An information consultation intelligent retrieval system and method based on cloud service

Abstract

The application relates to the field of information retrieval, in particular to an information consultation intelligent retrieval system and method based on cloud services, which comprises a cross-table index module, a retrieval characteristic module, a sentence processing module, a granularity retrieval module and a resource allocation module; the cross-table index module is used for constructing an index structure; the retrieval characteristic module is used for synthesizing a characteristic vector and outputting a retrieval basis vector; the sentence processing module is used for receiving a consultation text and projecting a granularity text; the granularity retrieval module is used for retrieving resources and outputting a retrieval result; and the resource allocation module is used for constructing a vector tree and allocating retrieval resources; the application can improve retrieval precision in a domain, relieve data expansion pressure of a large-scale database, enhance retrieval generalization capability, improve comprehensiveness of retrieval content, reduce time complexity, facilitate maintenance and expansion, reduce the cost of retrieval semantic calculation, improve retrieval efficiency, and realize more stable retrieval performance.

Description

Technical Field

[0001] This invention relates to the field of information retrieval, specifically to an intelligent information retrieval system and method based on cloud services. Background Technology

[0002] Information retrieval systems are technologies that search for and return information needed by users from large-scale databases. By providing accurate database query results, they filter out valuable information from a large amount of data and are widely used in search engines, recommendation systems, and large-scale model applications. Data in large-scale databases is usually stored in the form of tables or sub-databases. Data classification limits the retrieval space and improves retrieval efficiency.

[0003] With the exponential growth of data volume, database structures based on intensive retrieval are prone to significant latency due to the large number of data table types, leading to information overload. Furthermore, some cross-domain retrieval problems struggle to obtain accurate results. When faced with natural language queries, language models are required for indexing and compilation. However, the retrieval process is often plagued by language bias, resulting in insufficient semantic understanding and inadequate command differentiation. This leads to low utilization of training data and unstable retrieval performance.

[0004] Furthermore, for cloud-based query services, the concurrency of search requests is high. Local information retrieval systems need to retrieve content from a large amount of text to be searched, which can easily lead to imbalances in the allocation of computing resources. This results in large fluctuations in the quality of the data generated from the search results, low tolerance for query noise, reduced utilization of storage resources, increased computing costs, and reduced retrieval efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent information retrieval system and method based on cloud services to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, the present invention provides the following technical solution: an intelligent information retrieval system based on cloud services, comprising: a cross-table indexing module, a retrieval feature module, a statement processing module, a granular retrieval module, and a resource allocation module;

[0007] The cross-table index module is used to build an index structure within the database, record the position of each data item in the data table fields, use a composite anomaly graph to test the correlation between data table nodes, build a primary and foreign key relationship table based on the relationship between data tables, construct cross-table query statements, and perform vector encoding on the semantic text in the primary and foreign key relationship table to obtain feature vectors.

[0008] The retrieval feature module is used to synthesize feature vectors between related tables using the alternating language method. The synthesis ratio is determined by the ratio of the related data volume. Alternating sentence vectors are generated, and the training corpus is reconstructed using the alternating sentence vectors. Through confusing negative sampling, a pre-retrieval model containing a projection layer and a nonlinear function layer is trained so that the model outputs retrieval basis vectors for data items that are related.

[0009] The statement processing module is used to receive the consultation statement input by the user from the cloud, compile the consultation statement using a large language model, use the search command as a positive sample and the non-search command as a negative sample, decompose the statement through comparative learning, extract the search words, assign feature values ​​to the search words of different parts of speech, aggregate the granular text according to the feature values, and project it into the database according to the text type to obtain the text vector.

[0010] The granularity retrieval module is used to perform similar vector search on text vectors and retrieval basis vectors, associate each granular text and data item, perform single table retrieval when there is no association, allocate retrieval resources for each associated data item according to the degree of association through the nearest neighbor index, output all retrieval results, construct the query database corresponding to the user using the retrieval results, generate new retrieval basis vectors using the database, compile the database and output retrieval results.

[0011] The resource allocation module is used to model the text retrieval process based on the nearest neighbor index corresponding to the text. Through a tree convolutional network, with the text vector as the root node, each retrieval basis vector as a branch node, and all data tables corresponding to the retrieval basis vectors as leaf nodes, indexes are inserted in the branch nodes. Information from different index columns is aggregated to represent multi-column indexes, and a vector tree is constructed. A resource cost model is established based on the amount of retrieval data in each branch of the vector tree. The input query statement is evaluated, the computational cost of the query statement is assessed, and the query time under standard computational power is output as the query's runtime cost for resource allocation.

[0012] Furthermore, the cross-table indexing module includes: a data testing unit and a relation processing unit;

[0013] The data testing unit is used to create an index table in the database, record the value of the data item, table name, field name and primary key information, use a graph database to record the primary key and foreign key information of each pair of related tables, generate SQL join query statements based on the primary and foreign key relationship table, and merge the related table data.

[0014] The association processing unit is used to construct a composite anomaly graph, where nodes represent each data table, edges represent the associated data items between tables, and the length of the edge represents the amount of associated data. Nodes with edge lengths greater than a threshold are used as associated tables. The BERT language model is used to extract the table name, field name, and annotations corresponding to the semantic text of the data items, and the semantic text in the cross-table query results is converted into numerical vectors.

[0015] Furthermore, the retrieval feature module includes: a semantic synthesis unit and a vector matching unit;

[0016] The semantic synthesis unit is used to alternately combine the text of the associated fields of the primary key table and the foreign key table for each primary-foreign key relationship, determine the alternation ratio and the length of the alternating text based on the ratio of the amount of associated data in the two tables to the total amount of data in the tables, encode the alternating text, and obtain an alternating statement vector.

[0017] The vector matching unit is used to take the alternating statement vector as a positive sample, randomly replace the fields in the association table to generate negative samples, construct a training set, train the pre-retrieval model, make the positive samples close to the output result, and the negative samples far from the output result, and output the retrieval basis vector.

[0018] Furthermore, the statement processing module includes: a cloud interaction unit, a word segmentation and compilation unit, and a data projection unit;

[0019] The cloud-based interactive unit is used to build a cloud-based interactive platform. Users input consultation statements into the interactive window, and the large language model is used for comparative learning to classify the input text, obtain valid search instructions, and use the spaCy word segmentation algorithm to segment the search instructions to obtain a list of fields.

[0020] The word segmentation and compilation unit is used to label the part-of-speech of each field in the field list according to the degree of similarity of the pre-retrieval information. The pre-retrieval information includes: data table fields, data item names, time ranges and storage locations. Based on the part-of-speech features, each retrieval word segment is aggregated into a larger granular text unit to form retrieval granular text.

[0021] The data projection unit is used to map each granular text to a database table, and to project the granular text according to the correlation between the granular text and each table. The projected text is then vectorized to obtain a text vector, where each element in the text vector represents the correlation between the granular text and the corresponding data table.

[0022] Furthermore, the granularity retrieval module includes: an association retrieval unit and a resource update unit;

[0023] The associated retrieval unit is used to calculate the cosine similarity or Euclidean distance between the text vector and the retrieval basis vector using the nearest neighbor search algorithm, sort the data items according to the calculation results, output the data table or primary-foreign key relationship table most relevant to the text vector, perform retrieval within the table, and output the retrieval results.

[0024] The resource update unit is used to generate a query data table based on the user's search results, record the fields, conditions and return results of the user's query, construct a new search basis vector using the generated query data table, fit it with a large model, upload the search results to the cloud platform in natural language and output them to the user.

[0025] Furthermore, the resource allocation module includes: a process modeling unit, an information aggregation unit, and a resource allocation unit;

[0026] The process modeling unit is used to model the text retrieval process through a vector tree, establish a connection between each branch node and the index column, and use a convolutional network to weight and combine the features of each layer node and the information of the parent and child nodes and insert them into the index, so that the text vector can be quickly matched with the data table through the index.

[0027] The information aggregation unit is used to insert multiple index columns in the branch node for queries containing multiple search conditions. For each search process, it calculates the data volume overhead of each search process based on the total search criteria vector and data table field index data volume and index efficiency of the vector tree.

[0028] The resource allocation unit is used to simulate the query process based on a vector tree structure after the input query statement, calculate the computing power requirement for the query statement to be executed in the database by the data volume overhead and the length of the query text, output the execution time of the query under standard computing power conditions, and dynamically adjust the allocation of cloud transmission or local computing power resources.

[0029] A cloud-based intelligent information retrieval method includes the following steps:

[0030] Step S1. Construct an index structure in the database, record the position of each data item in the data table, test the correlation between data table nodes, construct a primary and foreign key relationship table, and perform vector encoding on the semantic text in the primary and foreign key relationship table to obtain feature vectors;

[0031] Step S2. Synthesize the feature vector between related data tables by the ratio of the amount of related data, generate alternating statement vectors, train the pre-retrieval model, input data items with related relationships, merge the alternating statement vectors in the tables where the data items are located, and output the retrieval basis vector;

[0032] Step S3. Receive the user's input query statement, compile and decompose the query statement, extract the search words, assign feature values ​​to the search words with different parts of speech, aggregate the granular text according to the feature values, project the granular text into the database, and obtain the text vector;

[0033] Step S4. Using the nearest neighbor index, perform similarity vector calculation on the text vector and the retrieval basis vector, sort the data items according to the calculation results, output the data related to the text vector, construct the query database using the retrieval results, generate new retrieval basis vectors, and output the retrieval results;

[0034] Step S5. Using a tree convolutional network, with the text vector as the root node, each retrieval basis vector as a branch node, and all data tables corresponding to the retrieval basis vectors as leaf nodes, an index is inserted into the branch nodes to construct a vector tree. The query process is simulated based on the vector tree structure, and the data volume cost is output. The allocation of retrieval resources is dynamically adjusted according to the data volume cost.

[0035] Furthermore, step S1 includes:

[0036] Step S11. Create an index table in the database to record the values ​​of data items, table names, field names, and primary key information. Use a graph database to record the primary key and foreign key information of each pair of related tables. Generate an SQL join query statement based on the primary and foreign key relationship table and merge the related table data.

[0037] Step S12. Construct a composite anomaly graph, where nodes represent each data table, edges represent related data items between tables, and the length of the edge represents the amount of related data. Nodes with edge lengths greater than a threshold are designated as related tables. The BERT language model is used to extract the table name, field name, and comments corresponding to the semantic text of the data items. The semantic text in the cross-table query results is converted into numerical vectors.

[0038] Furthermore, step S2 includes:

[0039] Step S21. For each primary-foreign key relationship, alternately combine the text of the related fields in the primary key table and the foreign key table. Determine the alternation ratio and the length of the alternation text based on the ratio of the amount of related data in the two tables to the total amount of data in the tables. Encode the alternation text to obtain the alternation statement vector.

[0040] Step S22. Use the alternating statement vector as positive samples, randomly replace the fields in the association table to generate negative samples, build a training set, train the pre-retrieval model, make the positive samples close to the output results, and the negative samples far from the output results, and output the retrieval basis vector.

[0041] Furthermore, step S3 includes:

[0042] Step S31. Build a cloud-based interactive platform. Users input their query statements into the interactive window. The large language model is used for comparative learning to classify the input text, obtain valid search instructions, and use the spaCy word segmentation algorithm to segment the search instructions to obtain a list of fields.

[0043] Step S32. Label the part-of-speech tags of each field in the field list according to the degree of similarity of the pre-retrieval information. The pre-retrieval information includes: data table fields, data item names, time ranges and storage locations. Based on the part-of-speech features, aggregate each retrieval word segment into a larger granular text unit to form retrieval granular text.

[0044] Step S33. Map each granular text to a database table, and project it according to the correlation between the granular text and each table. Vectorize the projected text to obtain a text vector. Each element in the text vector represents the correlation between the granular text and the corresponding data table.

[0045] Furthermore, step S4 includes:

[0046] Step S41. Use the nearest neighbor search algorithm to calculate the cosine similarity or Euclidean distance of the text vector and the retrieval basis vector, sort the data items according to the calculation results, output the data table or primary-foreign key relationship table most relevant to the text vector, perform the retrieval in the table, and output the retrieval results;

[0047] Step S42. Based on the user's search results, generate a query data table, record the fields, conditions and results of the user's query, construct a new search basis vector using the generated query data table, fit it with a large model, upload the search results to the cloud platform in natural language and output them to the user.

[0048] Furthermore, step S5 includes:

[0049] Step S51. Model the text retrieval process using a vector tree, establish a connection between each branch node and the index column, and use a convolutional network to weight and combine the features of each layer node with the information of the parent and child nodes and insert them into the index, so that the text vector can be quickly matched with the data table through the index.

[0050] Step S52. For queries containing multiple search conditions, insert multi-column indexes within the branch nodes. For each search process, calculate the data volume overhead of each search process based on the total search criteria vector and data table fields of the vector tree, as well as the index data volume and index efficiency.

[0051] Step S53. After inputting the query statement, simulate the query process based on the vector tree structure, calculate the computing power requirement for executing the query statement in the database by the data volume overhead and the length of the query text, output the execution time of the query under standard computing power conditions, and dynamically adjust the allocation of cloud transmission or local computing power resources.

[0052] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0053] 1. This invention provides a method for constructing an index structure within a database, building a primary and foreign key relationship table based on the relationships between data tables, constructing cross-table query statements, synthesizing feature vectors between related tables, training a pre-retrieval model, adapting vector representations for different retrieval scenarios, improving the retrieval model's ability to distinguish similar data, improving intra-domain retrieval accuracy, alleviating the data expansion pressure of large-scale databases, enhancing generalization ability in cross-domain data retrieval, and exhibiting high scalability and adaptability.

[0054] 2. This invention uses a large language model to compile consultation statements, thereby decomposing the statements, extracting search terms, allocating search resources according to the degree of relevance, and outputting search results. It optimizes the instruction decomposition process, reduces instruction noise, improves the comprehensiveness of search content while maintaining search accuracy, has low time complexity, is easy to maintain and expand, reduces the overhead of search semantic calculation, improves search efficiency, and enhances the scalability of the database and the knowledge transfer capability of the consultation process.

[0055] 3. This invention models the text retrieval process, constructs a vector tree, establishes a resource cost model based on the retrieval data volume of the vector tree, evaluates the computational cost of each query statement, determines the runtime cost of the query, and thus intelligently allocates computing resources. This achieves end-to-end optimization of cloud platform retrieval services, resulting in more stable retrieval performance and improved user retrieval satisfaction. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a schematic diagram of the structure of an intelligent information retrieval system based on cloud services according to the present invention;

[0058] Figure 2 This is a schematic diagram illustrating the steps of an intelligent information retrieval method based on cloud services according to the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Please see Figures 1 to 2 The present invention provides a technical solution: an intelligent information retrieval system based on cloud services, comprising: a cross-table indexing module, a retrieval feature module, a statement processing module, a granular retrieval module, and a resource allocation module;

[0061] The cross-table index module is used to build an index structure within the database, record the position of each data item in the data table fields, use a composite anomaly graph to test the correlation between data table nodes, build a primary and foreign key relationship table based on the relationship between data tables, construct cross-table query statements, and perform vector encoding on the semantic text in the primary and foreign key relationship table to obtain feature vectors.

[0062] The cross-table index module includes: a data testing unit and a relation processing unit;

[0063] The data testing unit is used to create an index table in the database, record the value of the data item, table name, field name and primary key information, use a graph database to record the primary key and foreign key information of each pair of related tables, generate SQL join query statements based on the primary and foreign key relationship table, and merge the related table data.

[0064] The association processing unit is used to construct a composite anomaly graph, where nodes represent each data table, edges represent the associated data items between tables, and the length of the edge represents the amount of associated data. Nodes with edge lengths greater than a threshold are used as associated tables. The BERT language model is used to extract the table name, field name, and annotations corresponding to the semantic text of the data items, and the semantic text in the cross-table query results is converted into numerical vectors.

[0065] The retrieval feature module is used to synthesize feature vectors between related tables using the alternating language method. The synthesis ratio is determined by the ratio of the related data volume. Alternating sentence vectors are generated, and the training corpus is reconstructed using the alternating sentence vectors. Through confusing negative sampling, a pre-retrieval model containing a projection layer and a nonlinear function layer is trained so that the model outputs retrieval basis vectors for data items that are related.

[0066] The retrieval feature module includes: a semantic synthesis unit and a vector matching unit;

[0067] The semantic synthesis unit is used to alternately combine the text of the associated fields of the primary key table and the foreign key table for each primary-foreign key relationship, determine the alternation ratio and the length of the alternating text based on the ratio of the amount of associated data in the two tables to the total amount of data in the tables, encode the alternating text, and obtain an alternating statement vector.

[0068] The vector matching unit is used to take the alternating statement vector as a positive sample, randomly replace the fields in the association table to generate negative samples, construct a training set, train the pre-retrieval model, make the positive samples close to the output result, and the negative samples far from the output result, and output the retrieval basis vector.

[0069] The statement processing module is used to receive the consultation statement input by the user from the cloud, compile the consultation statement using a large language model, use the search command as a positive sample and the non-search command as a negative sample, decompose the statement through comparative learning, extract the search words, assign feature values ​​to the search words of different parts of speech, aggregate the granular text according to the feature values, and project it into the database according to the text type to obtain the text vector.

[0070] The statement processing module includes: a cloud interaction unit, a word segmentation and compilation unit, and a data projection unit;

[0071] The cloud-based interactive unit is used to build a cloud-based interactive platform. Users input consultation statements into the interactive window, and the large language model is used for comparative learning to classify the input text, obtain valid search instructions, and use the spaCy word segmentation algorithm to segment the search instructions to obtain a list of fields.

[0072] The word segmentation and compilation unit is used to label the part-of-speech of each field in the field list according to the degree of similarity of the pre-retrieval information. The pre-retrieval information includes: data table fields, data item names, time ranges and storage locations. Based on the part-of-speech features, each retrieval word segment is aggregated into a larger granular text unit to form retrieval granular text.

[0073] The data projection unit is used to map each granular text to a database table, and to project the granular text according to the correlation between the granular text and each table. The projected text is then vectorized to obtain a text vector, where each element in the text vector represents the correlation between the granular text and the corresponding data table.

[0074] The granularity retrieval module is used to perform similar vector search on text vectors and retrieval basis vectors, associate each granular text and data item, perform single table retrieval when there is no association, allocate retrieval resources for each associated data item according to the degree of association through the nearest neighbor index, output all retrieval results, construct the query database corresponding to the user using the retrieval results, generate new retrieval basis vectors using the database, compile the database and output retrieval results.

[0075] The granularity retrieval module includes: an association retrieval unit and a resource update unit;

[0076] The associated retrieval unit is used to calculate the cosine similarity or Euclidean distance between the text vector and the retrieval basis vector using the nearest neighbor search algorithm, sort the data items according to the calculation results, output the data table or primary-foreign key relationship table most relevant to the text vector, perform retrieval within the table, and output the retrieval results.

[0077] The resource update unit is used to generate a query data table based on the user's search results, record the fields, conditions and return results of the user's query, construct a new search basis vector using the generated query data table, fit it with a large model, upload the search results to the cloud platform in natural language and output them to the user.

[0078] The resource allocation module is used to model the text retrieval process based on the nearest neighbor index corresponding to the text. Through a tree convolutional network, with the text vector as the root node, each retrieval basis vector as a branch node, and all data tables corresponding to the retrieval basis vectors as leaf nodes, indexes are inserted in the branch nodes. Information from different index columns is aggregated to represent multi-column indexes, and a vector tree is constructed. A resource cost model is established based on the amount of retrieval data in each branch of the vector tree. The input query statement is evaluated, the computational cost of the query statement is assessed, and the query time under standard computational power is output as the query's runtime cost for resource allocation.

[0079] The resource allocation module includes: a process modeling unit, an information aggregation unit, and a resource allocation unit;

[0080] The process modeling unit is used to model the text retrieval process through a vector tree, establish a connection between each branch node and the index column, and use a convolutional network to weight and combine the features of each layer node and the information of the parent and child nodes and insert them into the index, so that the text vector can be quickly matched with the data table through the index.

[0081] The information aggregation unit is used to insert multiple index columns in the branch node for queries containing multiple search conditions. For each search process, it calculates the data volume overhead of each search process based on the total search criteria vector and data table field index data volume and index efficiency of the vector tree.

[0082] The resource allocation unit is used to simulate the query process based on a vector tree structure after the input query statement, calculate the computing power requirement for the query statement to be executed in the database by the data volume overhead and the length of the query text, output the execution time of the query under standard computing power conditions, and dynamically adjust the allocation of cloud transmission or local computing power resources.

[0083] A cloud-based intelligent information retrieval method includes the following steps:

[0084] Step S1. Construct an index structure in the database, record the position of each data item in the data table, test the correlation between data table nodes, construct a primary and foreign key relationship table, and perform vector encoding on the semantic text in the primary and foreign key relationship table to obtain feature vectors;

[0085] Step S1 includes:

[0086] Step S11. Create an index table in the database to record the values ​​of data items, table names, field names, and primary key information. Use a graph database to record the primary key and foreign key information of each pair of related tables. Generate an SQL join query statement based on the primary and foreign key relationship table and merge the related table data.

[0087] Step S12. Construct a composite anomaly graph, where nodes represent each data table, edges represent related data items between tables, and the length of the edge represents the amount of related data. Nodes with edge lengths greater than a threshold are designated as related tables. The BERT language model is used to extract the table name, field name, and comments corresponding to the semantic text of the data items. The semantic text in the cross-table query results is converted into numerical vectors.

[0088] Step S2. Synthesize the feature vector between related data tables by the ratio of the amount of related data, generate alternating statement vectors, train the pre-retrieval model, input data items with related relationships, merge the alternating statement vectors in the tables where the data items are located, and output the retrieval basis vector;

[0089] Step S2 includes:

[0090] Step S21. For each primary-foreign key relationship, alternately combine the text of the related fields in the primary key table and the foreign key table. Determine the alternation ratio and the length of the alternation text based on the ratio of the amount of related data in the two tables to the total amount of data in the tables. Encode the alternation text to obtain the alternation statement vector.

[0091] Step S22. Use the alternating statement vector as positive samples, randomly replace the fields in the association table to generate negative samples, build a training set, train the pre-retrieval model, make the positive samples close to the output results, and the negative samples far from the output results, and output the retrieval basis vector.

[0092] Step S3. Receive the user's input query statement, compile and decompose the query statement, extract the search words, assign feature values ​​to the search words with different parts of speech, aggregate the granular text according to the feature values, project the granular text into the database, and obtain the text vector;

[0093] Step S3 includes:

[0094] Step S31. Build a cloud-based interactive platform. Users input their query statements into the interactive window. The large language model is used for comparative learning to classify the input text, obtain valid search instructions, and use the spaCy word segmentation algorithm to segment the search instructions to obtain a list of fields.

[0095] Step S32. Label the part-of-speech tags of each field in the field list according to the degree of similarity of the pre-retrieval information. The pre-retrieval information includes: data table fields, data item names, time ranges and storage locations. Based on the part-of-speech features, aggregate each retrieval word segment into a larger granular text unit to form retrieval granular text.

[0096] Step S33. Map each granular text to a database table, and project it according to the correlation between the granular text and each table. Vectorize the projected text to obtain a text vector. Each element in the text vector represents the correlation between the granular text and the corresponding data table.

[0097] Step S4. Using the nearest neighbor index, perform similarity vector calculation on the text vector and the retrieval basis vector, sort the data items according to the calculation results, output the data related to the text vector, construct the query database using the retrieval results, generate new retrieval basis vectors, and output the retrieval results;

[0098] Step S4 includes:

[0099] Step S41. Use the nearest neighbor search algorithm to calculate the cosine similarity or Euclidean distance of the text vector and the retrieval basis vector, sort the data items according to the calculation results, output the data table or primary-foreign key relationship table most relevant to the text vector, perform the retrieval in the table, and output the retrieval results;

[0100] Step S42. Based on the user's search results, generate a query data table, record the fields, conditions and results of the user's query, construct a new search basis vector using the generated query data table, fit it with a large model, upload the search results to the cloud platform in natural language and output them to the user.

[0101] Step S5. Using a tree convolutional network, with the text vector as the root node, each retrieval basis vector as a branch node, and all data tables corresponding to the retrieval basis vectors as leaf nodes, an index is inserted into the branch nodes to construct a vector tree. The query process is simulated based on the vector tree structure, and the data volume cost is output. The allocation of retrieval resources is dynamically adjusted according to the data volume cost.

[0102] Step S5 includes:

[0103] Step S51. Model the text retrieval process using a vector tree, establish a connection between each branch node and the index column, and use a convolutional network to weight and combine the features of each layer node with the information of the parent and child nodes and insert them into the index, so that the text vector can be quickly matched with the data table through the index.

[0104] Step S52. For queries containing multiple search conditions, insert multi-column indexes within the branch nodes. For each search process, calculate the data volume overhead of each search process based on the total search criteria vector and data table fields of the vector tree, as well as the index data volume and index efficiency.

[0105] Step S53. After inputting the query statement, simulate the query process based on the vector tree structure, calculate the computing power requirement for executing the query statement in the database by the data volume overhead and the length of the query text, output the execution time of the query under standard computing power conditions, and dynamically adjust the allocation of cloud transmission or local computing power resources.

[0106] Example: Construct a database index, record the location of data items, test the correlation between data table nodes, construct a primary and foreign key relationship table, perform vector encoding on semantic text, synthesize feature vectors between related tables using alternating language method, output retrieval basis vectors, user input query statements, extract retrieval words, aggregate granular text to obtain text vectors, retrieve related data items, output retrieval results and construct user query database, generate new retrieval basis vectors, and output retrieval results.

[0107] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0108] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cloud-based intelligent information retrieval method, characterized in that, The method includes the following steps: Step S1. Construct an index structure in the database, record the position of each data item in the data table, test the correlation between data table nodes, construct a primary and foreign key relationship table, and perform vector encoding on the semantic text in the primary and foreign key relationship table to obtain feature vectors; Step S2. Synthesize the feature vector between related data tables by the ratio of the amount of related data, generate alternating statement vectors, train the pre-retrieval model, input data items with related relationships, merge the alternating statement vectors in the tables where the data items are located, and output the retrieval basis vector; Step S3. Receive the user's input query statement, compile and decompose the query statement, extract the search words, assign feature values ​​to the search words with different parts of speech, aggregate the granular text according to the feature values, project the granular text into the database, and obtain the text vector; Step S4. Using the nearest neighbor index, perform similarity vector calculation on the text vector and the retrieval basis vector, sort the data items according to the calculation results, output the data related to the text vector, construct the query database using the retrieval results, generate new retrieval basis vectors and output the retrieval results; Step S5. Using a tree convolutional network, with the text vector as the root node, each retrieval basis vector as a branch node, and all data tables corresponding to the retrieval basis vectors as leaf nodes, an index is inserted into the branch nodes to construct a vector tree. The query process is simulated based on the vector tree structure, and the data volume cost is output. The allocation of retrieval resources is dynamically adjusted according to the data volume cost. Step S2 includes: Step S21. For each primary-foreign key relationship, alternately combine the text of the related fields in the primary key table and the foreign key table. Determine the alternation ratio and the length of the alternation text based on the ratio of the amount of related data in the two tables to the total amount of data in the tables. Encode the alternation text to obtain the alternation statement vector. Step S22. Use the alternating statement vector as positive samples, randomly replace the fields in the association table to generate negative samples, build a training set, train the pre-retrieval model, make the positive samples close to the output results, and the negative samples far from the output results, and output the retrieval basis vector.

2. The intelligent information retrieval method based on cloud services according to claim 1, characterized in that: Step S1 includes: Step S11. Create an index table in the database to record the values ​​of data items, table names, field names, and primary key information. Use a graph database to record the primary key and foreign key information of each pair of related tables. Generate an SQL join query statement based on the primary and foreign key relationship table and merge the related table data. Step S12. Construct a composite anomaly graph, where nodes represent each data table, edges represent related data items between tables, and the length of the edge represents the amount of related data. Nodes with edge lengths greater than a threshold are designated as related tables. The BERT language model is used to extract the table name, field name, and comments corresponding to the semantic text of the data items. The semantic text in the cross-table query results is converted into numerical vectors.

3. The intelligent information retrieval method based on cloud services according to claim 2, characterized in that: Step S3 includes: Step S31. Build a cloud-based interactive platform. Users input their query statements into the interactive window. The large language model is used for comparative learning to classify the input text, obtain valid search instructions, and use the spaCy word segmentation algorithm to segment the search instructions to obtain a list of fields. Step S32. Label the part-of-speech tags of each field in the field list according to the degree of similarity of the pre-retrieval information. The pre-retrieval information includes: data table fields, data item names, time ranges and storage locations. Based on the part-of-speech features, aggregate each retrieval word segment into a larger granular text unit to form retrieval granular text. Step S33. Map each granular text to a database table, and project it according to the correlation between the granular text and each table. Vectorize the projected text to obtain a text vector. Each element in the text vector represents the correlation between the granular text and the corresponding data table.

4. The intelligent information retrieval method based on cloud services according to claim 3, characterized in that: Step S4 includes: Step S41. Use the nearest neighbor search algorithm to calculate the cosine similarity or Euclidean distance of the text vector and the retrieval basis vector, sort the data items according to the calculation results, output the data table or primary-foreign key relationship table most relevant to the text vector, perform the retrieval in the table, and output the retrieval results; Step S42. Based on the user's search results, generate a query data table, record the fields, conditions and results of the user's query, construct a new search basis vector using the generated query data table, fit it with a large model, upload the search results to the cloud platform in natural language and output them to the user.

5. The intelligent information retrieval method based on cloud services according to claim 4, characterized in that: Step S5 includes: Step S51. Model the text retrieval process using a vector tree, establish a connection between each branch node and the index column, and use a convolutional network to weight and combine the features of each layer node with the information of the parent and child nodes and insert them into the index, so that the text vector can be quickly matched with the data table through the index. Step S52. For queries containing multiple search conditions, insert multi-column indexes within the branch nodes. For each search process, calculate the data volume overhead of each search process based on the total search criteria vector and data table fields of the vector tree, as well as the index data volume and index efficiency. Step S53. After inputting the query statement, simulate the query process based on the vector tree structure, calculate the computing power requirement for the query statement to be executed in the database by the data volume overhead and the length of the query text, output the execution time of the query under standard computing power conditions, and dynamically adjust the allocation of cloud transmission or local computing power resources.

6. A cloud-based intelligent information retrieval system, characterized in that, The system includes the following modules: cross-table index module, retrieval feature module, statement processing module, granular retrieval module, and resource allocation module; The cross-table index module is used to build an index structure within the database, record the position of each data item in the data table fields, use a composite anomaly graph to test the correlation between data table nodes, build a primary and foreign key relationship table based on the relationship between data tables, construct cross-table query statements, and perform vector encoding on the semantic text in the primary and foreign key relationship table to obtain feature vectors. The retrieval feature module is used to synthesize feature vectors between related tables using the alternating language method. The synthesis ratio is determined by the ratio of the related data volume. Alternating sentence vectors are generated, and the training corpus is reconstructed using the alternating sentence vectors. Through confusing negative sampling, a pre-retrieval model containing a projection layer and a nonlinear function layer is trained so that the model outputs retrieval basis vectors for data items that are related. The statement processing module is used to receive the consultation statement input by the user from the cloud, compile the consultation statement using a large language model, use the search command as a positive sample and the non-search command as a negative sample, decompose the statement through comparative learning, extract the search words, assign feature values ​​to the search words of different parts of speech, aggregate the granular text according to the feature values, and project it into the database according to the text type to obtain the text vector. The granularity retrieval module is used to perform similar vector search on text vectors and retrieval basis vectors, associate each granular text and data item, perform single table retrieval when there is no association, allocate retrieval resources for each associated data item according to the degree of association through the nearest neighbor index, output all retrieval results, construct the query database corresponding to the user using the retrieval results, generate new retrieval basis vectors using the database, compile the database and output retrieval results. The resource allocation module is used to model the text retrieval process based on the nearest neighbor index corresponding to the text. Through a tree convolutional network, with the text vector as the root node, each retrieval basis vector as a branch node, and all data tables corresponding to the retrieval basis vectors as leaf nodes, indexes are inserted in the branch nodes. Information from different index columns is aggregated to represent multi-column indexes, and a vector tree is constructed. A resource cost model is established based on the amount of retrieval data in each branch of the vector tree. The input query statement is evaluated, the computational cost of the query statement is assessed, and the query time under standard computational power is output as the query running cost for resource allocation. The retrieval feature module includes: a semantic synthesis unit and a vector matching unit; The semantic synthesis unit is used to alternately combine the text of the associated fields of the primary key table and the foreign key table for each primary-foreign key relationship, determine the alternation ratio and the length of the alternating text based on the ratio of the amount of associated data in the two tables to the total amount of data in the tables, encode the alternating text, and obtain an alternating statement vector. The vector matching unit is used to take the alternating statement vector as a positive sample, randomly replace the fields in the association table to generate negative samples, construct a training set, train the pre-retrieval model, make the positive samples close to the output result, and the negative samples far from the output result, and output the retrieval basis vector.

7. The cloud-based intelligent information retrieval system according to claim 6, characterized in that: The cross-table index module includes: a data testing unit and a relation processing unit; The data testing unit is used to create an index table in the database, record the value of the data item, table name, field name and primary key information, use a graph database to record the primary key and foreign key information of each pair of related tables, generate SQL join query statements based on the primary and foreign key relationship table, and merge the related table data. The association processing unit is used to construct a composite anomaly graph, where nodes represent each data table, edges represent the associated data items between tables, and the length of the edge represents the amount of associated data. Nodes with edge lengths greater than a threshold are used as associated tables. The BERT language model is used to extract the table name, field name, and annotations corresponding to the semantic text of the data items, and the semantic text in the cross-table query results is converted into numerical vectors.

8. The cloud-based intelligent information retrieval system according to claim 7, characterized in that: The statement processing module includes: a cloud interaction unit, a word segmentation and compilation unit, and a data projection unit; The cloud-based interactive unit is used to build a cloud-based interactive platform. Users input consultation statements into the interactive window, and the large language model is used for comparative learning to classify the input text, obtain valid search instructions, and use the spaCy word segmentation algorithm to segment the search instructions to obtain a list of fields. The word segmentation and compilation unit is used to label the part-of-speech of each field in the field list according to the degree of similarity of the pre-retrieval information. The pre-retrieval information includes: data table fields, data item names, time ranges and storage locations. Based on the part-of-speech features, each retrieval word segment is aggregated into a larger granular text unit to form retrieval granular text. The data projection unit is used to map each granular text to a database table, and to project the granular text according to the correlation between the granular text and each table. The projected text is then vectorized to obtain a text vector, where each element in the text vector represents the correlation between the granular text and the corresponding data table.

9. The cloud-based intelligent information retrieval system according to claim 8, characterized in that: The granularity retrieval module includes: an association retrieval unit and a resource update unit; The associated retrieval unit is used to calculate the cosine similarity or Euclidean distance between the text vector and the retrieval basis vector using the nearest neighbor search algorithm, sort the data items according to the calculation results, output the data table or primary-foreign key relationship table most relevant to the text vector, perform retrieval within the table, and output the retrieval results. The resource update unit is used to generate a query data table based on the user's search results, record the fields, conditions and return results of the user's query, construct a new search basis vector using the generated query data table, fit it with a large model, upload the search results to the cloud platform in natural language and output them to the user.

10. The cloud-based intelligent information retrieval system according to claim 9, characterized in that: The resource allocation module includes: a process modeling unit, an information aggregation unit, and a resource allocation unit; The process modeling unit is used to model the text retrieval process through a vector tree, establish a connection between each branch node and the index column, and use a convolutional network to weight and combine the features of each layer node and the information of the parent and child nodes and insert them into the index, so that the text vector can be quickly matched with the data table through the index. The information aggregation unit is used to insert multiple index columns in the branch node for queries containing multiple search conditions. For each search process, it calculates the data volume overhead of each search process based on the total search criteria vector and data table field index data volume and index efficiency of the vector tree. The resource allocation unit is used to simulate the query process based on a vector tree structure after the input query statement, calculate the computing power requirement for the query statement to be executed in the database by the data volume overhead and the length of the query text, output the execution time of the query under standard computing power conditions, and dynamically adjust the allocation of cloud transmission or local computing power resources.

Images