Large language model smart search grouping method, system and device and medium

Abstract

The invention relates to the technical field of internet retrieval, and particularly discloses a big language model smart search grouping method, system and device and a medium, the method comprises the steps of obtaining a search request submitted by a user, and extracting search state features used for representing a user search state; determining a preset number of search objects related to the search request based on the search state characteristics, and synchronously generating description information of each search object; constructing a search operation matrix, calculating a significance parameter of each search object at the current moment based on the search operation matrix, and updating a grouping weight of each search object; determining the search processing resource quantity of each search object according to the grouping weight, calling a large language model based on the search processing resource quantity, and generating and feeding back a search result; according to the method, differentiated search contents can be output for different search objects, and the coverage degree and the response accuracy of the search result to the actual search requirement of the user are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of Internet retrieval technology, specifically to a method, system, device, and medium for intelligent search grouping based on a large language model. Background Technology

[0002] In the context of internet technology in the era of big data, search functionality is a fundamental function of the vast majority of internet services. When responding to user search requests, existing search systems typically focus on a single search target, uniformly parsing the search request and returning a linearly sorted list of search results, or simply performing basic topic categorization during the results display phase. This search approach makes it difficult to simultaneously characterize the differences between multiple potential search objects in a single search process. Users often need to conduct multiple searches or repeatedly adjust their search criteria to gradually obtain search feedback from different directions or with different focuses, resulting in low search interaction efficiency.

[0003] In complex search scenarios, the same search request often corresponds to multiple semantic directions or potential search objects, such as different product categories, different usage scenarios, or different information levels. How to expand the feedback results of search needs to meet users' needs for multi-dimensional and multi-perspective search feedback is the technical problem that the present invention aims to solve. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, device, and medium for intelligent search grouping of large language models to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A smart search grouping method for large language models, the method comprising:

[0007] Get the search request submitted by the user, get the historical search behavior data corresponding to the search request, extract features from the search request and the historical search behavior data, and obtain search status features to characterize the user's search status.

[0008] Based on the search status characteristics, a preset number of search objects related to the search request are determined, and description information for each search object is generated synchronously; the description information includes at least the semantic and behavioral characteristics of the search object.

[0009] Behavioral data of each search object is collected at different time points based on a preset time period, and a search operation matrix is ​​constructed with time as the row and behavioral features as the column; the search operation matrix is ​​used to characterize the changes in the behavior of the search object over time.

[0010] The significance parameters of each search object at the current time are calculated based on the search operation matrix, and the grouping weights of each search object are updated according to the significance parameters.

[0011] The amount of search processing resources for each search object is determined based on the grouping weights. The large language model is then invoked based on the amount of search processing resources to generate and return search results. The amount of search processing resources is used at least to adjust the search expansion depth and search priority.

[0012] As a further aspect of the present invention: the steps of obtaining a user-submitted search request, obtaining historical search behavior data corresponding to the search request, and extracting features from the search request and historical search behavior data to obtain search status features characterizing the user's search status include:

[0013] Obtain user-submitted search requests and historical search behavior data;

[0014] The search request and historical search behavior data are preprocessed to obtain word groups; the text preprocessing includes at least word segmentation, noise reduction and keyword extraction.

[0015] The word groups corresponding to the search request are transformed to obtain semantic vectors, and the word groups corresponding to historical search behavior data are transformed to obtain behavioral features;

[0016] Statistical semantic vectors and behavioral features are used as search state features;

[0017] The semantic vectors and behavioral features are both text converted into numerical format.

[0018] As a further aspect of the present invention: the step of determining a preset number of search objects related to the search request based on the search state characteristics, and synchronously generating description information for each search object includes:

[0019] Based on the search status characteristics, recall a preset number of candidate search objects from a preset knowledge base;

[0020] Extract the semantic features of the content for each candidate search object;

[0021] Extract historical behavioral characteristics of the search object from the search logs;

[0022] By combining statistical semantic features and behavioral features, descriptive information about the search object can be obtained.

[0023] As a further aspect of the present invention: the step of collecting behavioral data of each search object at different time points based on a preset time period, and constructing a search operation matrix with time as the row and behavioral features as the column, includes:

[0024] Receive a set of behavioral features input from the management to characterize the behavior of the search object;

[0025] Behavioral data of each search object at different time points are collected based on a preset time period; the data type of the behavioral data is contained in a behavioral feature set;

[0026] Construct a matrix template using time as the row index and behavioral features as the column index;

[0027] By filling the matrix template with behavioral data from different time points, the search operation matrix is ​​obtained.

[0028] As a further aspect of the present invention: the step of calculating the significance parameter of each search object at the current time based on the search operation matrix, and updating the grouping weight of each search object according to the significance parameter includes:

[0029] Extract a submatrix containing the current time step from the search execution matrix;

[0030] Normalize the behavioral features in the submatrix;

[0031] Each behavioral feature is weighted according to a preset weight to obtain a comprehensive score;

[0032] The overall score is used as the salience parameter of the search object at the current moment;

[0033] Update the weight values ​​of the search objects in the grouping process based on the significance parameter.

[0034] As a further aspect of the present invention: the step of determining the search processing resource amount for each search object based on the grouping weight, and calling the large language model based on the search processing resource amount to generate and return the search results includes:

[0035] The behavior changes of each search object in different time periods are analyzed based on the search operation matrix, and the behavior deviation is calculated.

[0036] The behavioral deviation is used as a correction factor for the group weights, and the group weights are dynamically corrected to obtain the corrected group weights.

[0037] The search processing resource amount for each search object is determined based on the corrected grouping weights. Based on the search processing resource amount, the large language model is called to generate and return the search results.

[0038] Among them, behavioral deviation is used to characterize the degree of abnormality of the behavioral changes of the search object relative to the historical average behavior.

[0039] The present invention also provides a large language model intelligent search grouping system, the system comprising:

[0040] The search feature extraction module is used to obtain the search request submitted by the user, obtain the historical search behavior data corresponding to the search request, extract features from the search request and the historical search behavior data, and obtain search status features that characterize the user's search status.

[0041] The search object analysis module is used to determine a preset number of search objects related to the search request based on the search status characteristics, and synchronously generate description information for each search object; the description information includes at least the semantic features and behavioral features of the search object.

[0042] The search behavior analysis module is used to collect behavioral data of each search object at different time points based on a preset time period, and construct a search operation matrix with time as the row and behavioral features as the column; the search operation matrix is ​​used to characterize the changes in the behavior of the search object over time.

[0043] The group weight determination module is used to calculate the significance parameter of each search object at the current time based on the search operation matrix, and update the group weight of each search object according to the significance parameter;

[0044] The search result feedback module is used to determine the amount of search processing resources for each search object based on the grouping weight, call the large language model based on the amount of search processing resources, generate and feed back the search results; wherein, the amount of search processing resources is used at least to adjust the search expansion depth and search priority.

[0045] As a further aspect of the present invention: the search feature extraction module includes:

[0046] The data acquisition unit is used to acquire user-submitted search requests and historical search behavior data;

[0047] The preprocessing unit is used to perform text preprocessing on search requests and historical search behavior data to obtain word groups; the text preprocessing includes at least word segmentation, noise reduction and keyword extraction.

[0048] The text conversion unit is used to convert the word groups corresponding to the search request to obtain semantic vectors, and to convert the word groups corresponding to historical search behavior data to obtain behavioral features;

[0049] The first feature statistical unit is used to statistically analyze semantic vectors and behavioral features as search state features;

[0050] The semantic vectors and behavioral features are both text converted into numerical format.

[0051] The present invention also provides a computer device, which includes one or more processors and one or more memories. The one or more memories store at least one piece of program code. When the program code is loaded and executed by the one or more processors, it implements the large language model intelligent search grouping method.

[0052] The present invention also provides a storage medium storing at least one piece of program code, which, when loaded and executed by a processor, implements the large language model intelligent search grouping method.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention extracts joint features from search request and historical search behavior data to determine a preset number of search objects related to the search request, and constructs a search operation matrix based on the semantic and behavioral features of the search objects to achieve dynamic characterization of the behavior changes of the search objects. By calculating the salience parameters of the search objects at the current moment and updating the group weights, the search objects form a clear group structure in the search results, thereby simultaneously outputting search feedback results corresponding to multiple search objects in a single search request. It can output differentiated search content for different search objects, greatly improving the coverage and response accuracy of search results to the user's actual search needs. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0055] Figure 1 This is a flowchart of a grouping method for intelligent search of a large language model.

[0056] Figure 2 This is a block diagram of the composition structure of a large language model intelligent search grouping system. Detailed Implementation

[0057] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0058] Figure 1 The flowchart illustrates a grouping method for intelligent search of a large language model. In this embodiment of the invention, a grouping method for intelligent search of a large language model includes:

[0059] Step S100: Obtain the search request submitted by the user, obtain the historical search behavior data corresponding to the search request, extract features from the search request and the historical search behavior data, and obtain search status features to characterize the user's search status.

[0060] The search request refers to the original query information entered by the user when conducting a search. It is generally text. After receiving the user's search request, the system retrieves the historical search behavior data corresponding to the search request. This historical search behavior data refers to the data set generated by the user during the historical search process, which represents the evolution of the user's search intent and search interaction characteristics. In addition to historical search request data, it also includes search result interaction data and search context data. Furthermore, the scope of acquisition is not limited to the user's own historical search process. Based on pre-obtained data sharing permissions, it can also include the historical search data of all users with the same search request within a certain period of time. Feature extraction is performed on the search request and historical search behavior data to obtain search state features that represent the user's search state. The feature extraction process is mainly a text feature extraction process, and occasionally it involves feature extraction processes for other structures, such as images or audio. These are all conventional data processing processes and will not be elaborated here.

[0061] Step S200: Based on the search status features, determine a preset number of search objects related to the search request, and synchronously generate description information for each search object; the description information includes at least the semantic features and behavioral features of the search object;

[0062] Search status features represent the demand information (including some potential related demand information) corresponding to a user's search request. It is a large data volume. Based on the search status features, the search object is matched in the knowledge base (the data space of the response data), and the description information of each search object is queried synchronously. The description information includes at least the semantic features and behavioral features of the search object. In fact, in general, it only includes these two contents. The description information can also be understood as a collection of text content.

[0063] Step S300: Collect behavioral data of each search object at different time points based on a preset time period, and construct a search operation matrix with time as the row and behavioral features as the column; the search operation matrix is ​​used to characterize the changes in the behavior of the search object over time;

[0064] Then, using the search object as a benchmark, the behavioral data of the search object at different time points within a preset time period is obtained, and a search operation matrix is ​​constructed with time as the row and behavioral features as the column. Specifically, the behavioral data is generally determined by the description information of the search object and the search behavior log. These data are standardized to obtain analyzable time series structured data, such as click count, click-through rate, dwell time, and conversion behavior.

[0065] In the above content, the behavioral data of the search object is the statistical result of the search behavior of multiple users for the same search object. The statistical result is used to characterize the behavior of the search object in the overall search system, and it is not limited to the behavior of the current user. For example, when the search object is "a certain model of air purifier", the behavioral data source includes the click, dwell and conversion data of different users at different time periods.

[0066] Step S400: Calculate the significance parameter of each search object at the current time based on the search operation matrix, and update the grouping weight of each search object according to the significance parameter;

[0067] The search execution matrix is ​​analyzable user behavior data over a period of time. By performing matrix identification on the search execution matrix, parameters reflecting the importance of the search object are extracted, called saliency parameters. The saliency parameters are then numerically transformed to obtain the grouping weights used for grouping.

[0068] Step S500: Determine the amount of search processing resources for each search object based on the grouping weights, call the large language model based on the amount of search processing resources, generate and return the search results; wherein, the amount of search processing resources is used at least to adjust the search expansion depth and search priority.

[0069] The administrator pre-sets several weight ranges, with common weight ranges including low-weight reorganization, medium-weight reorganization, and high-weight reorganization. Different processing resource amounts are allocated to different weight ranges. After grouping the search objects based on the weight ranges, each group of search objects is processed using the corresponding processing resource amount. Specifically, the processing method involves calling a large language model based on the search processing resource amount to generate and return search results. The search processing resource amount is used at least to adjust the search expansion depth and search priority, thereby adjusting the speed and depth of the feedback results from the large language model.

[0070] In the above architecture, after a user enters a search request, they can get feedback results of search objects with different levels of relevance, and the data range of the feedback results is extremely wide.

[0071] Regarding step S100, the steps of obtaining the user-submitted search request, obtaining historical search behavior data corresponding to the search request, and extracting features from the search request and historical search behavior data to obtain search status features that characterize the user's search status include:

[0072] Obtain user-submitted search requests and historical search behavior data;

[0073] The search request and historical search behavior data are preprocessed to obtain word groups; the text preprocessing includes at least word segmentation, noise reduction and keyword extraction.

[0074] The word groups corresponding to the search request are transformed to obtain semantic vectors, and the word groups corresponding to historical search behavior data are transformed to obtain behavioral features;

[0075] Statistical semantic vectors and behavioral features are used as search state features;

[0076] The semantic vectors and behavioral features are both text converted into numerical format.

[0077] In one example of the technical solution of this invention, the process of determining search status features is described. The process involves acquiring user-submitted search requests and historical search behavior data, both of which are text data. The search requests and historical search behavior data undergo text preprocessing to obtain word groups. This text preprocessing includes at least word segmentation, noise reduction, and keyword extraction. Then, the search requests and historical search behavior data are analyzed separately. The word groups in the search requests are transformed to obtain semantic vectors, and the word groups corresponding to the historical search behavior data are transformed to obtain behavioral features. The semantic vectors and behavioral features are statistically analyzed and used as search status features. Here, both the semantic vectors and behavioral features are converted numerical text formats, such as word vectors.

[0078] Regarding step S200, the step of determining a preset number of search objects related to the search request based on the search state features, and synchronously generating description information for each search object, includes:

[0079] Based on the search status characteristics, recall a preset number of candidate search objects from a preset knowledge base;

[0080] Extract the semantic features of the content for each candidate search object;

[0081] Extract historical behavioral characteristics of the search object from the search logs;

[0082] By combining statistical semantic features and behavioral features, descriptive information about the search object can be obtained.

[0083] In one example of the technical solution of this invention, the process of determining the search object is described. The search state feature is a data structure in numerical format, which is used as a feature to match the search object in the knowledge base. The number of matched search objects is preset, and the matched search objects are called candidate search objects. The content semantic features of each candidate search object are extracted, and then the historical behavior features of the search object are extracted from the search log. The semantic features and behavior features are statistically analyzed to obtain the description information of the search object. The processes of extracting content semantic features and extracting historical behavior features are both conventional text processing processes, which do not need to be elaborated. However, for the search object, it is actually the smallest content unit that can be recalled, analyzed, compared, grouped, and output in the search system. This description is too abstract. Specifically, it is "content that can be displayed to the user", which can include text content, structured information, behavioral statistics data, and semantic vectors.

[0084] Regarding the content semantic features and historical behavior features mentioned above, content semantic features refer to the feature representations used to characterize the semantic meaning of the search object after semantic modeling of its own content. These can include text semantic vectors, which are similar to the semantic vectors mentioned above, and can also include topic distribution features, which are slightly more complex but are essentially word features that can be obtained using existing text processing techniques. Similarly, historical behavior features refer to the features used to characterize the behavior of the search object, obtained based on statistics of users' historical interaction behavior with the search object; for example, click count, click-through rate (click count / impression count), average dwell time, repeat visit rate, and conversion trigger rate.

[0085] Regarding step S300, the step of collecting behavioral data of each search object at different time points based on a preset time period, and constructing a search operation matrix with time as the row and behavioral features as the column, includes:

[0086] Receive a set of behavioral features input from the management to characterize the behavior of the search object;

[0087] Behavioral data of each search object at different time points are collected based on a preset time period; the data type of the behavioral data is contained in a behavioral feature set;

[0088] Construct a matrix template using time as the row index and behavioral features as the column index;

[0089] By filling the matrix template with behavioral data from different time points, the search operation matrix is ​​obtained.

[0090] In one example of the technical solution of this invention, the process of creating a search operation matrix is ​​described. The system receives a set of behavioral features input by the administrator to characterize the behavior of the search objects. Within a preset time period, behavioral data belonging to the set of behavioral features for each search object at different time points are collected. Using time as the row number and behavioral features as the column number, a matrix template is constructed. The behavioral data at different time points are then filled into the matrix template to obtain the search operation matrix. This process is actually based on the preset data template to statistically analyze the behavioral data, thereby obtaining matrix-form behavioral data related to the search objects.

[0091] Regarding step S400, the step of calculating the significance parameter of each search object at the current time based on the search operation matrix, and updating the grouping weight of each search object according to the significance parameter includes:

[0092] Extract a submatrix containing the current time step from the search execution matrix;

[0093] Normalize the behavioral features in the submatrix;

[0094] Each behavioral feature is weighted according to a preset weight to obtain a comprehensive score;

[0095] The overall score is used as the salience parameter of the search object at the current moment;

[0096] Update the weight values ​​of the search objects in the grouping process based on the significance parameter.

[0097] In one example of the technical solution of this invention, the process of determining the group weights is described. A sub-matrix containing the current time is extracted from the search operation matrix, and the behavioral features in the sub-matrix are normalized. This process ensures that the range of each value is within a preset interval, which is very convenient for subsequent calculations. Each behavioral feature is weighted according to the preset weights to obtain a comprehensive score. The comprehensive score is used as the significance parameter of the search object at the current time. Based on the significance parameter, the weight value of the search object in the grouping process is updated.

[0098] One feasible solution for the calculation process described above is as follows:

[0099] Centered on the current moment, select a preset number of time points, extract the data corresponding to the selected time points, and normalize them. The normalization method can be as follows:

[0100] In the formula, and These represent the minimum and maximum values ​​of the data, respectively. This represents the value after normalization. This represents the value before normalization. For the normalized values, the values ​​are summed using the preset weights of each parameter to obtain a comprehensive score, which serves as a significance parameter. The significance parameter needs to be mapped to the group weights, which can be done using a linear mapping method with preset mapping parameters.

[0101] As a preferred embodiment of the technical solution of the present invention, the steps of determining the search processing resource amount for each search object based on the grouping weight, calling the large language model based on the search processing resource amount, and generating and feeding back the search results include:

[0102] The behavior changes of each search object in different time periods are analyzed based on the search operation matrix, and the behavior deviation is calculated.

[0103] The behavioral deviation is used as a correction factor for the group weights, and the group weights are dynamically corrected to obtain the corrected group weights.

[0104] The search processing resource amount for each search object is determined based on the corrected grouping weights. Based on the search processing resource amount, the large language model is called to generate and return the search results.

[0105] Among them, behavioral deviation is used to characterize the degree of abnormality of the behavioral changes of the search object relative to the historical average behavior.

[0106] In one embodiment of the technical solution of this invention, an additional auxiliary scheme is provided for updating the group weights. Specifically, other parameters are introduced to adjust the group weights. The processing scheme is as follows: based on the search operation matrix, the behavior changes of each search object in different time periods are analyzed, and the behavior deviation is calculated; the behavior deviation is used as a correction factor for the group weights, and the group weights are dynamically corrected to obtain the corrected group weights; wherein, the behavior deviation is used to characterize the degree of abnormality of the behavior changes of the search objects relative to the historical average behavior.

[0107] After obtaining the corrected group weights, the group weight intervals are redefined, and the search processing strategy for each group weight interval is determined simultaneously. Then, based on the corrected group weights, each search object is mapped to a preset group weight interval. For search objects in different group weight intervals, the corresponding search processing strategies are determined respectively. Among them, the group weight interval includes at least a low weight interval, a medium weight interval, and a high weight interval. Different group weight intervals correspond to different search expansion depths, result grouping granularities, or search priorities.

[0108] After determining the search processing strategy for each search object, the search objects are filtered and stratified according to the group weight interval. Search objects in the high-weight interval and medium-weight interval are selected as key analysis objects, and input prompts are constructed and input into the large language model. The large language model performs semantic analysis and search path deduction on the key analysis objects, generates and feeds back the search results. The large language model is an existing model, and the accuracy and speed of the search results obtained vary depending on the amount of processing resources applied. At this time, one search request corresponds to multiple grouped search objects, and each search object corresponds to a feedback result of different depths. The final result is a grouped feedback result, which is highly referential.

[0109] Regarding the behavioral deviation mentioned above, the behavioral deviation is used to measure the degree of deviation of the current behavioral pattern of the search object from its historical stable behavioral pattern. The specific calculation process is as follows: determine the historical baseline vector of the search object, and for the current behavioral vector, directly calculate the vector distance. After the vector distance is calculated, introduce a linear mapping to obtain the correction factor. Here, the behavioral vector itself is a set of values, and its dimension is not limited, but the data structure must be consistent. That is, the data structure describing the behavior needs to be the same.

[0110] Figure 2 This is a block diagram illustrating the structural composition of a large language model intelligent search grouping system. As a preferred embodiment of the technical solution of this invention, the invention also provides a large language model intelligent search grouping system, wherein system 10 includes:

[0111] The search feature extraction module 11 is used to obtain the search request submitted by the user, obtain the historical search behavior data corresponding to the search request, extract features from the search request and the historical search behavior data, and obtain search status features that characterize the user's search status.

[0112] The search object analysis module 12 is used to determine a preset number of search objects related to the search request based on the search status characteristics, and synchronously generate description information for each search object; the description information includes at least the semantic features and behavioral features of the search object.

[0113] Search behavior analysis module 13 is used to collect behavioral data of each search object at different time points based on a preset time period, and construct a search operation matrix with time as the row and behavioral features as the column; the search operation matrix is ​​used to characterize the changes in the behavior of the search object over time.

[0114] The group weight determination module 14 is used to calculate the significance parameter of each search object at the current time based on the search operation matrix, and update the group weight of each search object according to the significance parameter.

[0115] The search result feedback module 15 is used to determine the amount of search processing resources for each search object based on the grouping weight, call the large language model based on the amount of search processing resources, generate and feed back the search results; wherein, the amount of search processing resources is used at least to adjust the search expansion depth and search priority.

[0116] Furthermore, the search feature extraction module 11 includes:

[0117] The data acquisition unit is used to acquire user-submitted search requests and historical search behavior data;

[0118] The preprocessing unit is used to perform text preprocessing on search requests and historical search behavior data to obtain word groups; the text preprocessing includes at least word segmentation, noise reduction and keyword extraction.

[0119] The text conversion unit is used to convert the word groups corresponding to the search request to obtain semantic vectors, and to convert the word groups corresponding to historical search behavior data to obtain behavioral features;

[0120] The first feature statistical unit is used to statistically analyze semantic vectors and behavioral features as search state features;

[0121] The semantic vectors and behavioral features are both text converted into numerical format.

[0122] Specifically, the search object analysis module 12 includes:

[0123] The object recall unit is used to recall a preset number of candidate search objects from a preset knowledge base based on search status characteristics.

[0124] The content feature extraction unit is used to extract the semantic features of each candidate search object.

[0125] The behavioral feature extraction unit is used to extract the historical behavioral features of the search object from the search log;

[0126] The second feature statistical unit is used to statistically analyze semantic and behavioral features to obtain descriptive information about the search object.

[0127] Furthermore, the search behavior analysis module 13 includes:

[0128] The behavior space receiving unit is used to receive a set of behavioral features input by the management to characterize the behavior of the search object;

[0129] A behavior data acquisition unit is used to collect behavior data of each search object at different time points based on a preset time period; the data type of the behavior data is contained in a set of behavior features.

[0130] The matrix template building unit is used to construct a matrix template using time as the row index and behavioral features as the column index.

[0131] The data filling unit is used to fill the matrix template with behavioral data from different time points to obtain the search operation matrix.

[0132] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent search grouping of large language models, characterized in that, The method includes: Get the search request submitted by the user, get the historical search behavior data corresponding to the search request, extract features from the search request and the historical search behavior data, and obtain search status features to characterize the user's search status. Based on the search status characteristics, a preset number of search objects related to the search request are determined, and description information for each search object is generated synchronously; the description information includes at least the semantic and behavioral characteristics of the search object. Behavioral data of each search object is collected at different time points based on a preset time period, and a search operation matrix is ​​constructed with time as the row and behavioral features as the column; the search operation matrix is ​​used to characterize the changes in the behavior of the search object over time. The significance parameters of each search object at the current time are calculated based on the search operation matrix, and the grouping weights of each search object are updated according to the significance parameters. The amount of search processing resources for each search object is determined based on the grouping weights. The large language model is then invoked based on the amount of search processing resources to generate and return search results. The amount of search processing resources is used at least to adjust the search expansion depth and search priority.

2. The intelligent search grouping method for large language models according to claim 1, characterized in that, The steps of obtaining the user-submitted search request, obtaining historical search behavior data corresponding to the search request, and extracting features from the search request and historical search behavior data to obtain search status features that characterize the user's search status include: Obtain user-submitted search requests and historical search behavior data; The search request and historical search behavior data are preprocessed to obtain word groups; the text preprocessing includes at least word segmentation, noise reduction and keyword extraction. The word groups corresponding to the search request are transformed to obtain semantic vectors, and the word groups corresponding to the historical search behavior data are transformed to obtain behavioral features; Statistical semantic vectors and behavioral features are used as search state features; The semantic vectors and behavioral features are both text converted into numerical format.

3. The intelligent search grouping method for large language models according to claim 1, characterized in that, The step of determining a preset number of search objects related to the search request based on the search state features, and synchronously generating description information for each search object, includes: Based on the search status characteristics, recall a preset number of candidate search objects from a preset knowledge base; Extract the semantic features of the content for each candidate search object; Extract historical behavioral characteristics of the search object from the search logs; By combining statistical semantic features and behavioral features, descriptive information about the search object can be obtained.

4. The intelligent search grouping method for large language models according to claim 1, characterized in that, The step of collecting behavioral data of each search object at different time points based on a preset time period, and constructing a search operation matrix with time as the row and behavioral features as the column, includes: Receive a set of behavioral features input from the management to characterize the behavior of the search object; Behavioral data of each search object at different time points are collected based on a preset time period; the data type of the behavioral data is contained in a behavioral feature set; Construct a matrix template using time as the row index and behavioral features as the column index; By filling the matrix template with behavioral data from different time points, the search operation matrix is ​​obtained.

5. The intelligent search grouping method for a large language model according to claim 1, characterized in that, The step of calculating the significance parameter of each search object at the current time based on the search operation matrix, and updating the grouping weight of each search object according to the significance parameter includes: Extract a submatrix containing the current time step from the search execution matrix; Normalize the behavioral features in the submatrix; Each behavioral feature is weighted according to a preset weight to obtain a comprehensive score; The overall score is used as the salience parameter of the search object at the current moment; Update the weight values ​​of the search objects in the grouping process based on the significance parameter.

6. The intelligent search grouping method for a large language model according to claim 5, characterized in that, The steps of determining the search processing resource amount for each search object based on the grouping weight, and calling the large language model based on the search processing resource amount to generate and return the search results include: The behavior changes of each search object in different time periods are analyzed based on the search operation matrix, and the behavior deviation is calculated. The behavioral deviation is used as a correction factor for the group weights, and the group weights are dynamically corrected to obtain the corrected group weights. The search processing resource amount for each search object is determined based on the corrected grouping weights. Based on the search processing resource amount, the large language model is called to generate and return the search results. Among them, behavioral deviation is used to characterize the degree of abnormality of the behavioral changes of the search object relative to the historical average behavior.

7. A large language model intelligent search grouping system, characterized in that, The system includes: The search feature extraction module is used to obtain the search request submitted by the user, obtain the historical search behavior data corresponding to the search request, extract features from the search request and the historical search behavior data, and obtain search status features to characterize the user's search status. The search object analysis module is used to determine a preset number of search objects related to the search request based on the search status characteristics, and synchronously generate description information for each search object; the description information includes at least the semantic features and behavioral features of the search object. The search behavior analysis module is used to collect behavioral data of each search object at different time points based on a preset time period, and construct a search operation matrix with time as the row and behavioral features as the column; the search operation matrix is ​​used to characterize the changes in the behavior of the search object over time. The group weight determination module is used to calculate the significance parameter of each search object at the current time based on the search operation matrix, and update the group weight of each search object according to the significance parameter; The search result feedback module is used to determine the amount of search processing resources for each search object based on the grouping weight, call the large language model based on the amount of search processing resources, generate and feed back the search results; wherein, the amount of search processing resources is used at least to adjust the search expansion depth and search priority.

8. The intelligent search and grouping system for a large language model according to claim 7, characterized in that, The search feature extraction module includes: The data acquisition unit is used to acquire user-submitted search requests and historical search behavior data; The preprocessing unit is used to perform text preprocessing on search requests and historical search behavior data to obtain word groups; the text preprocessing includes at least word segmentation, noise reduction and keyword extraction. The text conversion unit is used to convert the word groups corresponding to the search request to obtain semantic vectors, and to convert the word groups corresponding to historical search behavior data to obtain behavioral features; The first feature statistical unit is used to statistically analyze semantic vectors and behavioral features as search state features; The semantic vectors and behavioral features are both text converted into numerical format.

9. A computer device, characterized in that, The computer device includes one or more processors and one or more memories, wherein at least one piece of program code is stored in the one or more memories, and when the program code is loaded and executed by the one or more processors, it implements the intelligent search grouping method for large language models as described in claims 1-6.

10. A storage medium, characterized in that, The storage medium stores at least one piece of program code, which, when loaded and executed by the processor, implements the intelligent search grouping method for large language models as described in claims 1-6.

Images