Data retrieval and intelligent operating systems and their methods

Abstract

A data retrieval and intelligent operating system and method thereof, comprising: a statement analysis module for extracting features and performing semantic analysis on a query statement to decompose it into multiple subquery statements; a decision agent module for generating a classification decision for each subquery statement based on its semantic features, wherein each classification decision includes one of the multiple retrieval modes; a retrieval module including a multiple retrieval unit and a hybrid retrieval unit for retrieving each subquery statement according to the retrieval mode in the classification decision and generating multiple retrieval results; and a result review module for reviewing the subquery statements and the retrieval results of the retrieval unit and / or the hybrid retrieval unit, and outputting the retrieval results if the retrieval results are correct.

Description

Technical Field

[0001] This invention relates to a data retrieval technology, and more particularly to a data retrieval and intelligent operating system and its method. Prior Technology

[0002] Search techniques include Relational Database Management Systems (RDBMs), Retrieval-Augmented Generation (RAG), or search engines, each with its own advantages and disadvantages.

[0003] Taking a relational database management system as an example, although structured data retrieval is fast, highly consistent, and efficient in transaction management, it has the following drawbacks: it cannot handle unstructured data, such as documents, PDFs, chat logs, and video descriptions; it has poor semantic understanding, only allowing queries based on explicit conditions and unable to understand natural language or semantically ambiguous issues; it lacks scalability, resulting in high maintenance costs when dealing with large amounts of documents or multilingual content; and it has poor query flexibility, requiring users to clearly understand the database schema and field names, making it unfriendly to general users. While augmented generative retrieval technology can handle unstructured data, understand semantics, and query in natural language, its drawbacks include: if the retrieved content is not accurate enough, the model may generate incorrect answers; it does not support structured query logic, such as complex operations like precise statistics, joins, and sorting; it lacks data consistency and normalization, making it susceptible to outdated information and unable to update data in real time; it involves embedded computation and vector alignment, sometimes resulting in higher latency and query costs; and it depends on index quality, with vectorization quality and chunking methods affecting recall performance.

[0004] In view of this, the present invention addresses the shortcomings of the aforementioned prior art when used in a single application and meets future needs by proposing a data retrieval and intelligent operating system and its method, thereby improving the above-mentioned deficiencies. The specific architecture and implementation method will be described in detail below. Summary of the Invention

[0005] The main objective of this invention is to integrate various retrieval modes, break down query statements into multiple sub-queries using semantic features, automatically determine which retrieval mode is more appropriate for the task requirements of the sub-queries, and assign the retrieval tasks of the sub-queries to the corresponding retrieval units, thereby improving the decision-making and execution capabilities of the retrieval system and enhancing the accuracy of the retrieval.

[0006] To achieve the above objectives, the present invention provides a data retrieval and intelligent operating system, which is coupled to an input device, receives a query statement input by the input device, and executes the retrieval of the query statement through a processor. The data retrieval and intelligent operating system includes: a statement analysis module, which performs feature extraction and semantic analysis on the query statement to decompose the query statement into multiple sub-queries; and a decision agent module, connected to the statement analysis module, which utilizes an artificial intelligence agent (AI) to perform the retrieval. An agent determines the semantic features of subqueries and generates a classification decision for each subquery, with each classification decision containing one of multiple search patterns; a search module connects to the agent module to receive classification decisions, the search module includes multiple search units and a hybrid search unit, each search unit corresponding to a search pattern, and performs searches on each subquery using either the search unit or the hybrid search unit according to the search pattern in each classification decision, generating multiple search results, wherein the hybrid search unit performs searches using at least two search patterns; and a result review module connects to the search module, reviews the subqueries and the search results of the search units and / or the hybrid search units, and outputs the search results if the search results are correct.

[0007] According to an embodiment of the present invention, a preprocessing module is further included, connected between the input device and the statement analysis module. The preprocessing module receives the query statement input by the input device, converts the query statement into a structured query statement, and then sends it to the statement analysis module to decompose the structured query statement into subquery statements.

[0008] According to an embodiment of the present invention, the decision agent module includes: a prompt specification generation unit for generating prompt specifications; a comparison unit connected to the prompt specification generation unit, which compares the positive correlation between the subquery statement and the prompt specification through a comparison model and assigns a first relevance confidence value; a vector database for storing complex vector data; a segmentation unit connected to the vector database, which segments the query statement into complex segments as subquery statements, converts each segment into an embedding vector through an embedding model, performs similarity matching between the embedded vector and the vector data in the vector database, finds the complex vector with the highest vector numerical similarity, and assigns a second relevance confidence value based on the vector similarity; and a decision unit connected to the comparison unit and the segmentation unit, which stores a first threshold (one of the first relevance confidence values) and a second threshold (one of the second relevance confidence values), compares the first relevance confidence value with the first threshold, and compares the second relevance confidence value with the second threshold to generate a classification decision.

[0009] According to an embodiment of the present invention, the prompt specification generation unit extracts multiple types of metadata from a relational database management system and generates prompt specifications based on the metadata.

[0010] According to an embodiment of the present invention, the retrieval unit includes a relational database retrieval unit, an enhanced generative retrieval unit, and a web retrieval unit.

[0011] According to an embodiment of the present invention, when the first relevance confidence value is less than the first threshold and the second relevance confidence value is less than the second threshold, the decision unit generates a classification decision for retrieval by the network retrieval unit.

[0012] According to an embodiment of the present invention, when the first relevance confidence value is less than the first threshold and the second relevance confidence value is greater than the second threshold, the decision unit generates a classification decision to enhance the retrieval unit's retrieval.

[0013] According to an embodiment of the present invention, when the first relevance confidence value is greater than the first threshold and the second relevance confidence value is less than the second threshold, the decision unit generates a classification decision to perform a retrieval using the relational database retrieval unit.

[0014] According to an embodiment of the present invention, when the first relevance confidence value is greater than the first threshold and the second relevance confidence value is greater than the second threshold, the decision unit generates a classification decision to perform retrieval using a hybrid retrieval unit.

[0015] According to an embodiment of the present invention, the relational database retrieval unit includes an SQL query statement generator, which takes the prompt specification generated by the prompt specification generation unit as input to the SQL query statement generator to generate an SQL command, and performs a relational database retrieval based on the SQL command.

[0016] According to an embodiment of the present invention, the network retrieval unit includes: a URL selection unit, which performs semantic similarity comparison on a predefined plurality of credibility websites and their descriptions to select a plurality of highly relevant URLs from the credibility websites; a website query unit, which obtains a plurality of query results by searching for subquery statements through a search engine and extracts the first plurality of query results from the query results; and a processing unit, which connects the URL selection unit and the website query unit, inputs the highly relevant URLs and the first plurality of query results into a network retrieval model to perform retrieval and generate retrieval results.

[0017] According to an embodiment of the present invention, the hybrid retrieval unit connects the relational database retrieval unit, the enhanced generative retrieval unit, and the web retrieval unit, and simultaneously uses at least two of the relational database retrieval unit, the enhanced generative retrieval unit, and the web retrieval unit to perform retrieval.

[0018] This invention also provides a data retrieval and intelligent operation method, which is located in a server. The server is coupled to an input device, receives a query statement input by the input device, and executes the retrieval of the query statement through a processor. The data retrieval and intelligent operation method includes the following steps: using a statement analysis module to extract features and perform semantic analysis on the query statement to decompose the query statement into multiple sub-queries; using an AI agent in a decision agent module to judge the semantic features of the sub-queries to generate a classification decision for each sub-query statement, each classification decision including one of the multiple retrieval modes; using a retrieval module to retrieve each sub-query statement according to the retrieval mode in the classification decision and generate multiple retrieval results, wherein the retrieval module includes multiple retrieval units and a hybrid retrieval unit, each retrieval unit corresponding to one of the retrieval modes; and using a result review module to review the sub-queries and the retrieval results of the retrieval units and / or the hybrid retrieval units, and outputting the retrieval results when the retrieval results are correct. Simple Explanation of the Diagram

[0019] Figure 1 is an architecture diagram of the data retrieval and intelligent operating system of this invention; Figure 2 is a block diagram of the data retrieval and intelligent operating system of this invention; Figure 3 is a flowchart of the data retrieval and intelligent operation method of the present invention; Figure 4 is a block diagram of the data retrieval and intelligent operating system of this invention, which includes a detailed block diagram of the decision agent module; Figure 5 is a flowchart of the decision-making process of the decision-making unit in this invention; Figure 6 is a block diagram of the retrieval module in this invention. Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0022] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should also be further understood that the term "and / or" as used in this specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] This invention provides a data retrieval and intelligent operating system and its method. Utilizing large language model (LLM) technology, it can simultaneously process structured and unstructured documents using relational database management systems (RDBMS) and search-enhanced generation (RAG) techniques. Furthermore, by automatically determining task requirements through an AI agent, the system's decision-making and execution capabilities are improved, creating an intelligent and multifunctional data retrieval and operating system.

[0025] Please refer to Figure 1, which is an architecture diagram of the data retrieval and intelligent operating system 10 of this invention. The data retrieval and intelligent operating system 10 is housed in a server 20. The server 20 is coupled to an input device 30 to receive external operations. This input device 30 can be at least one of a mouse, keyboard, or handwriting tablet. The server 20 also includes a processor 40, which can be a central processing unit (CPU), digital signal processor (DSP), microprocessor (MPU), graphics processing unit (GPU), or microcontroller (MCU) to execute applications and perform computational processing. The processor 40 is connected to the data retrieval and intelligent operating system 10. When the server 20 receives a query statement input by the input device 30, the processor 40 executes the data retrieval and intelligent operating system 10 to perform a retrieval of the query statement.

[0026] Please refer to Figure 2, which is a block diagram of the data retrieval and intelligent operating system 10 of this invention. The data retrieval and intelligent operating system 10 includes a statement analysis module 12, a decision agent module 14, a retrieval module 16, and a result review module 18. The statement analysis module 12 is connected to the decision agent module 14, the decision agent module 14 is connected to the retrieval module 16, and the retrieval module 16 is connected to the result review module 18. The statement analysis module 12 performs feature extraction and semantic analysis on the query statement to break it down into multiple subquery statements. The decision agent module 14 includes an artificial intelligence agent (AI agent), which automatically judges the semantic features of the subquery statements to generate a classification decision for each subquery statement. Each classification decision includes one of multiple retrieval modes, but the classification decision for each subquery statement may include the same or different retrieval modes. The retrieval module 16 includes multiple retrieval units 162, 164, 166, etc., and a hybrid retrieval unit 168. Each retrieval unit 162, 164, 166 corresponds to one of the multiple retrieval modes. For example, retrieval unit 162 uses the first retrieval mode, retrieval unit 164 uses the second retrieval mode, retrieval unit 166 uses the third retrieval mode, and so on. The number of retrieval units corresponds to the types of retrieval modes used by the data retrieval and intelligent operating system 10. In one embodiment, the retrieval modes include Relational Database Management Systems (RDBMs), Retrieval-Augmented Generation (RAG) retrieval, web search retrieval, and a hybrid retrieval. The hybrid retrieval unit 168 combines two or more retrieval modes for retrieval. When the retrieval module 16 receives the classification decision generated by the decision agent module 14, it performs retrieval on each sub-query statement according to the retrieval mode in the classification decision and generates multiple retrieval results, with each retrieval unit corresponding to one of the retrieval modes. The result review module 18 reviews the search results of sub-query statements, search units, and / or mixed search units. If the search results are correct, the search results are output.

[0027] Figure 3 is a flowchart of the data retrieval and intelligent operation method of the present invention. Please refer to Figures 2 and 3 simultaneously. As shown in the figure, in step S10, the statement analysis module 12 first performs feature extraction and semantic analysis on the query statement to decompose the query statement into multiple subquery statements. In this step, the statement analysis module 12 uses a large language model to analyze and decompose the query statement. This large language model can be a multi-model large language model, such as GPT 4o, Google Gemini 2.5, etc. In step S12, the AI ​​agent of the decision agent module 14 judges the semantic features of the subquery statements, automatically determines the appropriate processing method (i.e., the appropriate retrieval mode) for each subquery statement, and generates a classification decision for each subquery statement accordingly. This classification decision includes the retrieval mode suitable for the subquery statement. In step S14, according to the retrieval mode in the classification decision, the multiple retrieval units and / or a hybrid retrieval unit in the retrieval module 16 respectively retrieve each subquery statement and generate multiple retrieval results. Furthermore, retrieval module 16 assigns the classification decision task to the retrieval unit corresponding to the retrieval mode. Assuming the classification decision of the first sub-query statement includes the first retrieval mode, retrieval module 16 assigns the first sub-query statement to retrieval unit 162 for retrieval. Finally, as described in step S16, result review module 18 reviews the sub-query statement and the retrieval results of retrieval units 162, 164, 166, and / or the mixed retrieval unit 168. If the retrieval result is correct, it is output. In step S16, result review module 18 reviews the input query statement and retrieval results through the Chain of Thought (CoT) capability of the large language model to determine whether the retrieval result is correct. This determination method is achieved by the Chain of Thought processing based on context, combined with the dialogue generated by the large language model during the automatic generation process, to help generate the judgment result.

[0028] Figure 4 is a block diagram of the data retrieval and intelligent operating system 10 of the present invention, which includes a detailed block diagram of the decision agent module 14. The data retrieval and intelligent operating system 10 further includes a preprocessing module 11, which is connected between the input device 30 and the statement analysis module 12. After receiving the query statement input by the input device 30, the preprocessing module 11 first converts the query statement into a structured query statement, and then transmits the structured query statement to the statement analysis module 12, allowing the statement analysis module 12 to more easily decompose the structured query statement into sub-queries. For example, the query statement may be input in text or voice form. The preprocessing module 11 can convert voice input into text format for subsequent processing, while text input can be processed directly. The preprocessing module 11 removes semantically ambiguous or invalid information from the query statement, and then uses Natural Language Processing (NLP) to analyze the syntax and semantics of the query statement. For example, suppose the query input by the input device 30 is "Query the list of employees and their details for a specific store (store 1)", the structured query after being processed by the preprocessing module 11 will become "Query store 1, list of employees".

[0029] The decision agent module 14 includes a prompt specification generation unit 142, a comparison unit 144, a vector database 146, a block unit 148, and a decision unit 149. The prompt specification generation unit 142 is connected to the comparison unit 144, the vector database 146 is connected to the block unit 148, and the comparison unit 144 and the block unit 148 are connected to the decision unit 149. The prompt specification generation unit 142 generates a prompt specification. It connects to the relational database management system (see the relational database retrieval unit 162 in Figure 6, which is installed within the relational database retrieval unit 162) via a connection string. It extracts various metadata, such as table / column descriptions, types, primary keys (pk), foreign keys (fk), and indexes, and generates the prompt specification. In relational database retrieval mode, this prompt specification serves as input to the SQL query generator to produce SQL commands, which will be described later. The comparison unit 144 compares the subquery statement with the prompt specification using a comparison model, determining the degree of positive correlation between the two and assigning a first correlation confidence value C1. This comparison model is a large language model, such as Google Gemini 2.5 or OpenAI GPT 4o. The vector database 146 stores complex vector data. The chunking unit 148 divides the query statement into complex chunks as subqueries, and converts each chunk into an embedding vector through an embedding model. The embedding vector is then matched with the vector data in the vector database 146 to find the complex vector with the highest numerical similarity. A second relevance confidence value C2 is then assigned based on the similarity of the vectors.

[0030] The decision unit 149 originally stores a first threshold T1, one of the first relevance confidence values, and a second threshold T2, one of the second relevance confidence values. The decision process of the decision unit 149 is shown in Figure 5. For example, in step S20, the first relevance confidence value C1 is compared with the first threshold T1, and the second relevance confidence value C2 is compared with the second threshold T2. Then, steps S22-S24 generate different classification decisions based on the two comparison results. The comparison results include four possibilities, such as the four comparison results in S22, each corresponding to a classification decision in S24. Referring also to Figure 6, if the first relevance confidence value C1 is less than the first threshold T1, and the second relevance confidence value C2 is less than the second threshold T2, the decision unit 149 generates a classification decision to perform a search using the network retrieval unit 166. If the first relevance confidence value C1 is less than the first threshold T1, and the second relevance confidence value C2 is greater than the second threshold T2, the decision unit 149 generates a classification decision to perform a search using the enhanced generation retrieval unit 164. If the first relevance confidence value C1 is greater than the first threshold T1, and the second relevance confidence value C2 is less than the second threshold T2, the decision unit 149 generates a classification decision to perform a search using the relational database retrieval unit 162. If the first relevance confidence value C1 is greater than the first threshold T1, and the second relevance confidence value C2 is greater than the second threshold T2, the decision unit 149 generates a classification decision to perform a search using the hybrid retrieval unit 168.

[0031] The aforementioned related database retrieval unit 162, enhanced generation retrieval unit 164, web retrieval unit 166, and hybrid retrieval unit 168 are all located in the retrieval module 16. The hybrid retrieval unit 168 connects the related database retrieval unit 162, the enhanced generation retrieval unit 164, and the web retrieval unit 166. If the classification decision selects to use the hybrid retrieval unit 168 for retrieval, at least two of the related database retrieval unit 162, the enhanced generation retrieval unit 164, and the web retrieval unit 166 will be used simultaneously for retrieval, and the results will then be returned to the hybrid retrieval unit 168.

[0032] The relational database retrieval unit 162 includes an SQL query statement generator 1622 and an SQL execution unit 1624. The SQL query statement generator 1622 takes the prompt specification generated by the prompt specification generation unit 142 as input to the SQL query statement generator 1622 to generate an SQL command, and the SQL execution unit 1624 performs relational database retrieval based on the SQL command.

[0033] Retrieval-Augmented Generation (RAG) is an AI technique that combines retrieval and generation, typically used to improve the accuracy of responses from Large Language Models (LLMs). The RAG unit 164 comprises three steps: retrieval, re-ranking, and confidence assessment, as well as retrieval generation. The retrieval step follows the same flow as module 144. In the second step, the complex vectors with the highest numerical similarity are re-ranked. A reranking model is used to recalculate the similarity of these segments and rank them according to relevance, placing the most relevant vectors at the top. The third step then calculates whether the most relevant segments are sufficiently reliable (compared to a set threshold). If the confidence value is higher than the threshold, these segments are provided as "contextual information" to the LLM to generate the final answer; however, if the confidence value is lower than the threshold, it may indicate that the query lacks sufficient supporting data, and an uncertain answer is returned to the user, or the user is asked to re-enter the query.

[0034] The web retrieval unit 166 includes a URL selection unit 1662, a website query unit 1664, and a processing unit 1666. The processing unit 1666 connects the URL selection unit 1662 and the website query unit 1664. The URL selection unit 1662 performs semantic similarity comparisons on a predefined plurality of trustworthy websites and their descriptions to select a plurality of highly relevant URLs from the trustworthy websites. The website query unit 1664 obtains multiple query results by searching a subquery statement through a search engine (e.g., Google Search API) and extracts the first plurality of query results from the query results. The processing unit 1666 receives the highly relevant URLs selected by the URL selection unit 1662 and the first plurality of query results obtained by the website query unit 1664, and inputs the highly relevant URLs and the first plurality of query results into a web retrieval model for retrieval. This web retrieval model is a large-scale language model, such as a generative language model.

[0035] For example, suppose the relational database management system already stores actor film performance records in MySQL format, and another database stores text files of actor personal information, such as PDF files, as well as online movie databases (IMDb) and Wikipedia. When a user enters the query "What is Tom Hanks' career background? What movies has he participated in? What new movies has he recently released?", firstly, the query analysis module 12 breaks it down into three sub-queries: "Provide the career background of 'Tom Hanks'", "Query the movies 'Tom Hanks' has participated in", and "Query the recent film works of 'Tom Hanks'". Next, the decision agent module 14 analyzes the appropriate retrieval method for each sub-query. For example, career background might come from internal documents or online resources, so a hybrid retrieval mode can be used, selecting enhanced generative retrieval and web retrieval. If internal documents are available, the enhanced generative retrieval unit 164 can retrieve them from the vector database 146; if no internal documents are available, the web retrieval unit 166 can search online. The query "Tom Hanks' film credits" can be retrieved using a relational database search mode, which provides a structured search for information on actors and their film credits. Recent film credits require consideration of the film's popularity, which is real-time information, necessitating online searches. Based on the classification decision of the decision agent module 14, retrieval module 16 assigns "Tom Hanks' career background" to the hybrid search unit 168, and selects the enhanced generation search unit 164 and the online search unit 166; "Tom Hanks' film credits" is assigned to the relational database search unit 162; and "Tom Hanks' recent film credits" is assigned to the online search unit 166. The result review module 18 uses a large language model's thought process to verify if the results match the user's query, ensuring the data source is correct, the query results are complete and consistent, and no additional corrections are needed. If any of these criteria are not met, the query must be retried. Finally, the large-scale language model converts the reviewed search results into human-readable natural language and outputs them. For example, the output might be: "Tom Hanks is an American actor who has won numerous awards, including an Academy Award. He has starred in many well-known films, including Apollo 13 and Saving Private Ryan. His recent works include Here."

[0036] However, the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention. Therefore, all equivalent variations or modifications made in accordance with the features and spirit described in the claims of this invention should be included within the scope of the patent application of this invention.

[0037] 10: Data Retrieval and Intelligent Operating Systems 11: Preprocessing Module 12: Statement Analysis Module 14: Decision Agent Module 142: Prompt Specification Generation Unit 144: Comparison Unit 146: Vector Database 148: Block Unit 149: Decision-making unit 16: Retrieval Module 162: Retrieval unit, related database retrieval unit 1622: SQL Query Generator 1624: SQL Execution Unit 164: Search unit, enhanced generative search unit 166: Search unit, online search unit 1662: URL Selection Unit 1664: Website Query Unit 1666: Processing Unit 168: Hybrid Retrieval Unit 18: Result Review Module 20: Server 30: Input device 40: Processor

Claims

1. A data retrieval and intelligent operating system, coupled to an input device, receiving a query statement input by the input device, and executing the retrieval of the query statement through a processor, the data retrieval and intelligent operating system comprising: A query analysis module extracts features and performs semantic analysis on the query to break it down into multiple subqueries. A decision agent module, connected to the query analysis module, uses an AI agent to determine the semantic features of the subqueries and generates a classification decision for each subquery. Each classification decision includes one of the multiple retrieval patterns. The decision agent module includes: a prompt specification generation unit to generate a prompt specification; a comparison unit, connected to the prompt specification generation unit, which compares the positive correlation between the subquery and the prompt specification using a comparison model and assigns a first relevance confidence value; a vector database storing multiple vector data; and a chunking unit, connected to the vector database, which divides the query into multiple chunks as subqueries and uses an embedding model. The model converts each segment into an embedding vector and performs similarity matching between the embedding vectors and the vector data in the vector database to find the complex vectors with the highest vector numerical similarity, and assigns a second relevance confidence value based on the similarity of the vectors; and a decision unit connects the comparison unit and the segmentation unit, the decision unit stores a first threshold of the first relevance confidence value and a second threshold of the second relevance confidence value, compares the first relevance confidence value with the first threshold, and compares the second relevance confidence value with the second threshold to generate the classification decision; A retrieval module, connected to the decision agent module to receive the classification decisions, includes multiple retrieval units and a hybrid retrieval unit. Each retrieval unit corresponds to a retrieval mode. Based on the retrieval mode in each classification decision, the retrieval unit or the hybrid retrieval unit is used to retrieve each sub-query statement and generate multiple retrieval results. The hybrid retrieval unit uses at least two retrieval modes for retrieval. A result review module, connected to the retrieval module, reviews the sub-query statements and the retrieval results of the retrieval units and / or the hybrid retrieval unit. If the retrieval results are correct, the retrieval results are output.

2. The data retrieval and intelligent operating system as described in claim 1 further includes a preprocessing module connected between the input device and the statement analysis module. The preprocessing module receives the query statement input by the input device, converts the query statement into a structured query statement, and then transmits it to the statement analysis module to decompose the structured query statement into the sub-query statements.

3. The data retrieval and intelligent operating system as described in claim 1, wherein the prompt specification generation unit extracts multiple types of metadata from a relational database management system and generates the prompt specification based on the metadata.

4. The data retrieval and intelligent operating system as described in claim 1, wherein the retrieval units include a relational database retrieval unit, an enhanced generative retrieval unit, and a web retrieval unit, and the hybrid retrieval unit is connected to the relational database retrieval unit, the enhanced generative retrieval unit, and the web retrieval unit, and simultaneously uses at least two of the relational database retrieval unit, the enhanced generative retrieval unit, and the web retrieval unit to perform retrieval.

5. The data retrieval and intelligent operating system as described in claim 4, wherein when the first relevance confidence value is less than the first threshold and the second relevance confidence value is less than the second threshold, the decision unit generates the classification decision to perform the retrieval by the network retrieval unit.

6. The data retrieval and intelligent operating system as described in claim 4, wherein when the first relevance confidence value is less than the first threshold and the second relevance confidence value is greater than the second threshold, the decision unit generates the classification decision to perform the retrieval using the enhanced generation retrieval unit.

7. The data retrieval and intelligent operating system as described in claim 4, wherein when the first relevance confidence value is greater than the first threshold and the second relevance confidence value is less than the second threshold, the decision unit generates the classification decision to perform the retrieval using the relational database retrieval unit.

8. The data retrieval and intelligent operating system as described in claim 1, wherein when the first relevance confidence value is greater than the first threshold and the second relevance confidence value is greater than the second threshold, the decision unit generates the classification decision to perform the retrieval using the hybrid retrieval unit.

9. The data retrieval and intelligent operating system as described in claim 4, wherein the relational database retrieval unit includes an SQL query generator, which takes the prompt specification generated by the prompt specification generation unit as input to the SQL query generator to generate an SQL command, and performs a relational database retrieval based on the SQL command.

10. The data retrieval and intelligent operating system as described in claim 4, wherein the network retrieval unit includes: A URL selection unit performs semantic similarity comparison on a predefined plurality of trustworthy websites and their descriptions to select a plurality of highly relevant URLs from the trustworthy websites; a website query unit obtains a plurality of query results by searching the subqueries through a search engine, and extracts the first plurality of query results from the query results; and a processing unit connects the URL selection unit and the website query unit, inputs the highly relevant URLs and the first plurality of query results into a web retrieval model to perform a retrieval and generate the retrieval results.

11. A data retrieval and intelligent operation method, located in a server, the server being coupled to an input device, receiving a query statement input by the input device, and executing the retrieval of the query statement through a processor, the data retrieval and intelligent operation method comprising the following steps: using a statement analysis module to extract features and perform semantic analysis on the query statement, so as to decompose the query statement into multiple subquery statements; using an artificial intelligence agent (AI agent) of a decision agent module to determine the semantic features of the subquery statements, so as to generate a classification decision for each of the subquery statements, each classification decision including one of the multiple retrieval modes; according to the retrieval mode in the classification decisions, using a multiple retrieval unit and / or a hybrid retrieval unit in a retrieval module to retrieve each of the subquery statements and generate multiple retrieval results, wherein... Each of these retrieval units corresponds to a retrieval mode, and the hybrid retrieval unit performs retrieval using at least two of these retrieval modes; and a result review module reviews the subquery statements and the retrieval results of the retrieval units and / or the hybrid retrieval unit, outputting the retrieval results if they are correct; wherein, the step of the decision agent module generating the classification decisions further includes: comparing the subquery statement with a pre-generated hint specification using a comparison model, and assigning a first relevance confidence value; dividing the query statement into multiple blocks using a chunking unit as the subquery statements, and using an embedding model (Embedding... The Model converts each segment into an embedding vector and performs similarity matching between the embedding vectors and complex vector data stored in a vector database to find the complex vectors with the highest vector numerical similarity. A second relevance confidence value is assigned based on the similarity of the vectors. A decision unit compares the first relevance confidence value with a first threshold and compares the second relevance confidence value with a second threshold to generate the classification decision.

12. The data retrieval and intelligent operation method as described in claim 11 further includes receiving the query statement input by the input device using a preprocessing module, converting the query statement into a structured query statement, and then transmitting it to the statement analysis module to decompose the structured query statement into the sub-query statements.

13. The data retrieval and intelligent operation method as described in claim 11, wherein the retrieval units include a relational database retrieval unit, an enhanced generation retrieval unit, and a web retrieval unit, and the hybrid retrieval unit performs retrieval using at least two of the relational database retrieval unit, the enhanced generation retrieval unit, and the web retrieval unit simultaneously.

14. The data retrieval and intelligent operation method as described in claim 13, wherein when the first relevance confidence value is less than the first threshold and the second relevance confidence value is less than the second threshold, the decision unit generates the classification decision to perform retrieval by the network retrieval unit.

15. The data retrieval and intelligent operation method as described in claim 13, wherein when the first relevance confidence value is less than the first threshold and the second relevance confidence value is greater than the second threshold, the decision unit generates the classification decision to perform the retrieval using the enhanced generation retrieval unit.

16. The data retrieval and intelligent operation method as described in claim 13, wherein when the first relevance confidence value is greater than the first threshold and the second relevance confidence value is less than the second threshold, the decision unit generates the classification decision to perform the retrieval using the relational database retrieval unit.

17. The data retrieval and intelligent operation method as described in claim 11, wherein when the first relevance confidence value is greater than the first threshold and the second relevance confidence value is greater than the second threshold, the decision unit generates the classification decision to perform retrieval using the hybrid retrieval unit.

18. The data retrieval and intelligent operation method as described in claim 13, wherein the retrieval system using the web retrieval unit includes the following steps: predefined multiple credibility websites and descriptions of such credibility websites; performing semantic similarity comparison between such subqueries and such credibility websites and their descriptions using a web retrieval model to select multiple highly relevant URLs from such credibility websites; obtaining multiple query results by searching such subqueries using a search engine; taking the first multiple entries of such query results as multiple website contents; then using the web retrieval model and a web crawler to retrieve such website contents from such subqueries and generate a website retrieval result; and performing a retrieval using such highly relevant URLs and the website retrieval result as keywords and generating such retrieval results.

Images