Content query method and device, electronic equipment and computer readable medium

Abstract

The invention discloses a content query method and device, electronic equipment and a computer readable medium. The method comprises the steps that to-be-queried content is acquired; in business data of at least one target business accessed to the search platform, business data matched with the to-be-queried content is searched based on a keyword search mode and a semantic search mode, and a first search result corresponding to the keyword search mode and a second search result corresponding to the semantic search mode are obtained; fusing the first search result and the second search result to obtain a fusion result; and obtaining a target query result corresponding to the to-be-queried content based on the fusion result. Therefore, access operations of different services can be realized through the search platform, so that the different services can use the search platform to realize a search function, and each access service can obtain the same search experience; in addition, the query result of the to-be-queried content is obtained in combination with keyword search and semantic search modes, and the search accuracy of the query result can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of mobile terminals, and more specifically, to a content query method, device, electronic device and computer-readable medium. Background Art

[0002] With the development of software and hardware technology, the functions of current mobile phones are becoming increasingly rich, and users can get rich experiences just through mobile phones. However, from another perspective, the increasing number of applications and functions has also led to the complexity of information in mobile phones. When users need a certain function or data, the current method usually requires users to enter a specific application and search for accurate keywords to find the required content, which is not a reasonable search method. Summary of the invention

[0003] The present application proposes a content query method, device, electronic device and computer-readable medium to improve the above-mentioned defects.

[0004] In a first aspect, the present application provides a content query method, which is applied to a search platform installed in an electronic device, and the method includes: obtaining content to be queried; searching for business data matching the content to be queried in business data of at least one target business connected to the search platform based on a keyword search method and a semantic search method, respectively, to obtain a first search result corresponding to the keyword search method and a second search result corresponding to the semantic search method; fusing the first search result and the second search result to obtain a fused result; and obtaining a target query result corresponding to the content to be queried based on the fused result.

[0005] In the second aspect, the present application also provides a content query device, which is applied to a search platform installed in an electronic device, and the device includes: an acquisition unit, a search unit, a fusion unit and a processing unit. The acquisition unit is used to acquire the content to be queried; the search unit is used to search for business data matching the content to be queried in the business data of at least one target business connected to the search platform based on a keyword search method and a semantic search method, respectively, to obtain a first search result corresponding to the keyword search method and a second search result corresponding to the semantic search method; the fusion unit is used to fuse the first search result and the second search result to obtain a fusion result; the processing unit is used to obtain a target query result corresponding to the content to be queried based on the fusion result.

[0006] In a third aspect, the present application also provides an electronic device comprising: one or more processors; a memory; and one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to execute the above method.

[0007] In a fourth aspect, the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a program code executable by a processor, and when the program code is executed by the processor, the processor executes the above method.

[0008] The content query method, device, electronic device and computer-readable medium provided by the present application obtain the content to be queried; in the business data of at least one target business connected to the search platform, search for business data matching the content to be queried based on the keyword search method and the semantic search method, respectively, to obtain a first search result corresponding to the keyword search method and a second search result corresponding to the semantic search method; merge the first search result and the second search result to obtain a merged result; and obtain the target query result corresponding to the content to be queried based on the merged result. Therefore, the access operation of different businesses can be implemented through the search platform, so that different businesses can use the search platform to implement the search function, and each access business can obtain the same search experience; in addition, combining the keyword search and semantic search methods to obtain the query results of the content to be queried can improve the search accuracy of the query results.

[0009] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 A method flow chart of a content query method provided by an embodiment of the present application is shown;

[0012] Figure 2 A method flow chart of a content query method provided by another embodiment of the present application is shown;

[0013] Figure 3 A schematic diagram showing a framework for creating a data index provided by an embodiment of the present application is shown;

[0014] Figure 4 A method flow chart of a content query method provided by another embodiment of the present application is shown;

[0015] Figure 5A schematic diagram showing a fusion process of multiple search methods provided in an embodiment of the present application is shown;

[0016] Figure 6 A method flow chart of a content query method provided by yet another embodiment of the present application is shown;

[0017] Figure 7 The overall architecture diagram corresponding to the content query method provided by an embodiment of the present application is shown;

[0018] Figure 8 A schematic diagram showing an application of a content query method provided by an embodiment of the present application is shown;

[0019] Fig. 9 A module block diagram of a content query device provided by an embodiment of the present application is shown;

[0020] Fig.10 A structural block diagram of an electronic device provided in an embodiment of the present application is shown;

[0021] Fig.11 A storage unit for storing or carrying program codes for implementing the method according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0022] In order to make those skilled in the art better understand the present application scheme, the technical scheme in the present application embodiment will be clearly and completely described below in conjunction with the drawings in the present application embodiment. Obviously, the described embodiment is only a part of the present application embodiment, rather than all the embodiments. The components of the present application embodiment usually described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the present application for protection, but merely represents the selected embodiment of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0023] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0024] With the development of software and hardware technology, the functions of current mobile phones are becoming increasingly rich, and users can get rich experiences just through mobile phones. However, from another perspective, the increasing number of applications and functions has also led to the complexity of information in mobile phones. When users need a certain function or data, the current method usually requires users to enter a specific application and search for accurate keywords to find a specific function or data; and when users do not know which application it comes from or the specific keywords, they cannot find the accurate target object through retrieval.

[0025] In short, the data and functions of current mobile phones are isolated from each other according to applications. Users cannot retrieve any content they want in their mobile devices through a unified search portal. At the same time, local searches on the device side usually only support precise searches based on keywords, or limited fuzzy searches based on synonyms and generalized synonyms; when users cannot remember the exact keywords, they cannot find the exact content. Usually, local searches on the device side do not support intelligent fuzzy searches based on natural language.

[0026] The inventors found in their research that the current search function has the following disadvantages:

[0027] 1) The data access method requires the access application to access the data of other applications one by one, which is time-consuming and labor-intensive to develop and maintain. It cannot provide a consistent search experience for other applications, and it cannot effectively guarantee data security.

[0028] 2) Based on the access method of search interface, it is limited by the search capabilities of each application, and it is impossible to obtain a consistent intelligent search experience; the search results of each application cannot be filtered, aggregated, and integrated more effectively;

[0029] 4) The end-side search method based on keywords usually only supports precise keyword search and limited generalized search of synonyms and near-synonyms; it does not support natural language search requests, and the retrieval effect is usually poor;

[0030] 5) Cloud-based natural language search can usually support fuzzy intelligent search based on natural language, thereby obtaining a better search experience; however, business data needs to be transferred to the cloud for index storage, which cannot effectively protect user data privacy; at the same time, due to network conditions, the search delay is long, which affects the search experience to a certain extent.

[0031] See also Figure 1 , Figure 1A content query method provided by an embodiment of the present application is shown, which is applied to a search platform installed in an electronic device. The search platform can be an application in the electronic device, which provides a unified interface, for example, a unified access SDK. The application installed in the electronic device can access the search platform through the unified interface. It can be understood that accessing the search platform means that the search platform has access rights to at least part of the data of the application accessing the search platform, and the at least part of the data is applied to the content query operation. Specifically, the method includes: S101 to S104.

[0032] S101: Obtain the content to be queried.

[0033] The content to be queried can be a text content, or a voice or an image, that is, when the user inputs the content to be queried, the user can input text content, or a voice or an image, and of course, can also input text, voice and image at the same time, without limitation. It can be understood that the content to be queried can be regarded as a query request initiated by the user, and the electronic device needs to identify the content to be queried, and then return the reply content corresponding to the content to be queried, that is, the target query result.

[0034] S102: In the business data of at least one target business connected to the search platform, search for business data matching the content to be queried based on a keyword search method and a semantic search method, respectively, to obtain a first search result corresponding to the keyword search method and a second search result corresponding to the semantic search method.

[0035] It can be understood that the search platform (Data Middle Platform, DMP) is a platform that integrates search services and provides index creation and search capabilities for other businesses. The search platform corresponds to an access business, and the business refers to the application or service that accesses the search platform, such as files, settings, notes and other applications. The target business can be at least part of the access business corresponding to the search platform, for example, it can be a business that matches the query scope corresponding to the content to be queried, where the query scope can be the range set by the user when entering the content to be queried, for example, the user sets the type of query data, for example, setting the query type to include all documents, or setting the query type to include all video-related data, then the target business determined by the search platform can be a business that matches the query scope in the access business, for example, an application with document processing function, or an application with video processing function.

[0036] It should be noted that the business data corresponding to the target business can be data stored in the local storage space of the electronic device, can be data generated by the user in the target business, or can be data sent by other terminals and received by the target business within a certain period of time. There is no limitation on this.

[0037] As an implementation method, the business data of at least one target business connected to the search platform can be named as a target data set, and the business data matching the content to be queried can be searched in the target data set based on a keyword search method to obtain a first search result. Specifically, keyword search (also known as search based on search terms) is a search method, and its basic principle is to find documents or information containing these keywords by matching the query terms (i.e., keywords) input by the user with the words contained in the data. For example, the user enters one or more keywords, and the search platform searches for data related to the query keywords in the pre-built index, and then sorts the results according to relevance and returns them to the user.

[0038] At the same time, it is also necessary to search for business data that matches the content to be queried in the target data set based on semantic search to obtain the second search result. Semantic search is a search method based on semantic understanding. Unlike keyword search, the goal of semantic search is to understand the semantic intent behind the query and return relevant content based on semantic similarity rather than literal matching. Semantic search relies on natural language processing (NLP) technology. For example, when a user enters the content to be queried, the search platform will first use NLP technology to perform semantic analysis on the query to understand the intent and meaning behind it. The search platform converts the query into one or more semantic vectors through a model and compares them with the pre-calculated document vectors in the business data. The semantic similarity of the document (usually calculated by cosine similarity or other distance metrics) determines its relevance to the query, sorts it according to relevance, and returns the result that best meets the user's intent.

[0039] S103: Merge the first search result and the second search result to obtain a fusion result.

[0040] It can be understood that the first search result represents the business data that matches the content to be queried and is found by the search platform through keyword search method, and the second search result fusion represents the business data that matches the content to be queried and is found by the search platform through semantic search method.

[0041] As an implementation method, the first search result and the second search result can be fused to obtain a fused result, which enables the fused result to give full play to the advantages of the two search methods, make up for the shortcomings of a single search method, and improve the relevance and accuracy of the search results. Specifically, keyword search can accurately match the results containing the query words in the document, which is suitable for scenarios where the query words are clear and unambiguous. Semantic search can understand the semantics and context of the query, and has good support for synonyms, near-synonyms and complex queries. It can handle users' fuzzy queries and understand the intentions behind them.

[0042] After the search results of the two are merged, the fusion results can make up for the limitations of the two. The fusion search results can ensure that users can accurately query the content containing keywords and that semantically similar content can also be covered, thereby improving the completeness of the results. For example, when a user queries "smartphone", keyword search can accurately find documents containing "smartphone", while semantic search can return results related to "mobile phone" or "mobile device", ensuring that users will not miss any relevant information.

[0043] S104: Obtaining a target query result corresponding to the to-be-queried content based on the fusion result.

[0044] As an implementation method, the search platform can directly return the fusion result as the target query result corresponding to the content to be queried to the user. Of course, it is also possible to further complete the search result based on the fusion result to obtain the target query result corresponding to the content to be queried. For example, based on the fusion result, combined with the large language model, the target query result corresponding to the content to be queried is obtained. For details, please refer to the subsequent embodiments.

[0045] Therefore, in an embodiment of the present application, access operations for different services can be implemented through the search platform, so that different services can use the search platform to implement search functions, and each access service can obtain the same search experience; in addition, combining keyword search and semantic search to obtain query results for the content to be queried can improve the search accuracy of the query results.

[0046] See also Figure 2 , Figure 2 A content query method provided by an embodiment of the present application is shown, which is applied to the above search platform. Specifically, the method includes: S201 to S206.

[0047] S201: Obtain the content to be queried.

[0048] S202: Obtain a data index corresponding to business data of at least one target business connected to the search platform, wherein the data index includes first index information and second index information.

[0049] It should be noted that the first index information includes preset keywords corresponding to the business data, and the second index information includes preset semantic features corresponding to the business data. That is to say, for the first index information, it includes multiple business data identities and preset keywords corresponding to the identities of each business data. The preset keywords can be keywords determined by the business data based on the keyword inverted index. Specifically, the business data corresponds to data description information, which is used to describe the content of the business data and can be regarded as a summary content corresponding to the business data. The preset pipe detection corresponding to the data description information of each business data is determined based on the keyword inverted index, so that the data description information corresponding to the preset keyword can be determined, and the data description information corresponds to the identity of the business data, so that the preset keywords corresponding to each business data can be determined. Similarly, for the second index information, the preset semantic features corresponding to each business data can also be established by semantic analysis of the data description information of the business data.

[0050] S203: Based on the keyword search method, search for business data matching the to-be-queried content in the first index information to obtain the first search result.

[0051] Specifically, the content to be queried is analyzed to extract keywords, and then the keywords of the content to be queried are matched with each preset keyword in the first index information to determine the first matching degree between the content to be queried and each business data, and the first search result is obtained based on the first matching degree corresponding to each business data.

[0052] S204: Based on the semantic search method, search the second index information for business data matching the content to be queried to obtain the second search result.

[0053] Similarly, the content to be queried is analyzed, the semantic features of the content to be queried are extracted, the semantic features of the content to be queried are matched with the preset semantic features in the second index information, the second matching degree corresponding to the content to be queried and each business data is determined, and the first search result is obtained based on the second matching degree corresponding to each business data.

[0054] As an implementation method, the keyword search algorithm can be a fusion of multiple different keyword search algorithms. That is to say, the keyword search method includes a first keyword search algorithm and a second keyword detection algorithm. Then, based on the keyword search method, the business data matching the content to be queried is searched in the first index information to obtain the first search result. The implementation method is to search for the business data matching the content to be queried in the first index information based on the first keyword search algorithm to obtain the first matching result; based on the second keyword search algorithm, the business data matching the content to be queried is searched in the first index information to obtain the second matching result; and the first matching result and the second matching result are merged to obtain the first search result.

[0055] It can be understood that the first keyword search algorithm and the second keyword search algorithm can be two different algorithms, and the first matching result includes the first matching degree between the content to be queried and each preset keyword in the first index information determined by the first keyword search algorithm, and the second matching result includes the second matching degree between the content to be queried and each preset keyword in the first index information determined by the second keyword search algorithm. Then, the first matching degree and the second matching degree corresponding to each business data can be fused, so as to merge the first matching result and the second matching result to obtain the first search result. For example, the weight corresponding to the first keyword search algorithm and the weight corresponding to the second keyword search algorithm are determined, and based on the weight corresponding to the first keyword search algorithm and the weight corresponding to the second keyword search algorithm, the first matching degree and the second matching degree of each business data are weighted and summed to obtain the first search result. The first search result is a query result that integrates the search results of the first keyword search algorithm and the second keyword search algorithm. The first search result corresponds to a keyword matching degree, and the keyword matching degree corresponding to the business data is the fusion result of the first matching degree and the second matching degree corresponding to the business data.

[0056] As an implementation method, the first keyword search algorithm is an exact matching algorithm, and the second keyword search algorithm is a fuzzy matching algorithm. For example, the fuzzy matching algorithm is a like search, and the exact matching algorithm is a keyword search.

[0057] Like search is a search method based on fuzzy matching in the database. It is usually used in SQL databases to query whether the data contains specific character patterns through the LIKE operator. Like search usually uses wildcards for partial matching. Like search can be regarded as a fuzzy match. Like search does not require exact matching and can use wildcards for partial matching. Keyword search (sometimes also called full-text search) is an exact keyword matching search for text, usually used for more complex text content retrieval. Compared with like search, keyword search focuses on quickly finding words related to the query from a large amount of text, and usually uses technologies such as inverted index to improve search efficiency.

[0058] In the embodiment of the present application, in the first search result, the ranking priority of the second matching result is higher than the ranking priority of the first matching result, that is, the ranking priority of the like search is higher than the keyword search. It is understandable that when displaying the results of keyword matching, the ranking of the second matching result is higher than the first matching result. Assume that the second matching result includes 4 business data greater than the threshold value corresponding to the fuzzy matching algorithm, which are docID1, docID2, docID3, and docID4, which are sorted in sequence, and the first matching result includes 2, which are docID3 and docID5, which are sorted in sequence. In the first search result, the 4 second matching results and the 2 first matching results can be deduplicated first, that is, the remaining matching data after deduplication include docID1, docID2, docID3, docID4 and docID5. Regardless of whether the first matching degree corresponding to docID5 is greater than the second matching degrees of docID1, docID2, docID3, and docID4, docID5 is located after docID1, docID2, docID3, and docID4 in the sorting order. That is, the first search result is sorted as docID1, docID2, docID3, docID4, and docID5. In addition, in the keyword matching degrees corresponding to the first search result, docID1, docID2, docID3, docID4, and docID5 decrease in sequence. Specifically, by reasonably setting the weight corresponding to the first keyword retrieval algorithm and the weight corresponding to the second keyword retrieval algorithm, for example, reducing the weight corresponding to the first keyword retrieval algorithm and increasing the weight corresponding to the second keyword retrieval algorithm, the weight corresponding to the first keyword retrieval algorithm can be made smaller than the weight corresponding to the second keyword retrieval algorithm, thereby making the fused keyword matching degree corresponding to the first matching result in the first search result smaller than the fused keyword matching degree corresponding to the second matching result, and further making the ranking of the matching business data in the first matching result lower than the ranking of the matching business data in the second matching result.

[0059] It should be noted that reference may be made to subsequent embodiments for the fusion of the first keyword search algorithm and the second keyword search algorithm.

[0060] As an implementation mode, when obtaining the data index corresponding to the business data of at least one target business connected to the search platform, first determine whether the data index has been established. If the data index has been established, the data index can be used directly. If it has not been established, you can choose to create it in time or choose to create it later. Specifically, determine whether the data index corresponding to the business data of at least one target business connected to the search platform has been established; if it has been established, obtain the data index corresponding to the business data of at least one target business connected to the search platform; if it has not been established, establish the data index corresponding to the business data of at least one target business connected to the search platform.

[0061] In the embodiment of the present application, the access service will synchronize data with the search platform so that the search platform can better query the data in the access service. Therefore, when the access application synchronizes data with the search platform, the data index will be created simultaneously. Usually, the index library is created immediately before data retrieval or when the retrieval is triggered. The construction and maintenance process chain of the index library is relatively long and involves many modules.

[0062] See also Figure 3 , Figure 3 The creation framework of the data index is shown. The triggering timing of the maintenance (including creation and update) of the data index may include: active triggering of access services, triggering based on search requests, triggering of file resource managers, and daily maintenance.

[0063] Specifically, the active triggering of the access service refers to the index library of the access service actively maintaining the service through the SDK. Exemplarily, the maintenance of the index library includes adding indexes, deleting indexes and changing indexes for the data index of the access service. In the embodiment of the present application, the operations of adding indexes, deleting indexes and changing indexes can be uniformly named as update operations of data indexes. For example, services such as notes and photo albums are applicable to this triggering time. The index creation timing of this process is completely controlled by the access service. In addition, triggering based on search requests means that when the access service initiates a search request, the DMP has no index data. For example, the user clears the DMP data in the application management, or has never used the relevant functions. In this case, the DMP will enter the initial indexing process. In addition, the file explorer trigger refers to the file explorer (Metis) detecting changes in external resources in real time, such as adding new files on the mobile phone. The file explorer (Metis) will pull up the DMP to receive file change events and enter the immediate index maintenance task. Furthermore, daily maintenance refers to the daily tasks set by DMP triggering daily tasks during the daily off-peak hours. In this case, DMP will clean up invalid indexes and continue unfinished index tasks and index maintenance tasks. An invalid index means that the data corresponding to the index is invalid data, which may be data that has been deleted.

[0064] Therefore, there are two ways for applications to access DMP, one is semi-managed and the other is fully managed. Among them, the semi-managed method means that the access application party establishes and maintains the index (addition, deletion, modification, etc.) by itself, and the DMP is responsible for daily index maintenance (cleaning invalid data, updating keywords and vector indexes, etc.); the fully managed method is to delegate all indexing work to the DMP, and the DMP implements the establishment and maintenance of the index; the fully managed method requires the access application party to provide a data acquisition interface so that the DMP can read relevant data; it is understandable that when the application accesses the search platform, it can specify the scope of the search, that is, to search only its own application data or to search global data; the search platform will perform permission control and data isolation according to the access method.

[0065] The access application customizes the index database table structure stored in the DMP, namely the metainfo table, based on its own data characteristics. The metainfo table is a database table used to store and manage metadata information of the data. It usually contains attributes, characteristics, and other auxiliary information describing the data (such as title, author, tag, creation time, file type, etc.), rather than directly storing the data itself. These metadata can enable more efficient data retrieval, management, and analysis. In other words, when the access application stores data in the DMP, it usually determines the required metadata fields based on its own needs and data characteristics.

[0066] In addition, the access application defines the columns that need to be searched for keywords and vectors according to the metainfo table structure. Exemplarily, according to the requirements of the access application, the DMP system allows indexes to be configured for certain columns in the table. For example, these columns may include: columns corresponding to fields such as title, content, author, and tags, which can be indexed for keyword search, where each field of title, content, author, and tags corresponds to one column.

[0067] Furthermore, for columns with very long text, segmentation parameters need to be configured. That is, for very long text columns (such as content), the access application can configure segmentation parameters to divide the text into multiple smaller blocks. These text blocks will be stored in the chunkinfo table associated with the metainfo table for subsequent retrieval.

[0068] It is understandable that DMP will first synchronize the original data to the metainfo table, and then divide the columns that need to be segmented into text chunks according to the configuration, cut the text into text chunks that meet the length requirements, and store them in the chunkinfo table; then, according to the configuration, establish a keyword inverted index for the columns that need to be searched for keywords; and establish a vector index for the columns that need to be searched for vectors. The aforementioned text refers to the data of certain columns in the metainfo table, which may contain longer text fields (such as description information, notes, comments, content, etc.), that is, the description information of certain columns.

[0069] Specifically, DMP will first synchronize the original data (such as logs, files, event data, etc.) provided by the application to the metainfo table. At this time, the data is directly stored in the original format, and then the data is converted into a format that conforms to the metainfo table structure. The chunkinfo table is a table used to store segmented text blocks. For columns marked as requiring segmentation, DMP will segment their text data into blocks and store each text block in the chunkinfo table. Each text block may contain part of the original text. There will be a certain order mark between blocks, and each text block corresponds to the position of the original text. Chunking can improve retrieval efficiency and avoid processing overly long texts.

[0070] In the metainfo table, DMP will create an inverted index for columns that require keyword search (such as event_type, device_type, etc.) according to configuration requirements. In this way, when searching, the system can quickly find data records that match the query terms. Among them, the keyword inverted index refers to recording all documents or entries containing a certain keyword. For columns that require vector retrieval, DMP will create a special vector index. Vector indexes are used to efficiently find vectors with high similarity, and usually use specific algorithms (such as HNSW, IVF, etc.) to optimize retrieval efficiency.

[0071] It should be noted that for the index of the column that needs to be segmented, the index will be performed on the data after segmentation rather than the original data. The same applies to retrieval, that is, after the segment is retrieved, it will be mapped to the metainfo original data item. The daily task of DMP is to clean up the invalid index data and continue the unfinished data indexing and index maintenance tasks.

[0072] That is to say, the object of the index is the data after segmentation. When the original text data (such as a long text field in a column) is segmented, it is divided into multiple small blocks, and each small block may be stored in a different record. Although the original text field may be very long, the data after segmentation is actually composed of multiple shorter text blocks. In this case, in order to improve the efficiency of search or query, the index will be established on the segmented data, rather than directly on the original long text field. It is understandable that the segmented text data is usually more suitable for indexing than the original long text, because the content of each block is shorter, and full-text indexing, keyword search and other operations can be performed more efficiently. Exemplarily, assume that there is a long text field named event_description, and it has been segmented into multiple small blocks according to the configuration and stored in the chunkinfo table. For each small block (assuming it is named chunk_text), assuming that event_description is divided into three blocks and stored in the chunkinfo table, an index can be established on the chunk_text field of the chunkinfo table, so that each block can be searched more quickly when querying. If no chunking is performed, a full-text index may be created directly for the original event_description field. In this way, when querying, the search engine needs to process longer text, which may affect performance. For example, if a text is divided into chunk 1 and chunk 2, the local search will first query chunk 1, and then map chunk 1 to the original text and the corresponding specific text location.

[0073] In addition, when creating and maintaining data indexes, it is also necessary to consider whether the current state of the electronic device is suitable for the creation or maintenance of the data index. In other words, the creation or maintenance of the data index requires determining whether the electronic device meets the corresponding conditions. In other words, DMP creates data indexes for local note management settings in batches, and when processing each batch of data, it determines factors such as the phone's battery level, storage, and screen to reduce power consumption and the effects of phone heating and freezing.

[0074] It is understandable that, considering that the workload is relatively large when a data index is created for the first time and it is usually triggered by the user, different conditions can be set for the first creation and non-first creation to determine whether the objective conditions for executing the data index creation operation are met.

[0075] For the data index creation process, when a data index creation request is detected, determine whether the creation request is the first creation; if it is the first creation, execute the data index creation operation when it is determined that the electronic device meets the first condition; if it is not the first creation, execute the data index creation operation, and in the process of executing the data index creation operation, if it is detected that the electronic device meets the second condition, interrupt the data index creation operation. In other words, for a data index creation operation that is not the first time, the data index creation operation can be executed first, and during the execution process, determine whether the electronic device meets the second condition. If the second condition is met, terminate the data index creation operation, and then wait for the next task trigger before continuing the execution. It can be understood that the data index creation request can be triggered by the user or by the application. For details, please refer to the above explanation on the triggering timing.

[0076] As an implementation method, for the initial creation of a data index, the method of setting the first condition also needs to refer to whether the current time is in the daytime time period or the nighttime time period, because the data indexing operation has different requirements on the state of the electronic device for the daytime time period and the nighttime time period. The setting of the first condition not only needs to take into account the power consumption of the electronic device, but also needs to ensure that the data index creation operation can be successfully executed.

[0077] Specifically, if this creation request is the first creation, the target time period corresponding to the current moment is determined. If the target time period is within the daytime time period, the first condition is determined to include at least one of the remaining power of the electronic device being greater than a first power threshold, the shell temperature of the electronic device being less than a first temperature threshold, the CPU load of the electronic device being less than a first load threshold, and the remaining storage space of the electronic device being greater than a first space threshold.

[0078] The daytime time period can be set based on actual usage needs, for example, it can be set based on the user's living habits, for example, the time period when the user uses the mobile phone more frequently during the day, and there is no limitation on this. Exemplarily, the first power threshold, the first temperature threshold, the first load threshold and the first space threshold can also be set based on actual usage needs. For example, the first power threshold is 80% (i.e., 80% of the maximum power), the first temperature threshold is 41 degrees Celsius, and the first load threshold is 80%. 80% refers to 80% of the CPU usage rate, then 100% CPU load means that all CPU cores are fully occupied, and the first space threshold is 500MB.

[0079] It can be understood that if the moment of obtaining the creation request for the first index is within the daytime period, the corresponding first condition may be at least one of the remaining power of the electronic device being greater than the first power threshold, the shell temperature of the electronic device being less than the first temperature threshold, the CPU load of the electronic device being less than the first load threshold, and the remaining storage space of the electronic device being greater than the first space threshold, that is, it may be any one or a combination of multiple ones. In an embodiment of the present application, the first condition corresponding to the daytime period may include the remaining power of the electronic device being greater than the first power threshold, the shell temperature of the electronic device being less than the first temperature threshold, the CPU load of the electronic device being less than the first load threshold, and the remaining storage space of the electronic device being greater than the first space threshold, that is, when the remaining power of the electronic device is greater than the first power threshold, the shell temperature of the electronic device is less than the first temperature threshold, the CPU load of the electronic device is less than the first load threshold, and the remaining storage space of the electronic device is greater than the first space threshold, it is determined that the electronic device meets the first condition.

[0080] It is understandable that if the target time period is within the night time period, the first condition is determined to include at least one of the electronic device being in an idle state, the electronic device being in a screen-off state, the electronic device being in a charging state and the remaining power being greater than a second power threshold, the housing temperature of the electronic device being less than a first temperature threshold, the CPU load of the electronic device being less than a first load threshold, and the remaining storage space of the electronic device being greater than a first space threshold; if it is determined that the electronic device meets the first condition, the data index creation operation is performed. Similarly, the second power threshold can also be set based on actual usage, for example, the second power threshold can be 20%. Then the first condition corresponding to the night time period may also be at least one of the multiple conditions, or a combination of multiple conditions. In the embodiment of the present application, the first condition corresponding to the night time period may include that the electronic device is in an idle state, the electronic device is in an off-screen state, the electronic device is in a charging state and the remaining power is greater than the second power threshold, the shell temperature of the electronic device is less than the first temperature threshold, the CPU load of the electronic device is less than the first load threshold, and the remaining storage space of the electronic device is greater than the first space threshold. That is to say, when it is determined that the electronic device is in an idle state, the electronic device is in an off-screen state, the electronic device is in a charging state and the remaining power is greater than the second power threshold, the shell temperature of the electronic device is less than the first temperature threshold, the CPU load of the electronic device is less than the first load threshold, and the remaining storage space of the electronic device is greater than the first space threshold, it is determined that the electronic device meets the first condition during the night time period. Among them, the night time period can be 2-5 am.

[0081] Then, after the first condition is determined, if it is determined that the electronic device satisfies the first condition, a data index creation operation is performed.

[0082] As an implementation method, if the detected data index creation request is not the first creation, it is also necessary to determine the second condition based on the current time period of the electronic device. Specifically, if it is not the first creation, determine the target time period corresponding to the current moment; if the target time period is within the daytime time period, determine the second condition including that the shell temperature of the electronic device is greater than or equal to the first temperature threshold, the temperature difference of the shell temperature rise of the electronic device during the data index creation operation is greater than the second temperature threshold, or the power consumption of the electronic device during the data index creation operation exceeds the second power threshold; if the target time period is within the night time period, determine the second condition including that the shell temperature of the electronic device is greater than or equal to the first temperature threshold, the power consumption of the electronic device during the data index creation operation exceeds the second power threshold, or the electronic device is in a bright screen state; perform the data index creation operation, and in the process of performing the data index creation operation, if it is detected that the electronic device meets the second condition, interrupt the data index creation operation.

[0083] It is understandable that the second temperature threshold and the second power threshold can also be set based on actual usage requirements, for example, the second temperature threshold is 2 degrees Celsius, and the second power threshold is 2%. In the case of non-first creation, if the current time is in the daytime time period, it can be determined that the electronic device meets any one of the following conditions: the shell temperature of the electronic device is greater than or equal to the first temperature threshold, the temperature difference of the shell temperature rise of the electronic device during the data index creation operation is greater than the second temperature threshold, and the power consumption of the electronic device during the data index creation operation exceeds the second power threshold, and the electronic device meets any one of the following conditions: the second condition is determined. In the case of non-first creation, if the current time is in the night time period, it can be determined that the electronic device meets any one of the following conditions: the shell temperature of the electronic device is greater than or equal to the first temperature threshold, the power consumption of the electronic device during the data index creation operation exceeds the second power threshold, and the electronic device is in a bright screen state, and the second condition is determined.

[0084] As another embodiment, for the update operation of the data index, when an update request for an established data index is detected, it is determined whether the update request is the first update; if it is the first update, the update operation of the data index is performed when it is determined that the electronic device satisfies a third condition; if it is not the first update, the update operation of the data index is performed when it is determined that the electronic device satisfies a fourth condition.

[0085] It can be understood that the update operation of the data index may refer to the addition of indexes, modification of indexes, and other operations on the established data indexes. If it is the first update, the third condition is determined to include at least one of the remaining power of the electronic device being greater than the second power threshold, the CPU load of the electronic device being less than the first load threshold, and the search operation being performed based on the data index within the specified time period; when it is determined that the electronic device meets the third condition, the update operation of the data index is performed. In other words, the third condition is at least one of the multiple conditions such as the remaining power of the electronic device being greater than the second power threshold, the CPU load of the electronic device being less than the first load threshold, and the search operation being performed based on the data index within the specified time period, and it can also be any combination of multiple conditions. In the embodiment of the present application, the third condition is that the remaining power of the electronic device is greater than the second power threshold, the CPU load of the electronic device is less than the first load threshold, and the search operation being performed based on the data index within the specified time period is satisfied at the same time. Among them, the search operation being performed based on the data index within the specified time period means that within the specified time period, the electronic device has performed a query operation in the data index to be updated based on the user's query request, and the purpose is to ensure that the data index to be updated has been used. The specified time period may be within 7 days corresponding to the current time, may be 7 consecutive days before the current time, or may be 7 consecutive days including the current time.

[0086] If it is not the first update, the fourth condition includes at least one of the following: the time difference between the current moment and the last time the data index update operation was performed is greater than the duration threshold, the number of executions of the update operation in the current time period corresponding to the current moment is less than the number threshold, the data to be updated this time is less than the specified number threshold, and the current moment is in the night time period; if it is determined that the electronic device meets the fourth condition, the data index update operation is performed. Among them, the duration threshold, the number threshold and the specified number threshold can be set based on actual usage. For example, the duration threshold is 10s, the number threshold is 4, and the specified number threshold is 50. That is to say, for non-first updates, the limit is triggered once every 10s (file), no more than four times and no more than 50 documents / notes (<20s, <2mAH) in one minute, and updates are made at night when the device is idle.

[0087] S205: Merge the first search result and the second search result to obtain a fused result.

[0088] S206: Obtain a target query result corresponding to the content to be queried based on the fusion result.

[0089] It should be noted that, for the contents not described in detail in the above steps, reference can be made to the above embodiments and they will not be described again here.

[0090] Therefore, it can be seen that the retrieval method provided in the embodiment of the present application is a hybrid retrieval that includes a variety of different search algorithms. Specifically, the implementation method of the hybrid retrieval can refer to the subsequent embodiments.

[0091] See also Figure 4 , Figure 4 A content query method provided by an embodiment of the present application is shown, which is applied to the above search platform. Specifically, the method includes: S401 to S406.

[0092] S401: Obtain the content to be queried.

[0093] S402: In the business data of at least one target business connected to the search platform, business data matching the content to be queried is searched based on a keyword search method and a semantic search method, respectively, to obtain a first search result corresponding to the keyword search method and a second search result corresponding to the semantic search method.

[0094] S403: Determine a first weight corresponding to the first search result and a second weight corresponding to the second search result.

[0095] S404: reordering the plurality of first matching data and the second matching data based on the first weight and the second weight to obtain a shuffled result, wherein the shuffled result includes a plurality of matching data that are sequentially ordered.

[0096] S405: Taking the top N matching data in the shuffled result as the fusion result, where N is an integer greater than 1.

[0097] As an implementation method, the fusion of the first search result and the second search result is to recalculate the score of each matching data by weighted summing the first matching data and the second matching data to obtain a scoring value for each matching data. For example, a weighted fusion operation is performed based on the first weight, the keyword matching degree corresponding to the first matching data, the second weight, and the semantic matching degree corresponding to the second matching data to obtain a scoring value corresponding to each matching data; multiple matching data are sorted based on the scoring value, and the top N matching data are used as the fusion result, where N is an integer greater than 1.

[0098] It should be noted that the first matching data in the first search result refers to the matching data of the successful match determined based on the keyword search method. For example, based on the keyword search method, the keyword matching degree between each business data and the content to be queried can be determined, and the business data can be sorted based on the keyword matching degree. The first number of business data with the highest ranking is used as the first matching data. Thus, the first matching data corresponds to the keyword matching degree. Similarly, the second matching data in the second search result refers to the matching data of the successful match determined based on the semantic search method. It can be understood that the first matching data and the second matching data can be the same data, and the first matching data and the second matching data embody different names. For example, a text docID is recorded as the first matching data in the first search result and as the second matching data in the second search result.

[0099] As an implementation method, a weighted fusion operation is performed based on the first weight, the keyword matching degree corresponding to the first matching data, the second weight, and the semantic matching degree corresponding to the second matching data to obtain an implementation method for the scoring value corresponding to each matching data, and the identity identifier corresponding to the first matching data is determined. The identity identifier is recorded as the identity information of the first matching data. For example, the aforementioned docID is the identity identifier. Thus, the keyword matching degree and semantic matching degree corresponding to each business data can be determined. Then, for each business data, the corresponding scoring value is the product of the first weight and the keyword matching degree, and then the product is summed with the product of the second weight and the semantic matching degree. The result is the scoring value corresponding to the business data.

[0100] It can be understood that the first matching data in the first search result may be the business data that has been screened after the keyword matching degree is calculated for each business data, that is, the business data is sorted based on the keyword matching degree mentioned above, and the first number of business data with the highest sorting degree is taken as the first matching data. Similarly, the second matching data in the second search result is also the business data after screening. When the first search result and the second search result are merged, some matching data may not have corresponding keyword matching degree or semantic matching degree, and the non-corresponding matching degree can be set to zero.

[0101] As an implementation method, the first weight and the second weight may be the same or different. If the first weight and the second weight are different, the second weight may be set to be greater than the first weight. Of course, the first weight and the second weight may also be determined based on the first search result. Specifically, the implementation method of determining the first weight corresponding to the first search result and the second weight corresponding to the second search result is to determine the reference matching degree based on at least one of the keyword matching degrees in the first search result; if the reference matching degree is less than a preset threshold, the second weight is set to be greater than the first weight; if the reference matching degree is greater than or equal to the preset threshold, the first weight and the second weight are set to be the same. Among them, the reference matching degree may be the average value of the keyword matching degrees corresponding to each first matching data in the first search result. Of course, it may also be the maximum value of the keyword matching degrees corresponding to each first matching data. That is to say, the implementation method of determining the reference matching degree based on at least one of the keyword matching degrees in the first search result may be to determine the largest keyword matching degree in the first search result as the reference matching degree.

[0102] As another implementation, the plurality of first matching data and second matching data may be reordered based on the first weight and the second weight to obtain a shuffled result, wherein the first weight and the second weight affect the order of the respective corresponding matching data in the shuffled result. For example, the reordering may be performed based on the following formula (1).

[0103] rrf_score = weight*(1 / (rank+c)) (1)

[0104] In formula (1), c is the fusion constant, which can be set to 60, rank is the sequence number of the matching data in its corresponding queue, and weight is the first weight or the second weight.

[0105] That is, the first search result includes a plurality of first matching data which are sequentially sorted, and the second search result includes a plurality of second matching data which are sequentially sorted.

[0106] For the first search result, the weight in formula (1) is the first weight, and each first matching data in the first search result corresponds to a serial number in the first search result and an identity identifier. For the first search result, each matching data therein can be traversed to calculate the rrf_score, i.e., the first score, of each matching data in the recall path. Taking the first matching data in the first search result as an example, each document corresponds to a docID, and each document corresponds to a sorted serial number in the first search result. Based on the above formula (1), the rrf_score corresponding to each docID can be obtained. Similarly, for the second search result, the weight in formula (1) is the second weight, which can also obtain the rrf_score, i.e., the second score, corresponding to each matching data in the second search result. The first score and the second score of each matching data are accumulated, i.e., the rrf_score in the two recall paths is accumulated to the corresponding docID, thereby obtaining the score value corresponding to each matching data, and then sorting each matching data based on the score value of each matching data, thereby obtaining a mixed arrangement result. Then, the top N matching data in the shuffled result are taken as the fusion result, where N is an integer greater than 1.

[0107] S406: Obtain a target query result corresponding to the to-be-queried content based on the fusion result.

[0108] It can be seen that the content query method provided in the embodiment of the present application uses a mixed search of multiple search methods. Specifically, after receiving the search request, the DMP will first perform word segmentation on the search query, and the obtained term term includes time words and common words, among which time words can be used for time-based conditional search, and common words are mainly used for keyword search methods. Exemplarily, it is assumed that in the present application, the keyword search method includes two different keyword search algorithms, namely, an exact matching algorithm and a fuzzy matching algorithm (for example, a like search). Compared with a like search, the exact matching algorithm here belongs to an exact search, that is, an exact matching algorithm based on keywords. Then, the keyword search method is to merge the search results of the exact search and the fuzzy search. Therefore, common words are mainly used for exact matching algorithms and like searches, etc., and synonyms and near-synonyms are also generalized to achieve fuzzy search based on keywords.

[0109] When the front-end application requests DMP to perform a search, DMP will perform keyword search and vector semantic search from different resource contents according to the specified search scope (such as settings, notes, files, etc., you can specify cross-application multi-resource search, or limit the search to a single resource), and then perform mixed sorting.

[0110] As an implementation mode, it is assumed that the keyword search method (i.e., traditional search) mentioned in this application includes a first keyword search algorithm and a second keyword search algorithm, wherein the first keyword search algorithm is an exact matching algorithm, the second keyword search algorithm is a fuzzy matching algorithm, and the voice search method is a vector search algorithm, please refer to Figure 5 The fusion process of the multiple search methods is as follows: Figure 5 shown.

[0111] Figure 5 In the above, keyword search is the exact matching algorithm, and like search is the fuzzy matching algorithm. The matching degree of ike search results is calculated as the proportion of hit text to the original text, that is, the hit ratio. The hit ratio refers to the ratio of the part of the query text (or keyword) that appears in the target text to the total length of the target text. For example, if you search for "apple" and the target text is "I love apple pie", the hit part is "apple". The hit ratio can be calculated by dividing the number of characters that appear in the text of "apple" by the total number of characters in the target text. In addition, the keyword matching degree of the exact matching algorithm, that is, the ordinary keyword matching algorithm, is calculated as TF-IDF score * hit term number. The exact matching algorithm is usually used to calculate the matching degree of keywords. Its core idea is to measure the importance of each term based on keyword frequency (TF) and inverse document frequency (IDF), and to obtain the final matching degree through the weighted scores of these importance. Among them, TF (Term Frequency) refers to the frequency of a word in a document. Words with high TF values ​​appear more frequently in documents, which usually means that the word is more important in the current document. IDF (Inverse Document Frequency) indicates the importance of a word in the entire document set. Words with high IDF values ​​usually appear less frequently in many documents, indicating that they are relatively rare in the corpus and therefore have a higher weight.

[0112] It can be seen that the entire search process consists of two parts, namely the keyword search part and the semantic search part.

[0113] For keyword search:

[0114] A like fuzzy search is performed in the relevant keyword search column of the resource database table, where the relevant keyword search column refers to a column in the index database of a certain application to be queried for keyword search. For example, the application is a setting application, and the content to be queried this time is used to obtain the relevant content of a certain setting item. Then the relevant keyword search column can refer to a column in the index database corresponding to the setting application for keyword search, such as the setting item description column. In other words, in the data index, a like fuzzy search is performed on the keyword corresponding to a certain search column corresponding to the application to be queried.

[0115] A keyword search is performed in the relevant keyword search column of the resource database table to perform an exact match. The explanation of the relevant keyword search column can be referred to the above content and will not be repeated here.

[0116] For example, if the application to be queried is a settings application and the content to be queried is a settings item, in the keyword exact match search method, only the original segmentation results in the segmentation results are used to perform keyword retrieval, and the synonyms of the segmentation results and the time and location words of NER are not used. This is because the query of mobile phone settings items does not have time and location characteristics, for example, users will not say "yesterday's XX settings".

[0117] After the like fuzzy search and the keyword exact match search, the search results corresponding to the keyword exact match search and the search results corresponding to the like fuzzy search are obtained. Then, the search results corresponding to the keyword exact match search and the search results corresponding to the like fuzzy search are filtered respectively to obtain the first matching result corresponding to the keyword exact match search and the second matching result corresponding to the like fuzzy search. Among them, the filtering operation is used to filter the non-existent setting items, wherein, due to the change of data, the DMP will not delete the relevant data index immediately, but mark it as invalid; the cleanup task will be uniformly performed during idle time to avoid affecting user use; when performing a search, it is necessary to filter the invalid data in the search results, that is, filter to the invalid search results, wherein invalid means that the data corresponding to the search result has been deleted. Then, the first matching result corresponding to the keyword exact match search and the second matching result corresponding to the like fuzzy search are de-duplicated and merged, and the merged result is named the first search result. The merge operation of the first matching result and the second matching result can refer to the aforementioned content, and will not be repeated here.

[0118] After merging, the first search result saves a first preset number of matching data. As mentioned above, the first search result includes multiple first matching data and the keyword matching degree corresponding to each first matching data. The number of first matching data is the first preset number. For example, the first preset number is 6, that is, the results of the like search and the keyword exact match search are merged, and the merged result retains the top 6 search results.

[0119] like Figure 5 As shown, the semantic search method is vector semantic search. A vector index is established according to the relevant vector index column of the resource database table, and a vector search is performed in the corresponding vector index table. The establishment of the vector index can refer to the aforementioned embodiment and will not be repeated here. The vector semantic search can be scored using the embedding vector score. Specifically, in the resource database table, the embedding column is used to store the embedded vector of each resource (ie, business data). After the vector index is established, vector search can be performed. During retrieval, it is necessary to use the query vector and the embedded vector of each resource in the database to perform similarity calculation. In vector retrieval, scoring is usually based on the similarity between the query vector and the embedded vector in the database, thereby obtaining the similarity corresponding to each business data corresponding to the vector semantic search, that is, the semantic matching degree, and then, a filtering operation is performed, that is, filtering the non-existent business data (eg, setting items), thereby obtaining the second search result.

[0120] After obtaining the first search result and the second search result, a shuffling strategy is executed. For example, RRF shuffling may be used, that is, RRF uses the inverse of the ranking sequence to merge the rankings of the results of different retrieval strategies.

[0121] Specifically, the recall strategy weights and fusion constants are initialized. Among them, there are three strategies for weight use: Adaptive strategy: applied to the setting item, check the maximum matching degree in the traditional search, if it is lower than the set de-weighting threshold, set the traditional search and vector search weights to [0.3, 0.7], otherwise set to [0.5, 0.5]. Among them, the maximum matching degree in the traditional search refers to the aforementioned reference matching degree, and the de-weighting threshold can be the aforementioned preset threshold. Therefore, the weight setting method below the set de-weighting threshold and not lower than the de-weighting threshold can refer to the aforementioned content, which will not be repeated here. Specifically, in the case of setting the second weight greater than the first weight, the second weight can be set to 0.7 and the first weight is 0.3. In addition, the weight strategy can also include: fixed average strategy: fixed to [0.5, 0.5], applied to chunk shuffling in document search; priority strategy: fixed one weight to 0.7, applied to title shuffling in document search, for example, the second weight corresponding to the vector can be set to 0.7.

[0122] The implementation method of the weighted fusion operation based on the first weight, the keyword matching degree corresponding to the first matching data, the second weight and the semantic matching degree corresponding to the second matching data can be to calculate the scores of the docID in the two-way recall based on the above formula (1), and add the calculation results of the two ways to the corresponding docID. If the evaluation adopts the reranker model for re-ranking, the results of the multi-way recall are directly merged after deduplication, and sorted according to the score of the reranker model.

[0123] Therefore, the embodiments of the present application can integrate multiple search methods to obtain the final search results, thereby achieving a smarter and more accurate search effect.

[0124] See also Figure 6 , Figure 6 A content query method provided by an embodiment of the present application is shown, which is applied to the above search platform. Specifically, the method includes: S601 to S604.

[0125] S601: Obtain the content to be queried.

[0126] S602: In the business data of at least one target business connected to the search platform, search for business data matching the content to be queried based on a keyword search method and a semantic search method, respectively, to obtain a first search result corresponding to the keyword search method and a second search result corresponding to the semantic search method.

[0127] S603: Merge the first search result and the second search result to obtain a fusion result.

[0128] S604: Obtaining an output result of the large language model based on the fusion result through the large language model, and obtaining a target query result corresponding to the content to be queried based on the output result.

[0129] That is, the fusion result is used as the input of the large language model, and then the output result of the large language model is obtained, and the target query result corresponding to the content to be queried is obtained based on the output result. As an implementation method, the fusion result and the content to be queried can be assembled into a prompt vector prompt, and the prompt vector prompt is input into the large language model (for example, LLM model).

[0130] It is understandable that the large language model can be deployed on an electronic device or on the cloud. If it is deployed on the cloud, the implementation method of S604 can be to send the fusion result to the cloud, trigger the cloud to generate prompt information based on the fusion result, input the prompt information into the large language model, and obtain the target query result corresponding to the content to be queried based on the output result of the large language model.

[0131] like Figure 7 As shown, it is assumed that the large language model is deployed in the cloud.

[0132] Depend on Figure 7 It can be seen that different services can access the DMP. Figure 7 In the example, the execution flow of this method can be:

[0133] 1) The front-end application initiates intelligent search through the RAGAgent interface provided by the fusion search service-search middle platform (DMP, referred to as DMP below);

[0134] 2) If the DMP has established relevant data indexes, go to 8); otherwise, go to 3);

[0135] 3) Prompt the user that the index has not been established yet and immediate indexing is required;

[0136] 4) If the user chooses to perform indexing later, the user is prompted that the indexing is not yet ready and the process terminates;

[0137] 5) If the user selects instant indexing, go to 6);

[0138] 6) DMP reads data of access resources through the data interface;

[0139] 7) DMP performs keyword inverted indexing and vector indexing on the data of access resources;

[0140] 8) DMP performs keyword inverse search and vector search based on the query string passed by the foreground application;

[0141] 9) DMP merges and sorts the results of inverted search and vector search to generate local search results;

[0142] 10) DMP passes local search results to the AI ​​cloud service;

[0143] 11) AI cloud service uses the Reranker model to rerank the results;

[0144] 12) AI cloud service enhances the prompt words based on the re-ranking results and passes them to the LLM large model;

[0145] 13) The large model generates the final search answer in a streaming manner;

[0146] 14) The AI ​​cloud service streams the answer back to the DMP;

[0147] 15) The DMP passes the answer to the front-end application and associates the final answer with local resources;

[0148] 16) The foreground application presents the interface.

[0149] It should be noted that the specific implementation of the above steps has been described in the above embodiments and will not be repeated here. It can be seen that on the basis of the local search results, the large language model in the cloud can also be combined to obtain more accurate reply content as the target query result corresponding to the query content.

[0150] Combined with the above Figure 5 The steps shown in the figure can obtain the fusion result of local multi-resource retrieval through the above steps. Then, for the fused result, the top N1 (i.e. the top N ranked ones mentioned above) results can be retained, and the cloud interface can be called to transmit the search request (including the content to be queried) and the fusion result to the cloud for further retrieval and sorting.

[0151] After the cloud receives the search request and the fused search results, it first uses the rerank model to sort the fused results according to the relevance of the search request and the search result items, and retains the top N2 results. In other words, after the cloud receives the request, it will use the rerank model to reorder the matching data in the transmitted fused results. The reranking model is usually based on machine learning or deep learning models, and the training goal is to optimize based on the relevance of the query and the search results.

[0152] The filtered results and the search request are then assembled into a prompt, and the LLM interface is called to obtain the LLM output results. LLM can more intelligently analyze the correlation between natural language-based search requests and the searched items, thereby obtaining more accurate retrieval results. However, the processing delay of LLM is usually long, so the output results of LLM are returned to the client side in the form of streaming data for display.

[0153] See also Figure 8 , an example is used to illustrate the embodiment of the present application, Figure 8 The interface displayed is the interface of the full search application. Figure 8 As shown in (a), the interface displays recommended content 801, and the user can click on the recommended content 801. The clicked recommended content will be used as the current content to be queried. The cloud side can configure a daily / 7-day upper limit. If the number of times exceeds this limit, the recommended content will no longer be displayed within the specified time. In addition, the cloud side pre-stores about 50 recommended words as recommended content, and the project will no longer push recommended words if the effective time exceeds this time period. Figure 8 As shown in (b), the interface displays an identification control 802. After the content to be queried is input, the query operation can be triggered by clicking the identification control 802. As an implementation, when it is detected that the input content exceeds a specified number of characters, the large model can be preloaded in advance before the content input is completed. Figure 8As shown in (c), the interface displays a control 803. Clicking the control stops the target query result generation. Figure 8 (d) as shown.

[0154] Therefore, in the embodiment of the present application, the front-end access application can use this solution to search from settings, notes, documents and other content at one time, and filter, summarize and merge the results to obtain a cross-application search experience; and, it integrates multiple retrieval technologies, such as time retrieval, conditional retrieval, keyword inverse retrieval and semantic vector retrieval, etc., and can perform multiple retrieval strategies for the same data source at the same time, thereby obtaining richer retrieval results.

[0155] Furthermore, the embodiments of the present application can support customized search strategies and search scopes for services; when a service is accessed, it can specify the search scope and search strategy at the same time, so as to achieve differentiated search effects for different data features of different services, and better adapt to the needs of the service itself. In addition, the aforementioned reranker model and LLM model can also be further deployed on the end-side device, so that the entire solution can be completely end-side, independent of the network, and can effectively avoid the situation where intelligent search is unavailable in no network or weak network environment; furthermore, after obtaining the local fusion search results, the search results can be directly presented to the foreground application instead of continuing to be transmitted to the cloud for refined search, which can achieve lower search latency and resource consumption.

[0156] See also Fig. 9 , which shows a structural block diagram of a content query device 900 provided in an embodiment of the present application. The device may include: an acquisition unit 901, a search unit 902, a fusion unit 903 and a processing unit 904.

[0157] The acquisition unit 901 is used to acquire the content to be queried.

[0158] The search unit 902 is used to search for business data matching the content to be queried in the business data of at least one target business connected to the search platform based on a keyword search method and a semantic search method, respectively, to obtain a first search result corresponding to the keyword search method and a second search result corresponding to the semantic search method.

[0159] Furthermore, the search unit 902 is also used to obtain a data index corresponding to business data of at least one target business connected to the search platform, wherein the data index includes first index information and second index information, wherein the first index information includes preset keywords corresponding to the business data, and the second index information includes preset semantic features corresponding to the business data; based on the keyword search method, in the first index information, search for business data matching the content to be queried to obtain the first search result; based on the semantic search method, in the second index information, search for business data matching the content to be queried to obtain the second search result.

[0160] Furthermore, the search unit 902 is also used to search for business data matching the content to be queried in the first index information based on the first keyword retrieval algorithm to obtain a first matching result; to search for business data matching the content to be queried in the first index information based on the second keyword retrieval algorithm to obtain a second matching result; and to merge the first matching result and the second matching result to obtain the first search result.

[0161] Furthermore, the first keyword search algorithm is an exact matching algorithm, the second keyword search algorithm is a fuzzy matching algorithm, and in the first search result, the sorting priority of the second matching result is higher than the sorting priority of the first matching result.

[0162] Furthermore, the search unit 902 is also used to determine whether a data index corresponding to the business data of at least one target business connected to the search platform has been established; if it has been established, obtaining the data index corresponding to the business data of at least one target business connected to the search platform; if it has not been established, establishing a data index corresponding to the business data of at least one target business connected to the search platform.

[0163] Furthermore, the search unit 902 is also used to determine whether the creation request is the first creation when a data index creation request is detected; if it is the first creation, then the data index creation operation is performed when it is determined that the electronic device satisfies the first condition; if it is not the first creation, the data index creation operation is performed, and in the process of executing the data index creation operation, if it is detected that the electronic device satisfies the second condition, the data index creation operation is interrupted.

[0164] Furthermore, the search unit 902 is further configured to determine the target time period corresponding to the current moment if it is the first creation;

[0165] If the target time period is within the daytime time period, determining the first condition includes at least one of the remaining power of the electronic device being greater than a first power threshold, the shell temperature of the electronic device being less than a first temperature threshold, the CPU load of the electronic device being less than a first load threshold, and the remaining storage space of the electronic device being greater than a first space threshold; if the target time period is within the night time period, determining the first condition includes at least one of the electronic device being in an idle state, the electronic device being in a screen-off state, the electronic device being in a charging state and the remaining power being greater than a second power threshold, the shell temperature of the electronic device being less than a first temperature threshold, the CPU load of the electronic device being less than a first load threshold, and the remaining storage space of the electronic device being greater than a first space threshold; when determining that the electronic device satisfies the first condition, executing a data index creation operation.

[0166] Furthermore, the search unit 902 is further configured to determine the target time period corresponding to the current moment if it is not the first creation;

[0167] If the target time period is within the daytime time period, determining the second condition includes that the shell temperature of the electronic device is greater than or equal to the first temperature threshold, the temperature difference of the shell temperature rise of the electronic device during the data index creation operation is greater than the second temperature threshold, or the power consumption of the electronic device during the data index creation operation exceeds the second power threshold; if the target time period is within the night time period, determining the second condition includes that the shell temperature of the electronic device is greater than or equal to the first temperature threshold, the power consumption of the electronic device during the data index creation operation exceeds the second power threshold, or the electronic device is in a screen-on state; executing the data index creation operation, and during the execution of the data index creation operation, if it is detected that the electronic device meets the second condition, interrupting the data index creation operation.

[0168] Furthermore, the search unit 902 is also used to determine whether an update request for an established data index is a first update when an update request is detected; if it is the first update, then when it is determined that the electronic device satisfies a third condition, the data index update operation is performed; if it is not the first update, then when it is determined that the electronic device satisfies a fourth condition, the data index update operation is performed.

[0169] Furthermore, the search unit 902 is also used to determine, if it is the first update, a third condition including that the remaining power of the electronic device is greater than a second power threshold, the CPU load of the electronic device is less than a first load threshold, and at least one of a search operation has been performed based on the data index within a specified time period; and when it is determined that the electronic device satisfies the third condition, an update operation of the data index is performed.

[0170] Furthermore, if it is not the first update, the search unit 902 is also used to determine at least one of the fourth conditions, including that the time difference between the current moment and the moment when the data index update operation was last performed is greater than a duration threshold, the number of executions of the update operation within the current time period corresponding to the current moment is less than a number threshold, the data to be updated this time is less than a specified number threshold, and the current moment is in the night time period; if it is determined that the electronic device meets the fourth condition, the data index update operation is performed.

[0171] The fusion unit 903 is used to fuse the first search result and the second search result to obtain a fusion result.

[0172] Furthermore, the first search result includes multiple first matching data and a keyword matching degree corresponding to each first matching data, the second search result includes multiple second matching data and a semantic matching degree corresponding to each second matching data, and the fusion unit 903 is also used to determine a first weight corresponding to the first search result and a second weight corresponding to the second search result; based on the first weight and the second weight, the multiple first matching data and the second matching data are reordered to obtain a mixed result, wherein the mixed result includes multiple matching data sorted in sequence; and the top N matching data in the mixed result are used as the fusion result, wherein N is an integer greater than 1.

[0173] Furthermore, the first weight and the second weight are the same.

[0174] Furthermore, the second weight is greater than the first weight.

[0175] Furthermore, the fusion unit 903 is also used to determine a reference matching degree based on the matching degree of at least one of the keywords in the first search result; if the reference matching degree is less than a preset threshold, the second weight is set to be greater than the first weight; if the reference matching degree is greater than or equal to the preset threshold, the first weight and the second weight are set to be the same.

[0176] Furthermore, the fusion unit 903 is further configured to determine the maximum keyword matching degree in the first search result as a reference matching degree.

[0177] The processing unit 904 is used to obtain a target query result corresponding to the to-be-queried content based on the fusion result.

[0178] Furthermore, the processing unit 904 is also used to obtain an output result of the large language model based on the fusion result through the large language model, and obtain a target query result corresponding to the content to be queried based on the output result.

[0179] Furthermore, the processing unit 904 is also used to send the fusion result to the cloud, trigger the cloud to generate prompt information based on the fusion result, input the prompt information into the large language model, and obtain the target query result corresponding to the content to be queried based on the output result of the large language model.

[0180] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here.

[0181] In several embodiments provided in the present application, the coupling between modules may be electrical, mechanical or other forms of coupling.

[0182] In addition, each functional module in each embodiment of the present application can be integrated into a processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The above integrated modules can be implemented in the form of hardware or software functional modules.

[0183] Please refer to Fig.10 , which shows a structural block diagram of an electronic device provided in an embodiment of the present application. The electronic device 100 may be an electronic device capable of running applications, such as a smart phone, a tablet computer, an e-book, etc. The electronic device 100 in the present application may include one or more of the following components: a processor 110, a memory 120, and one or more applications, wherein the one or more applications may be stored in the memory 120 and configured to be executed by one or more processors 110, and the one or more programs are configured to execute the method described in the aforementioned method embodiment.

[0184] The processor 110 may include one or more processing cores. The processor 110 uses various interfaces and lines to connect various parts of the entire electronic device 100, and executes various functions and processes data of the electronic device 100 by running or executing instructions, programs, code sets or instruction sets stored in the memory 120, and calling data stored in the memory 120. Optionally, the processor 110 can be implemented in at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). The processor 110 can integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. Among them, the CPU mainly processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing display content; and the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 110, but may be implemented separately through a communication chip.

[0185] The memory 120 may include a random access memory (RAM) or a read-only memory (ROM). The memory 120 may be used to store instructions, programs, codes, code sets or instruction sets. The memory 120 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the following various method embodiments, etc. The data storage area may also store data (such as a phone book, audio and video data, chat record data) created by the electronic device 100 during use.

[0186] Please refer to Fig.11 , which shows a structural block diagram of a computer-readable medium provided in an embodiment of the present application. The computer-readable medium 1100 stores program codes, which can be called by a processor to execute the method described in the above method embodiment.

[0187] The computer readable medium 1100 may be an electronic memory such as a flash memory, an EEPROM (electrically erasable programmable read-only memory), an EPROM, a hard disk, or a ROM. Optionally, the computer readable medium 1100 includes a non-transitory computer-readable storage medium. The computer readable medium 1100 has storage space for program code 1110 that performs any method step in the above method. These program codes can be read from or written to one or more computer program products. The program code 1110 can be compressed, for example, in an appropriate form.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A content query method, characterized in that: Applied to a search platform installed in an electronic device, the method comprises: Get the content to be queried; In the business data of at least one target business connected to the search platform, searching for business data matching the content to be queried based on a keyword search method and a semantic search method, respectively, to obtain a first search result corresponding to the keyword search method and a second search result corresponding to the semantic search method; Merging the first search result and the second search result to obtain a fused result; A target query result corresponding to the content to be queried is obtained based on the fusion result.

2. The method according to claim 1, characterized in that: The step of searching, in the business data of at least one target business connected to the search platform, business data matching the content to be queried based on a keyword search method and a semantic search method, respectively, to obtain a first search result corresponding to the keyword search method and a second search result corresponding to the semantic search method, comprises: Acquire a data index corresponding to business data of at least one target business accessing the search platform, wherein the data index includes first index information and second index information, wherein the first index information includes preset keywords corresponding to the business data, and the second index information includes preset semantic features corresponding to the business data; Based on the keyword search method, searching the first index information for business data matching the content to be queried, and obtaining the first search result; Based on the semantic search method, business data matching the to-be-queried content is searched in the second index information to obtain the second search result.

3. The method according to claim 2, characterized in that The keyword search method includes a first keyword search algorithm and a second keyword detection algorithm. Based on the keyword search method, searching the first index information for business data matching the content to be queried to obtain the first search result includes: Based on the first keyword search algorithm, searching the first index information for business data matching the content to be queried to obtain a first matching result; Based on the second keyword search algorithm, searching the first index information for business data matching the content to be queried to obtain a second matching result; The first matching result and the second matching result are combined to obtain the first search result.

4. The method according to claim 3, characterized in that The first keyword search algorithm is an exact matching algorithm, the second keyword search algorithm is a fuzzy matching algorithm, and in the first search result, the sorting priority of the second matching result is higher than the sorting priority of the first matching result.

5. The method according to claim 2, characterized in that: The acquiring of a data index corresponding to business data of at least one target business accessing the search platform includes: Determining whether a data index corresponding to business data of at least one target business accessing the search platform has been established; If it has been established, obtaining a data index corresponding to the business data of at least one target business connected to the search platform; If not established, establish a data index corresponding to the business data of at least one target business connected to the search platform.

6. The method according to claim 2, characterized in that Also includes: When a data index creation request is detected, determining whether the creation request is a first creation; If it is the first creation, then when it is determined that the electronic device meets the first condition, a data index creation operation is performed; If it is not the first creation, the data index creation operation is executed, and during the execution of the data index creation operation, if it is detected that the electronic device meets the second condition, the data index creation operation is interrupted.

7. The method according to claim 6, characterized in that If it is the first creation, then when it is determined that the electronic device meets the first condition, performing a data index creation operation includes: If it is the first creation, determine the target time period corresponding to the current time; If the target time period is within the daytime time period, determining the first condition includes at least one of the remaining power of the electronic device being greater than a first power threshold, the housing temperature of the electronic device being less than a first temperature threshold, the CPU load of the electronic device being less than a first load threshold, and the remaining storage space of the electronic device being greater than a first space threshold; If the target time period is within the night time period, determining the first condition includes at least one of the following: the electronic device is in an idle state, the electronic device is in a screen-off state, the electronic device is in a charging state and the remaining power is greater than a second power threshold, the housing temperature of the electronic device is less than a first temperature threshold, the CPU load of the electronic device is less than a first load threshold, and the remaining storage space of the electronic device is greater than a first space threshold; When it is determined that the electronic device satisfies the first condition, a data index creation operation is performed.

8. The method according to claim 6, characterized in that If it is not the first creation, then executing the data index creation operation, and in the process of executing the data index creation operation, if it is detected that the electronic device meets the second condition, interrupting the data index creation operation, including: If it is not the first creation, determine the target time period corresponding to the current time; If the target time period is within the daytime time period, determining the second condition includes that the housing temperature of the electronic device is greater than or equal to the first temperature threshold, the temperature difference of the housing temperature of the electronic device during the data index creation operation is greater than the second temperature threshold, or the power consumption of the electronic device during the data index creation operation exceeds the second power threshold; If the target time period is within the night time period, determining the second condition includes that the housing temperature of the electronic device is greater than or equal to the first temperature threshold, the power consumption of the electronic device during the data index creation operation exceeds the second power threshold, or the electronic device is in a screen-on state; The data index creation operation is executed, and during the execution of the data index creation operation, if it is detected that the electronic device meets the second condition, the data index creation operation is interrupted.

9. The method according to claim 2, characterized in that: Also includes: In case an update request for an established data index is detected, determining whether the update request is a first update; If it is the first update, then when it is determined that the electronic device meets the third condition, an update operation of the data index is performed; If it is not the first update, then when it is determined that the electronic device meets the fourth condition, an update operation of the data index is performed.

10. The method according to claim 9, characterized in that If it is the first update, then when it is determined that the electronic device meets the third condition, performing an update operation of the data index includes: If it is the first update, determining the third condition includes at least one of the remaining power of the electronic device being greater than the second power threshold, the CPU load of the electronic device being less than the first load threshold, and a search operation having been performed based on the data index within a specified time period; When it is determined that the electronic device satisfies the third condition, an update operation of the data index is performed.

11. The method according to claim 9, characterized in that If it is not the first update, then when it is determined that the electronic device meets the fourth condition, performing an update operation of the data index includes: If it is not the first update, the fourth condition is determined to include at least one of the following: the time difference between the current time and the time when the data index update operation was last performed is greater than the duration threshold, the number of executions of the update operation in the current time period corresponding to the current time period is less than the number threshold, the data to be updated this time is less than the specified number threshold, and the current time period is in the night time period; When it is determined that the electronic device satisfies the fourth condition, an update operation of the data index is performed.

12. The method according to claim 1, characterized in that The first search result includes a plurality of first matching data that are sequentially sorted, the second search result includes a plurality of second matching data that are sequentially sorted, and the fusion of the first search result and the second search result to obtain a fusion result includes: Determine a first weight corresponding to the first search result and a second weight corresponding to the second search result; Based on the first weight and the second weight, reorder the plurality of first matching data and the second matching data to obtain a shuffle result, wherein the shuffle result includes the plurality of matching data that are sequentially ordered; The top N matching data in the shuffled result are used as the fusion result, where N is an integer greater than 1.

13. The method according to claim 12, characterized in that The first weight and the second weight are the same.

14. The method according to claim 12, characterized in that The second weight is greater than the first weight.

15. The method according to claim 12, characterized in that The determining a first weight corresponding to the first search result and a second weight corresponding to the second search result includes: Determining a reference matching degree based on at least one of the keyword matching degrees in the first search result; If the reference matching degree is less than a preset threshold, setting the second weight to be greater than the first weight; If the reference matching degree is greater than or equal to a preset threshold, the first weight is set to be the same as the second weight.

16. The method according to claim 15, characterized in that The determining a reference matching degree based on at least one of the keyword matching degrees in the first search result includes: The maximum keyword matching degree in the first search result is determined as a reference matching degree.

17. The method according to any one of claims 1 to 16, characterized in that: The obtaining, based on the fusion result, a target query result corresponding to the to-be-queried content includes: Based on the fusion result, the output result of the large language model is obtained through the large language model, and the target query result corresponding to the content to be queried is obtained based on the output result.

18. The method according to claim 17, characterized in that The step of obtaining an output result of the large language model based on the fusion result by using the large language model, and obtaining a target query result corresponding to the content to be queried based on the output result includes: The fusion result is sent to the cloud, triggering the cloud to generate prompt information based on the fusion result, inputting the prompt information into a large language model, and obtaining a target query result corresponding to the content to be queried based on an output result of the large language model.

19. A content query device, characterized in that: The device is applied to a search platform installed in an electronic device, and comprises: An acquisition unit, used for acquiring content to be queried; A search unit, configured to search for business data matching the content to be queried in the business data of at least one target business connected to the search platform based on a keyword search method and a semantic search method, respectively, to obtain a first search result corresponding to the keyword search method and a second search result corresponding to the semantic search method; A fusion unit, configured to fuse the first search result and the second search result to obtain a fusion result; A processing unit is used to obtain a target query result corresponding to the content to be queried based on the fusion result.

20. An electronic device, characterized in that: include: one or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to execute the method according to any one of claims 1-18.

21. A computer readable medium, characterized in that The computer-readable medium stores a program code executable by a processor, and when the program code is executed by the processor, the processor executes the method according to any one of claims 1 to 18.

Images