (according to the first prompt word), and then further determine the user's intent through the user's search behavior and other user behavior characteristics (according to the second prompt word) to give a more accurate commodity object recommendation scheme that meets the user's intent.

[0038] The large model reply result and recommended search category displayed on the client front end can be displayed in different ways according to the user intent identified by the large language model. For example, when it is judged as a single category intent (single category search), the user can be prompted to select platform filtering conditions for further selection and search (user interaction initiates search), and the client backend can call tools based on the search request initiated by the user to search for agents, and the tool call can call the parallel search, filtering option search, and merchant search capabilities of the search engine, and then give the result of the single category platform search (single category intent case, with platform filtering options, SpecificGenreIntentSearch). When it is judged as a multi-category intent (multi-category search), the user can be asked whether to search for a supplier that can produce several categories at the same time, and the user can further initiate a search according to the prompt (user interaction initiates search), and the client backend can call tools based on the search request initiated by the user to search for agents, and the tool call can call the parallel search, filtering option search, and merchant search capabilities of the search engine, and then give the result of the supplier search (multi-category intent case, MultipleGenreIntentMerchantSearch). That is, the client backend can request the intelligent search agent to call tools during the process of single category platform search, supplier search, or obtaining large model answers and recommended content, such as calling search engines for parallel search (Parallel Search), filtering option search, or merchant search, or calling comparison tools, etc., and the present disclosure does not make any limitation thereto.

[0039] After requesting the intelligent search agent to make a tool call, such as calling a search engine for a parallel search, a filter item search, or a merchant search, etc., the commodity search result such as a supplier search or a single product category platform search, etc. can be returned to the client backend. The client backend can package the search result (search result packaging processing), and return the packaged search result to the client front end for commodity display. The user can further operate on the client front end, such as selecting a commodity for commodity comparison (user selects commodity comparison), etc. Then, through the client backend (commodity comparison, CompareProduction), the intelligent search agent is requested to make a tool call to call the comparison tool (R-Lab Product Comparision Application) to compare commodities. This step of commodity comparison can also be obtained by generating a prompt word (Prompt concatenation) and interacting with a large language model to obtain a commodity comparison result. Correspondingly, the comparison tool can process the generated comparison result or the commodity comparison result generated by the large language model (model reply processing) or result packaging, etc., and then return the result to the client backend. The client backend can further package the commodity comparison result, and return it to the client front end for display of the commodity comparison result.

[0040] FIG. 2 shows an example of a client-side front-end page of a search scheme for a commodity object provided in an embodiment of the present disclosure. In the search scheme for a commodity object provided by the present disclosure, the client-side front-end page can include an input box for a user to input search content, and the search results of the commodity object can be displayed in the style shown in FIG. 2. The input box of the client-side front-end page can be marked with the word "smart search" to prompt the user to input text, voice, picture, etc. for smart search; the input box can be displayed with "history", "disclaimer" and other controls or entrances for the user to view the search history (History), disclaimer (Disclaimer) terms, etc. For example, the user inputs "I manage a clothing store located in Country A. I am in search of women's autumn and winter clothing and planning to order at least 100 pieces of woman's sweaters." in the input box (the language used in this example is Chinese, English-speaking regions can use "I manage a clothing store located in Country A. I am in search of women's autumn and winter clothing and planning to order at least 100 pieces of woman's sweaters.", other language regions can also use the language used in the region, or multiple language types can be provided in the same region for users to choose, the same below, the present disclosure does not make any limitation), according to the search results obtained by the search scheme process for a commodity object provided by the present disclosure, such as interaction with a large language model, the search intent, procurement recommendation, commodity screening rule, analysis result, recommendation, etc. returned by the large language model can be displayed on the client-side front-end page. For example, the client-side front-end page can display the search intent returned by the large language model "Based on your needs, it is recommended that you consider the following factors when sourcing women's sweaters", where "women's sweaters" is the smart search summarizing performed by the large language model according to the user's input "I manage a clothing store located in Country A. I am in search of women's autumn and winter clothing and planning to order at least 100 pieces of woman's sweaters", i.e. the intent extraction and structuring of the large model. The client-side front-end page can also display the procurement recommendation returned by the large language model "Style: According to the preferences of the target market, different styles can be chosen, such as simple, classic, fashionable, loose, etc. to meet the needs of different customers. Design: Various design elements, such as patterns, collars and sleeves, can make the sweater more attractive.Considering the autumn and winter weather, you can choose a thicker style or a detachable collar design. (For English-speaking regions, this can be displayed as "Style: According to the target market preferences, different styles can be selected, such as simple, classic, fashionable, loose, etc., to meet the needs of different customers. Design: Various design elements, like patterns, collars, and sleeves, can make sweaters more appealing. Considering the weather in autumn and winter, you can choose a thicker style or a detachable collar design.") (Large Model: Specialized Knowledge Purchasing Advice Know-How). Furthermore, the client-side front-end page can also display the product filtering rules returned by the large language model: "The agent has collected a variety of selected products to meet your needs. Additionally, you can also choose: from experienced custom manufacturers with dedicated production lines." [1] Delivery will be on time. [2] Add to your order; Easy Cashback [3] Flights can return to local warehouses for free (in English-speaking regions, this may be displayed as "Agent collects a variety of selected products that can satisfy your needs. Besides, you can also choose to: Source from Verified Custom Manufactures with dedicated production lines; Add On-time Dispatch to your order; Filter for Easy Return to return to local warehouses for free") etc. (Client-side backend: Rules are revealed through platform filtering). Users can further select product filtering rules that meet their needs, and then call the large language model to delete product objects that do not meet the selected rules from the search results or re-search, in order to obtain product object search results that better meet the user's expectations.

[0041] Of course, the client-side front-end page can also display the analysis results returned by the large language model, such as "Products for 'women's sweaters with an MOQ of lower than 100' (in English regions, 'Products for 'women's sweaters with an MOQ of lower than 100' can be displayed)" - Women's Sweater 1 (with pictures), Price: 100-500 yuan, MOQ: 5; Women's Sweater 2 (with pictures), Price: 30-50 yuan, MOQ: 50; Women's Sweater 3 (with pictures), Price: 1000-5000 yuan, MOQ: 30" (analyze and display results, Apply and show results), and if there are too many results, a portion of the results can be displayed in a collapsed manner, and a "View more results" control or entry can be provided for users to view. At the same time, the client-side front-end page can also display the recommended content returned by the large language model, i.e., the client-side backend and the product side: category attribute / query recommendation, which can display controls or entries corresponding to the titles of the recommended content for users to click and view, such as displaying "Search for zippered cashmere sweaters" (in English regions, "Search for zippered cashmere sweaters" can be displayed), "Search for casual street style sweaters" (in English regions, "Search for casual street style sweaters" can be displayed), "Show other recommendations" (in English regions, "Show other recommendations" can be displayed), and the like. Controls or entries, and users can view the corresponding zippered cashmere sweaters, casual street style sweaters, and the like by clicking such controls or entries. "Like" (in English regions, "Like" can be displayed), "Dislike" (in English regions, "Dislike" can be displayed) controls can be provided for users to evaluate the recommended content displayed, and the user's search, evaluation, and other data can be used as training samples to further train the large language model, or as user preference and other user behavior features to return to the large language model, so that the large language model can generate more user-expected recommended products.

[0042] FIG. 3 shows a schematic diagram of the overall architecture of a search scheme for a commodity object provided in an embodiment of the present disclosure. In the overall architecture of the embodiment of the present disclosure, a client front end, a client backend (magellan), an intelligent search agent interface service (R-Lab-Agent-API service), a tool calling capability (star river capability), etc. can be included. Among them, the client front end can provide AI search floor, AI search entrance, AI dialogue entrance, etc. Capability. The client backend can provide intelligent search service (Smart Search service) based on the user input (User input) of the client front end, request processing (SSE output) related work content, such as query processing (Query processing), model calling, user information, data assembly, etc. According to the session information input by the client front end, the session and the buried point information can be output, which is used for online data tracking collection, and the mechanism of deep understanding of some core large language models records the dialogue process of multiple dialogues between the user and the AI, commodity interaction (whether the user purchases, browses some commodities, etc.), and statistics to the large language model for continuous optimization of the large language model. The client backend can initiate an intelligent search agent request (R-Lab Agent request) to the intelligent search agent interface service, and the intelligent search agent interface service can perform Agent intent recognition, Agent intent extraction, Agent tool calling, etc. according to the request of the client backend. Agent intent recognition can refer to outputting intent keywords according to input session information, that is, identifying which type of intent (single category search, multi-category search, or fuzzy search without specifying category) the user belongs to (single category search, multi-category search, or fuzzy search without specifying category) and the keywords of the intent, such as bowl, chopsticks, tablecloth, etc. in multi-category search (multi-intent), through the large language model to understand the user's session; Agent intent extraction can refer to outputting the calling capability tool according to the input intent keywords, calling different tool capabilities by identifying different intents, and supplementing the parameter information of the tool calling. Different tool capabilities can be different intent type corresponding processes, and the parameter information can be auxiliary functions for intent keywords (such as low price, cost-effective), analysis of other user behaviors in addition to the current session, determination of user consumption habits, etc. These analyses and determined habits will become parameters passed to the tool, such as real-time analysis of user behavior or user historical session, etc.; Agent tool calling can refer to outputting corresponding data information according to the parameter information of the input tool, such as commodity search, commodity screening, etc. Through tool calling, specific commodity data / suggestions are obtained, and the commodity is searched and then screened. The basis for commodity screening can be determined according to the aforementioned parameters (such as price, minimum order quantity MOQ, delivery time).

[0043] Among them, Agent intent recognition can include session management (creation / switching / lease renewal) and the like, such as managing the sessions of robots A (chat A), B (chat B), and C (chat C) with different users respectively, creating, switching, or renewing the lease of robots, and the like; Agent intent extraction can include purchase intent, service intent, commodity intent, or comparison intent, and the like; Agent tool invocation can include parameter mapping, request processing, point recording, or content aggregation, and the like. In addition, the tool invocation capability (star river capability) in the overall architecture can be invoked by the intelligent search agent, and after invocation, it can provide commodity search, trend purchase, or commodity comparison capabilities based on tool asynchronous requests (Tools asynchronous requests).

[0044] Correspondingly, based on the above architecture, a search scheme of a commodity object can be provided for a user in four steps. Among them, the first step is search service processing (request processing, data assembly); the second step is intelligent search agent intent analysis (Agent intent recognition, task allocation, instruction generation); the third step is intelligent search agent (Agent) tool execution (final execution of function instructions); the fourth step is search page rendering (streaming output, page rendering). In this process, the NCMS building platform (used to build the client front-end display interface) can be used to generate the display interface of the client front-end, the NCMS building platform can input the feedback of the large language model and tool information, and output the interactive components of the client front-end page, so that the client front-end becomes a visual page feedback to the user through data rendering. After the above steps, in the search results of the commodity object displayed on the client front-end through page response (streaming rendering), intent analysis floor, rule screening floor, and commodity search floor can be included, for example, as shown in the display results of FIG. 2, as shown in FIG. 3, the intent analysis floor can include large model intent extraction and structuring and the content of large model specialized knowledge purchase suggestion (Know How); the rule screening floor can include the content screened by the client back-end rule exposure platform; the commodity search floor can include the content of analyzing and displaying the results, and the like. The same parts in FIG. 3 as in FIG. 2 are not described here.

[0045] FIG. 4 shows a schematic diagram of an intent understanding process in a search scheme for a commodity object according to an embodiment of the present disclosure. As shown in FIG. 4, according to the natural language input by the client front-end user, a large language model can be called to analyze the user's natural language input, identify the user's intent, and perform intent classification. For example, according to the content of the user's dialogue with the professional procurement assistant AI, the large language model can be called to identify the conversation intent, and then the large language model can be called to generate a prompt according to the conversation information and the conversation scenario, and the large language model can be called to search for commodity results that meet the conversation intent. The generated prompt can be generated according to a pre-trained prompt template and user information, and the large language model can be a large language model pre-trained using multiple conversation scenarios. The large language model classifies the intent, which can be roughly classified into three categories, namely single-category search (user input determines a single product category), multi-category search (user actively inputs multi-category intent), or fuzzy search without specifying the category (user input or based on the scene / identity to determine a broad product category intent).

[0046] According to different intent classification, single-category search can correspond to a single-category search process, and multi-category search or fuzzy search without specifying the category can correspond to a search process of multiple commodity object categories. Specifically, as shown in FIG. 4, in the case of intent classification as single-category search (user input determines a single product category), the corresponding single-category search process can be procurement suggestion and platform mind screening. Such suggestions can be pre-trained on a large language model using network public data, purchase records of other buyers who purchase the same category of commodity objects, or other buyers who input the same keywords, so that the large language model can give such suggestions. In the case of intent classification as multi-category search (user actively inputs multi-category intent), the corresponding search process of multiple commodity object categories can be parallel retrieval of multiple categories (for example, 3 categories of commodity objects can be simultaneously and in parallel retrieved), that is, when processing the user's query request, multiple search tasks can be simultaneously executed to quickly and efficiently retrieve relevant commodities from multiple intents (queries) and return the results to the user, thereby improving the user experience. In the case of intent classification as fuzzy search without specifying the category (user input or based on the scene / identity to determine a broad product category intent), the corresponding search process of multiple commodity object categories can include parallel retrieval of subdivided categories derived from a broad product category, and give disassembly / procurement / warehousing suggestion, that is, on the basis of parallel retrieval of multiple categories, further suggestions can be given, for example, outputting the multiple commodity object categories corresponding to the fuzzy search without specifying the category and the procurement and combination purchase scheme of the multiple commodity object categories.

[0047] Finally, after the above search process ends, the user can continue to input through the client front end or give other selectable controls to click to guide the user to further understand the user's intention and classify the intention. After the above search process is repeated several times, the real intention of the user can be more accurately judged, and more accurate commodity suggestions can be given. In this process, the answers of the large language model and the contents of the recommended words can also be returned to the client front end to be displayed to the user for viewing and timely correction of the corresponding content.

[0048] The execution subject of the embodiments of the present disclosure can be an application program, a service, an instance, a functional module in the form of software, a virtual machine (VM), a container, or a cloud server, etc., or a hardware device (such as a server or a terminal device) or a hardware chip (such as a CPU, a GPU, a FPGA, a NPU, an AI accelerator card, or a DPU) with a data processing function. The device for realizing commodity object search can be deployed on the computing device of the application party providing the corresponding service or the cloud computing platform providing computing power, storage, and network resources. The mode of the cloud computing platform providing services to the outside world can be IaaS (Infrastructure as a Service, Infrastructure as a Service), PaaS (Platform as a Service, Platform as a Service), SaaS (Software as a Service, Software as a Service), or DaaS (Data as a Service, Data as a Service). Taking the platform providing SaaS software as a service (Software as a Service) as an example, the cloud computing platform can use its own computing resources to provide the training of the commodity object search model or the functional execution of the commodity object search module. The specific application architecture can be built according to the service requirements. For example, the platform can provide a construction service based on the above-mentioned model to the application party or individual using the platform resources, and further call the above-mentioned model and realize the function of online or offline commodity object search based on the commodity object search request submitted by the related client or server device.

[0049] 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 for analysis, stored data, displayed data, etc.) involved in the present disclosure are all information and data authorized by the user or authorized by all parties, and the collection, use, and processing of related data need to comply with relevant laws, regulations, and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose authorization or refusal.

[0050] The technical solutions of the present disclosure and how the technical solutions of the present disclosure solve the foregoing technical problems will be described in detail below with specific embodiments. The several specific embodiments listed can be combined with each other, and the same or similar concepts or processes can not be described again in some embodiments. The embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0051] The embodiment of the present disclosure provides a search method for a commodity object. As shown in FIG. 5, a flowchart of a search method 500 for a commodity object according to an embodiment of the present disclosure, the method 500 can include:

[0052] In step S501, according to a search request triggered based on a client, search basis data for searching a commodity object is obtained.

[0053] The search basis data is used as a basis for screening target commodity objects that meet the user's intention when searching for commodity objects. According to the search basis data, the scene and user intention involved in the conversation content of the client can be identified, and then the commodity meeting the user's intention can be searched.

[0054] The search basis data includes but is not limited to real-time conversation content of the client and historical storage data related to the real-time conversation content, for example, content input by the user to the front end of the client (non-text form input data such as voice, picture, etc. can be converted into text form), answers to the content input by the user by the artificial customer service or answers generated by the artificial intelligence AI (which can be realized by calling a large language model) according to the content input by the user, user behavior data, scene information data, commodity information data, etc. stored in the back end or cloud database related to the above-mentioned content.

[0055] In one possible implementation, the search basis data for searching the commodity object can be obtained by obtaining the image, voice and / or text input based on the client.

[0056] That is, the large language model can convert the text, voice, picture and other forms of commodity search requests input by the user from the front end of the client or output by the artificial customer service, AI, etc. to the front end of the client into text form, and the content converted into text form can be used as search basis data for searching the commodity object. Taking the purchase of a buyer (user) through a shopping website as an example, the specific scene can be that the user has a multi-round conversation with the artificial customer service or AI, and the user, the artificial customer service or AI, etc. can input images, voices and / or texts to the front end of the client for single-round or multi-round conversation during the multi-round conversation. The large language model can take the content of the single-round or multi-round conversation obtained by the front end of the client as the search basis data for searching the commodity object.

[0057] In step S502, a first prompt word corresponding to a commodity object search scene is generated based on the search basis data, and the first prompt word is used to interact with a large language model to obtain search intent information corresponding to the current commodity object search.

[0058] The first prompt word can be a prompt word with search basis data, and can extract keywords from text, pictures, voice, etc. obtained by the client, or extract keywords from the conversation content of the user and the AI dialogue, or use all the content obtained by the front end of the client as the search basis data. The first prompt word can be generated in combination with a first type of prompt word template used to request recognition of a search intent and the search basis data.

[0059] The large language model involved in the embodiments of the present disclosure can be a large-scale artificial intelligence model, especially a model in the field of natural language processing (NLP). Such a model has a very large number of parameters, which can reach tens of billions or even hundreds of billions of parameters. Due to its large size, these models usually require very high computing resources for training and running. That is, the large language model can be a deep learning model or a neural network model trained using a large amount of text data, which works by analyzing a large amount of text data and learning the patterns of language use, and can generate natural language text or understand the meaning of language text. The generated text is very close to what humans say or write. It is mainly divided into models based on autoencoders, sequence-to-sequence models, models based on Transformers, recurrent neural network models, and hierarchical models, etc. Different large language models are good at processing different types of tasks, for example, a model based on an autoencoder encodes the input text into a lower-dimensional representation, and then generates new text from the representation, which is suitable for processing tasks such as text summarization or content generation, and can be used for content generation of various content types related to the list; a model based on Transformer adopts a neural network architecture and is good at understanding long-distance dependencies in text data, which is suitable for processing various language tasks such as generating text, translating languages, and answering questions, and can be used for content generation of various content types related to the list. In specific implementations, a corresponding large language model can be selected as needed, and the present disclosure does not limit this.

[0060] The training data required for training the large language model can be a single data point or sample in a large data set, which contains at least one question and an expected answer. During the training of the large language model, the large language model can also be fine-tuned, that is, an additional training is performed on the basis of a pre-trained large model to adapt to a specific task. The pre-trained model used can be trained using a large amount of data in advance, and the pre-trained model has learned a general feature representation. The purpose of fine-tuning is to allow the large language model to further utilize a small amount of target task data to optimize and improve.

[0061] The execution subject in the embodiments of the present disclosure can utilize the inductive reasoning capability of a large language model (LLM), complete the task of complex commodity object search by assigning roles and context information to the execution subject and equipping it with corresponding tool plugins, provide more accurate commodity object search results, make the search of commodity objects more intelligent, and improve the user experience.

[0062] For example, the search intent information can include at least one of a search intent type, commodity object description information, a commodity object category, a commodity object price, and a commodity object order quantity. The search intent type can include single-category search, multi-category search, or fuzzy search without specifying a category.

[0063] The search intent type can include single-category search (user input determines a single category), multi-category search (user actively inputs multi-category intent), or fuzzy search without specifying a category (user input or based on scene / identity, etc. to determine a broad category intent). Taking a user as a buyer of a shopping website as an example, the single-category search can correspond to the single intent of the buyer. The buyer has a single purpose or goal in the purchase decision-making process, which means that the buyer has a clear, dominant demand or problem when looking for a solution, and their decision-making process will mainly revolve around this single intent. For example, a buyer may only focus on purchasing yoga pants, which is the single intent of the buyer. The multi-category search can correspond to the multi-intent of the buyer. The buyer has multiple related goals or needs when purchasing, and they will consider multiple factors when evaluating and selecting suppliers. In this case, the decision-making process of the buyer will be more complex. For example, a buyer who is an auto parts dealer in a certain place may purchase multiple different parts according to the car brand. The multi-category search can correspond to the broad intent of the buyer. The buyer can show a more generalized and non-specific purchase goal or need in the purchase decision-making process. This intent is not as specific and clear as single intent or multi-intent, so such a buyer needs more guidance and information to help them narrow down the selection and ultimately make a more specific purchase decision. For example, a retailer may purchase different categories according to trend changes. According to different intent classifications, single-category search can correspond to a single-category search process, and multi-category search or fuzzy search without specifying a category can correspond to a search process of multiple commodity object categories.

[0064] The commodity object description information can be a property description of the commodity object, which can be in text or image form. For example, the text can be user input text or voice converted text. The commodity object category can refer to the industry or type to which the commodity object belongs. The commodity object category can be divided according to market general cognition or website historical data, and the present disclosure does not make any limitation on this. The commodity object price can be the marked price of the commodity object, such as the unit price or wholesale price. The commodity object order quantity can refer to the minimum or maximum order quantity of the commodity object, and the present disclosure does not make any limitation on this.

[0065] In a possible implementation, the large language model is a multi-layer Transformer network structure, and the training data of the large language model includes access behavior data of a commodity platform.

[0066] The multi-layer Transformer network structure can be a network trained based on a Qwen base (a basic pre-training language model, a base model). The Qwen can be a pre-training model with a multi-layer Transformer network structure. Similar to other large language models, it can handle various NLP tasks such as text generation, summarization, and question answering. Its unique feature is its flexibility and scalability, which can be optimized for specific tasks through Fine-Tuning (fine-tuning steps). In the training process, four types of models can be included, namely pre-training models (Pretrain Models), training reward models (RM Models), supervised fine-tuning models (SFT Models), and human feedback reinforcement learning models (RLHF Models). The pre-training models (Pretrain Models) can include code-specific models (Code-Qwen) and visual language models (Qwen-VL). The training reward models (RM Models) can include preference model pre-training (Qwen-PMP) and training reward models (Qwen-RM). The supervised fine-tuning models (SFT Models) can include chat models (Qwen-Chat), code-specific chat models (Code-Qwen-Chat), math-specific models (Math-Qwen-Chat), and visual language chat models (Qwen-VL-Chat). The human feedback reinforcement learning models (RLHF Models) can include human feedback reinforcement learning (Qwen-Chat-RLHF).

[0067] In the embodiments of the present disclosure, large language models (LLMs) such as GPT-4 (intelligent robots), BERT (unsupervised self-supervised models), and T5 (sequence-to-sequence models) can be natural language processing (NLP) models based on deep learning, which are pre-trained on large-scale text data and can understand, generate, and translate text. These models usually use the Transformer architecture, which performs well in handling long-range dependencies and parallelizing computations. The Transformer architecture can be composed of an Encoder and a Decoder, and the core part can be Self-Attention, which allows the model to consider the context information of the entire sequence when processing each word. Self-Attention can be used to weight the key information at different positions within the sequence, enhancing the model's ability to capture context; Positional Encoding can be used in Self-Attention to capture position information in the sequence by adding position information to each word.

[0068] In the embodiments of the present disclosure, a large language model can be trained using a supervised fine-tuning method. During training, access data including a commodity platform can be used as training data, and diversity on the data can be ensured, such as using training data of different types, different tasks, different domains, and multiple languages. In processing the training data, the training data can be first de-duplicated, and low-quality data can be removed using rule-based and machine learning methods. Multiple models can be used to score the data, and high-quality data can be added. Real data can also be used for training in the later training period. For example, expert data annotated by humans can be used to train a preliminary model, and real search data collected subsequently can be used to train the model.

[0069] In step S503, a second prompt word corresponding to the commodity object search scenario is generated based on the search basis data and the search intent information, and the large language model is interacted with through the second prompt word to obtain a search result corresponding to the current commodity object search.

[0070] The second prompt word can be generated based on the search basis data and the search intent information obtained as described above. By interacting with the large language model through the second prompt word, the user's intent can be further understood and more accurately identified based on more user and commodity related information, and a more accurate search result corresponding to the current commodity object search can be obtained.

[0071] For example, based on the search basis data, the first prompt word corresponding to the commodity object search scene can be generated in the following manner: combining the first type of prompt word template for requesting identification of the search intent and the search basis data, and generating the first prompt word corresponding to the commodity object search scene. Correspondingly, based on the search basis data and the search intent information, the second prompt word corresponding to the commodity object search scene can be generated in the following manner: combining the second type of prompt word template for requesting search of the commodity object and the search basis data, and generating the second prompt word corresponding to the commodity object search scene.

[0072] As described above, the large language model can simulate a conversation with a person, and obtain the output result by inputting the question. The input question can include an input prompt word. Different prompt words cover different amounts of information and different ways of expression, which have a great influence on the output result. However, the creation of the prompt word template can well solve these problems. However, it is still to be determined which template to use for different input contents. Therefore, the embodiments of the present disclosure can retrieve the prompt word template associated with the session message sent by the session requester of the current session from the prompt word database, and modify the session message according to the prompt word template, so that the prompt word template is adapted to the session message, thereby obtaining a better way of asking questions under the condition that the amount of information of the representation data of the session message sent by the session requester is certain, and obtaining a better output result of the large language model. After obtaining the matched prompt word template, the session message sent by the session requester can be combined with the prompt word template to generate a target prompt word for input of the large language model, so as to modify the session message. The prompt word database can be a database storing a plurality of prompt word templates. These prompt word templates can be prompt word templates created according to different conversation scenes, different conversation contents, or different conversation objects, etc. The cloud database can be used to store the prompt word templates. The present disclosure does not limit the type and storage form of the prompt word database. The session messages of different content types can have corresponding prompt word template libraries respectively. The prompt word database storing the prompt word templates can be pre-configured to generate the prompt word for input of the large language model based on the representation data of the session message sent by the session requester of the current session. The reply messages of different content types of conversation messages can call the same or different large language models to generate. One or more large language models can be called to generate the reply messages of one content type. The specific configuration can be determined according to actual needs.

[0073] In the embodiments of the present disclosure, the first prompt word can have a corresponding first prompt word template, and the manner of generating the first prompt word corresponding to the commodity object search scene can be generating the first prompt word corresponding to the commodity object search scene in combination with the first type of prompt word template for requesting to identify the search intent and the search basis data. Correspondingly, the second prompt word can have a corresponding second prompt word template, and the manner of generating the second prompt word corresponding to the commodity object search scene can be generating the second prompt word corresponding to the commodity object search scene in combination with the second type of prompt word template for requesting to search the commodity object and the search basis data.

[0074] In a possible implementation, the manner of generating the second prompt word corresponding to the commodity object search scene based on the search basis data and the search intent information can be generating the second prompt word corresponding to the commodity object search scene based on the commodity object access behavior of the search user and / or the commodity object search behavior and the search basis data and the search intent information.

[0075] The commodity object access behavior of the search user can refer to commodity source seeking, that is, seeking the source of the commodity object, for example, judging that the intent of the user is to seek the source according to the first prompt word described above, and finding the expected commodity object. The commodity object search behavior can refer to the historical search record or the expression manner, keyword, etc. when searching of a specific user or a specific commodity, which can be used as a basis to judge the commodity object category that the user is interested in, the user preference, etc.

[0076] In some embodiments, the manner of generating the second prompt word corresponding to the commodity object search scene based on the search basis data and the search intent information can first generate user behavior feature information based on the search user related user behavior information and / or user session information, and then generate the second prompt word corresponding to the commodity object search scene based on the user behavior feature information, the search basis data and the search intent information. The user behavior feature information represents the user's preference for the commodity object and / or the transportation process.

[0077] The user behavior feature information can represent the user's preference for the commodity object and / or the transportation process, and the user behavior feature information can be used as an auxiliary parameter. These parameters can be used for screening after searching, and then the second prompt word corresponding to the commodity object search scene can be generated based on the user behavior feature information, the search basis data and the search intent information.

[0078] For example, the parameter information can be the auxiliary function of the intent keyword (such as price, minimum order quantity MOQ, delivery time, cost performance, etc.), the analysis of other user behaviors except the current session, the determined consumption habits of the user, and these analyses and determined habits will become parameters transmitted to the tool, such as real-time analysis of user behavior or user historical session, etc.

[0079] Exemplarily, the manner of obtaining the search result corresponding to the current search of the commodity object by interacting with the large language model through the second prompt word can be that the second prompt word is input into the large language model, so that the large language model invokes a commodity search process corresponding to a search intent type indicated by the search intent information, and correspondingly, the commodity search process searches for commodity objects based on search basis data and search intent information, and the large language model deletes commodity objects that do not match the user behavior characteristic information from the search result; wherein the search intent type can include single-category search, multi-category search or fuzzy search of unspecified category, the single-category search corresponds to a single-category search process, and the multi-category search or the fuzzy search of unspecified category corresponds to a search process of multiple commodity object categories.

[0080] That is to say, in one manner, the large language model can narrow the range of the commodity search result generated according to the first prompt word according to the second prompt word, and delete commodity objects that do not match the user behavior characteristic information. Of course, the range of the searched commodity objects can also be expanded to find commodity objects that more match the user behavior characteristic information, and the present disclosure does not make any limitation on this.

[0081] In one possible implementation, in the case where the search intent type indicated by the search intent information is the fuzzy search of unspecified category, the large language model can further output multiple commodity object categories corresponding to the fuzzy search of unspecified category and procurement and combination purchase schemes of the multiple commodity object categories.

[0082] Taking giving a group purchase suggestion to a buyer of a shopping website as an example, in the case of the fuzzy search of unspecified category, the buyer often has procurement needs for multiple commodities, which need to be procured together or combined, such as a buyer of a manufacturing industry who may need to purchase raw materials, machine parts and safety equipment, a buyer of a catering industry who may need to order food, catering equipment and cleaning supplies, etc. In these cases, the buyer who has group purchase needs tends to find a supplier or platform that can one-stop supply multiple commodities to improve efficiency and reduce procurement costs, and therefore the large language model can be pre-trained using data of general group purchase in the industry or other data such as shopping records of buyers of the same industry on the shopping website, so that the large language model can give such group purchase suggestions, and can also give suppliers or platforms that can one-stop supply multiple commodities, that is, C-end commodities are searched while B-end suppliers are searched, so that the user can find the suggested suppliers to customize the group purchase needs.

[0083] In some embodiments, the large language model can further output comparison data of the search result in at least one dimension, the search intent information, and a commodity search process corresponding to the search intent type indicated by the search intent information.

[0084] The comparison data of the search result in at least one dimension can be comparison data of a plurality of retrieved commodity search results in dimensions of price, order quantity, delivery time, etc. The dimensions can be set in advance, and the present disclosure does not make any limitation in this regard.

[0085] The present disclosure also provides a search method for a commodity object. As shown in FIG. 6, a flowchart of a search method 600 for a commodity object according to an embodiment of the present disclosure is shown. The method 600 can include the following steps:

[0086] In step S601, search results corresponding to the current search for a commodity object are obtained according to a search request triggered by a client. The search results are obtained by interacting with a large language model using a second prompt word after the second prompt word is generated based on search basis data and search intent information corresponding to a search scenario for a commodity object. The search intent information is obtained by interacting with the large language model using a first prompt word after the first prompt word is generated based on search basis data for searching for a commodity object.

[0087] In step S602, the search results are displayed on the client.

[0088] Corresponding to the above examples and method embodiments provided by the present disclosure, the present disclosure also provides a search device for a commodity object. As shown in FIG. 7, a structural block diagram of a search device 700 for a commodity object according to an embodiment of the present disclosure is shown. The device 700 can include a search basis data obtaining module 701 configured to obtain search basis data for searching for a commodity object according to a search request triggered by a client. A search intent information obtaining module 702 is configured to generate a first prompt word corresponding to a search scenario for a commodity object based on the search basis data, and obtain search intent information corresponding to the current search for a commodity object by interacting with a large language model using the first prompt word. A search result obtaining module 703 is configured to generate a second prompt word corresponding to a search scenario for a commodity object based on the search basis data and the search intent information, and obtain search results corresponding to the current search for a commodity object by interacting with the large language model using the second prompt word.

[0089] In a possible implementation, the search basis data obtaining module 701 can include an image, voice and / or text obtaining sub-module configured to obtain an image, voice and / or text input by the client.

[0090] In a possible implementation, the search intention information obtaining module 702 can include a first prompt word generating sub-module configured to generate a first prompt word corresponding to the commodity object search scene in combination with a first type of prompt word template for requesting identification of a search intention and the search basis data. The search result obtaining module 703 can include a second prompt word generating sub-module configured to combine a second type of prompt word template, and generate a second prompt word corresponding to the commodity object search scene in combination with a second type of prompt word template for requesting search of a commodity object and the search basis data.

[0091] In a possible implementation, the search intention information includes at least one of a search intention type, commodity object description information, a commodity object category, a commodity object price, and a commodity object order quantity. The search intention type includes single-category search, multi-category search, or fuzzy search without a specified category.

[0092] In a possible implementation, the search result obtaining module 703 can include a second prompt word generating sub-module configured to generate a second prompt word corresponding to the commodity object search scene based on a commodity object access behavior and / or a commodity object search behavior of a search user, and the search basis data and the search intention information.

[0093] In a possible implementation, the search result obtaining module 703 can include a user behavior feature information obtaining sub-module configured to obtain user behavior feature information generated based on user behavior information and / or user session information related to a search user, the user behavior feature information representing a preference of the user for a commodity object and / or a transportation process. The search result obtaining module 703 can further include a second prompt word generating sub-module configured to generate a second prompt word corresponding to the commodity object search scene based on the user behavior feature information, the search basis data, and the search intention information.

[0094] In some embodiments, the search result obtaining module 703 can include a second prompt word inputting sub-module configured to input the second prompt word to the large language model, so that the large language model invokes a commodity search process corresponding to a search intention type indicated by the search intention information, the commodity search process performs commodity object search based on the search basis data and the search intention information, and the large language model deletes commodity objects that do not match the user behavior feature information from the search result. The search intention type includes single-category search, multi-category search, or fuzzy search without a specified category, the single-category search corresponds to a single-category search process, and the multi-category search or the fuzzy search without a specified category corresponds to a search process of multiple commodity object categories.

[0095] In a possible implementation, in a case where the search intention type indicated by the search intention information is a fuzzy search of an unspecified category, the large language model further outputs a plurality of commodity object categories corresponding to the fuzzy search of the unspecified category and a purchase and combination purchase scheme of the plurality of commodity object categories.

[0096] In a possible implementation, the large language model further outputs comparison data of the search result in at least one dimension, search intention information, and a commodity search process corresponding to the search intention type indicated by the search intention information.

[0097] In a possible implementation, the large language model is a multi-layer Transformer network structure, and training data of the large language model includes access behavior data of a commodity platform.

[0098] Corresponding to the above examples and method embodiments provided by the embodiments of the present disclosure, the embodiments of the present disclosure also provide a search device for a commodity object. As shown in FIG. 8, a structural block diagram of a search device 800 for a commodity object according to an embodiment of the present disclosure, the device 800 can include: a search result acquisition module 801, configured to acquire a search result corresponding to a current commodity object search according to a search request triggered based on a client; wherein the search result is obtained by interacting with a large language model through a second prompt word after the second prompt word is generated based on search basis data and search intention information corresponding to a commodity object search scenario; the search intention information is obtained by interacting with the large language model through a first prompt word after the first prompt word is generated based on search basis data for searching a commodity object corresponding to a commodity object search scenario; and a search result display module 802, configured to display the search result based on the client.

[0099] The functions of each module in each device of the embodiments of the present disclosure can be referred to the corresponding description in the above method, and have the corresponding beneficial effects, which will not be repeated here.

[0100] FIG. 9 is a block diagram of an electronic device for implementing the embodiments of the present disclosure. As shown in FIG. 9, the electronic device includes a memory 901 and a processor 902, and the memory 901 stores a computer program that can run on the processor 902. The processor 902 implements the method in the above embodiments when executing the computer program. The number of the memory 901 and the processor 902 can be one or more.

[0101] The electronic device further includes:

[0102] A communication interface 903 is configured to communicate with external devices and perform data interaction transmission.

[0103] If the memory 901, the processor 902 and the communication interface 903 are independently implemented, the memory 901, the processor 902 and the communication interface 903 can be connected with each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, only one thick line is used in FIG. 9, but it does not mean that there is only one bus or only one type of bus.

[0104] Optionally, in a specific implementation, if the memory 901, the processor 902 and the communication interface 903 are integrated on a chip, the memory 901, the processor 902 and the communication interface 903 can complete communication between each other through an internal interface.

[0105] The embodiment of the present disclosure provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the method provided in the embodiment of the present disclosure.

[0106] The embodiment of the present disclosure also provides a computer program product, which includes a computer program, and the computer program is executed by a processor to implement the method provided in any embodiment of the present disclosure.

[0107] The embodiment of the present disclosure also provides a chip, which includes a processor, is used for calling and running instructions stored in a memory, so that a communication device installed with the chip executes the method provided in the embodiment of the present disclosure.

[0108] The embodiment of the present disclosure also provides a chip, which includes an input interface, an output interface, a processor and a memory, the input interface, the output interface, the processor and the memory are connected through an internal connection path, and the processor is used for executing code in the memory, and when the code is executed, the processor is used for executing the method provided in the embodiment of the present disclosure.

[0109] It is to be understood that the above-mentioned processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. It is to be noted that the processor can be a processor supporting an advanced RISC machine (ARM) architecture.

[0110] Further, the memory can include a read-only memory and a random access memory, optionally. The memory can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. The non-volatile memory can include a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically EPROM (EEPROM), or a flash memory. The volatile memory can include a random access memory (RAM) used as an external cache. By way of example, but not limitation, many forms of RAM are available. For example, a static random access memory (SRAM), a dynamic random access memory (DRAM), a synchronous DRAM (SDRAM), a double data rate SDRAM (DDR SDRAM), an enhanced SDRAM (ESDRAM), a SyncLink DRAM (SLDRAM), and a direct Rambus RAM (DR RAM) can be used.

[0111] In the above-described embodiments, all or part can be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed by a computer, all or part generates the flow or function according to the present disclosure. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium.

[0112] In the description of the present disclosure, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. Also, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, different embodiments or examples described in the present disclosure and features of different embodiments or examples can be combined and combined by those skilled in the art without contradiction.

[0113] In addition, the terms "first", "second", etc. are used only for the purpose of description and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present disclosure, the meaning of "a plurality of" is two or more, unless otherwise explicitly specified.

[0114] Any process or method described in the flowchart or otherwise described herein can be understood as a representation of code, including one or more executable instructions for implementing the specific logical functions or steps, modules, segments or portions. And the scope of the preferred embodiments of the present disclosure includes additional implementations, in which the functions can be performed in an order other than that shown or discussed, including in a substantially simultaneous manner according to the functions involved or in reverse order.

[0115] The logic and / or steps described in the flowchart or otherwise described herein, for example, can be considered as a list of executable instructions for implementing the logic function, which can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus, such as a computer-based system, a system including a processor, or other system that can take instructions from an instruction execution system, device or apparatus, or in conjunction with these instructions execution system, device or apparatus.

[0116] It should be understood that each part of the present disclosure can be implemented in hardware, software, firmware or a combination thereof. In the above embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. All or part of the steps of the above-mentioned embodiment method can be completed by a program instructing the relevant hardware, which can be stored in a computer readable storage medium, and the program includes one of the steps of the method embodiment or a combination thereof when executed.

[0117] In addition, each functional unit in various embodiments of the present disclosure can be integrated into one processing module, or each unit can exist physically independently, or two or more units can be integrated into one module. The above integrated module can be implemented in the form of hardware or in the form of a software functional module. The above integrated module, if implemented in the form of a software functional module and sold or used as an independent product, can also be stored in a computer readable storage medium. The storage medium can be a read-only memory, a magnetic disk or an optical disk, etc.

[0118] The above is only exemplary embodiments of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any person skilled in the art can easily think of various changes or replacements within the technical scope disclosed in the present disclosure, and these should be covered within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. A search method of commodity objects, comprising: obtaining search basis data for searching commodity objects according to a search request triggered by a client; generating a first prompt word corresponding to a commodity object search scene based on the search basis data, and interacting with a large language model through the first prompt word to obtain search intent information corresponding to the current commodity object search; generating a second prompt word corresponding to the commodity object search scene based on the search basis data and the search intent information, and interacting with the large language model through the second prompt word to obtain a search result corresponding to the current commodity object search.

2. The method of claim 1, wherein, The search basis data for searching commodity objects is obtained according to a search request triggered by a client, comprising: obtaining an image, voice and / or text input by the client.

3. The method of claim 1 or 2, wherein, The first prompt word corresponding to the commodity object search scene is generated based on the search basis data, comprising: generating a first prompt word corresponding to a commodity object search scene in combination with a first type of prompt word template for requesting identification of search intent and the search basis data; The second prompt word corresponding to the commodity object search scene is generated based on the search basis data and the search intent information, comprising: generating a second prompt word corresponding to a commodity object search scene in combination with a second type of prompt word template for requesting search of commodity objects and the search basis data.

4. The method of any one of claims 1-3, wherein, The search intent information includes at least one of search intent type, commodity object description information, commodity object category, commodity object price, and commodity object order quantity. The search intent type includes single category search, multi-category search, or fuzzy search without specifying a category.

5. The method of claim 1 or 2, wherein, The second prompt word corresponding to the commodity object search scene is generated based on the search basis data and the search intent information, comprising: generating a second prompt word corresponding to a commodity object search scene based on search user's commodity object access behavior and / or commodity object search behavior, and search basis data and search intent information.

6. The method of claim 1 or 2, wherein, The second prompt word corresponding to the commodity object search scene is generated based on the search basis data and the search intent information, comprising: obtaining user behavior feature information generated based on search user-related user behavior information and / or user session information, the user behavior feature information representing the user's preference for commodity objects and / or transportation process; generating a second prompt word corresponding to a commodity object search scene based on the user behavior feature information, search basis data and search intent information.

7. The method of claim 6, wherein, The search result corresponding to the current commodity object search is obtained by interacting with the large language model through the second prompt word, comprising: inputting the second prompt word into the large language model to enable the large language model to call a commodity search process corresponding to the search intent type indicated by the search intent information, the commodity search process performing commodity object search based on the search basis data and search intent information, and the large language model deleting commodity objects from the search result that do not match the user behavior feature information; The search intent type includes single-category search, multi-category search, or fuzzy search without specifying a category. The single-category search corresponds to a single-category search process, and the multi-category search or the fuzzy search without specifying a category corresponds to a search process of multiple product object categories.

8. The method of any one of claims 1-7, wherein, In a case where the search intent type indicated by the search intent information is the fuzzy search without specifying a category, the large language model further outputs multiple product object categories corresponding to the fuzzy search without specifying a category and a purchase and combination purchase scheme of the multiple product object categories.

9. The method of any one of claims 1-8, wherein, The large language model further outputs comparison data of the search result in at least one dimension, search intent information, and a product search process corresponding to the search intent type indicated by the search intent information.

10. The method of any one of claims 1-9, wherein, The large language model is a multi-layer Transformer network structure, and training data of the large language model includes access behavior data of a product platform.

11. A search method of a product object, comprising: obtaining a search result corresponding to the current product object search according to a search request triggered by a client; wherein the search result is obtained by interacting with a large language model through a second prompt word after the second prompt word is generated based on search basis data and search intent information corresponding to a product object search scenario; and the search intent information is obtained by interacting with the large language model through a first prompt word after the first prompt word is generated based on search basis data for searching for a product object; displaying the search result based on the client.

12. A search device of a product object, comprising: a search basis data obtaining module configured to obtain search basis data for searching for a product object according to a search request triggered by a client; a search intent information obtaining module configured to generate a first prompt word corresponding to a product object search scenario based on the search basis data, and obtain search intent information corresponding to the current product object search by interacting with a large language model through the first prompt word; a search result obtaining module configured to generate a second prompt word corresponding to a product object search scenario based on the search basis data and the search intent information, and obtain a search result corresponding to the current product object search by interacting with the large language model through the second prompt word.

13. A search device of a product object, comprising: a search result obtaining module configured to obtain a search result corresponding to the current product object search according to a search request triggered by a client; wherein the search result is obtained by interacting with a large language model through a second prompt word after the second prompt word is generated based on search basis data and search intent information corresponding to a product object search scenario; and the search intent information is obtained by interacting with the large language model through a first prompt word after the first prompt word is generated based on search basis data for searching for a product object; a search result display module configured to display the search result based on the client.

14. An electronic device comprising a memory, a processor, and a computer program stored on the memory, the processor implementing the method of any one of claims 1-11 when executing the computer program.

15. A computer readable storage medium having stored therein a computer program, the computer program implementing the method of any one of claims 1-11 when executed by a processor.

16. A computer program product comprising a computer program, characterized in that, The computer program implements the method of any one of claims 1-11 when executed by a processor.

Images