Structured retrieval enhancement-based national defense mobilization knowledge base question and answer method and system

Abstract

The invention discloses a national defense mobilization knowledge base question answering method and system based on structured retrieval enhancement, and relates to the technical field of artificial intelligence, natural language processing and national defense informatization. Performing dynamic mode induction on the multi-source heterogeneous mobilization data by utilizing a large language model, and automatically extracting entity, attribute and relationship triples; constructing a hybrid index system comprising structured storage and unstructured storage; performing conflict resolution and confidence evaluation on the multi-source data through a weighted scoring algorithm based on timeliness, authority and consistency; and performing hybrid routing retrieval according to the query intention of the user to generate an answer with a traceability link. According to the method, the accuracy, the traceability and the logical reasoning ability of complex decision query are remarkably improved, and the robustness and the adaptability of mobilization decision are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, natural language processing and national defense informatization, more particularly to a national defense mobilization knowledge base question answering method and system based on structured retrieval enhancement. BACKGROUND

[0002] In modern national defense mobilization and national emergency response systems, the efficiency and accuracy of decision-making directly depend on the degree of mastery of massive logistics data. National defense mobilization data has high heterogeneity, dispersion and dynamics, and its complexity far exceeds that of general commercial application scenarios. Mobilization supply chain data is often scattered in government contracts, financial data and private databases of various suppliers, forming serious data islands.

[0003] Specifically, the data challenges in the field of mobilization mainly manifest in the following dimensions: First, the extreme diversity of data source formats. Critical mobilization information is not always stored in neat database tables, but is scattered in unstructured documents. For example, troubleshooting guides for equipment are usually in the form of technical manuals in PDF format; emergency action plans and their functional annexes are often text-intensive documents that describe complex process logic for command and control, communication and warning; and the performance of suppliers may only exist in non-standard press releases or email records.

[0004] Second, the semantic barriers of professional coding systems. Mobilization potential has unique material classification and coding standards. A typical potential material number (SN) contains rich information. General large language models (LLM) often treat these numbers as meaningless strings, making it difficult to understand the hierarchical relationships and interchangeability rules behind them, resulting in serious logical errors in material allocation recommendations.

[0005] Third, the "long whip effect" and data blind spots of the supply chain. In traditional hierarchical mobilization structures, managers can only see the data of their immediate subordinates, lacking visibility into secondary suppliers. This information asymmetry leads to the "long whip effect," where demand fluctuations are amplified at each level of the supply chain, causing excessive inventory or severe shortages. Existing RAG techniques struggle to handle complex aggregation queries, data conflicts and traceability requirements in the field of mobilization; Text-to-SQL techniques are limited by the structured cold start problem.

[0006] Therefore, how to propose an accurate, traceable and logically reasoning question answering method for complex aggregation queries is a problem that needs to be solved by those skilled in the art. SUMMARY

[0007] In view of the above problems, the present application is proposed in order to provide a defense mobilization knowledge base question answering method and system based on structured retrieval enhancement to overcome the above problems or at least partially solve the above problems.

[0008] In order to achieve the above-mentioned purpose, the present application adopts the following steps: S1, acquiring multi-source heterogeneous mobilization field data, the mobilization field data being unstructured documents, performing text block processing on the unstructured documents to obtain original text segments, performing dense vectorization processing on the original text segments to generate dense vector data, and performing sparse indexing on keywords in the original text segments; S2, performing semantic analysis and coding analysis on the unstructured documents by using a large language model, extracting a triple data set containing entities, attributes and relationships in the unstructured documents, and dynamically generating a structured record adapted to the triple data set; S3, constructing a structured data storage system and an unstructured data storage system, storing the structured record to the structured data storage system, calculating a confidence degree of the structured record based on a preset mathematical model and marking a conflict state, updating data in the structured storage system based on the data after the confidence degree calculation and the data after the conflict state marking, and storing the original document, the dense vector and the keywords of the sparse index of the original text segment into the unstructured data storage system; S4, receiving natural language query information input by a user, analyzing semantic features of the natural language query information through the structured data storage system and the unstructured storage system, outputting an intent type of the semantic features, performing query routing processing by using a large language model according to the intent type, receiving query results fed back by the structured data storage system and the unstructured storage system, calculating a confidence score of the query results based on a preset mathematical model, and retaining high-confidence data; S5, based on the high-confidence data, outputting a final answer through large language model enhancement, and displaying a traceability link of a data source on the final answer.

[0009] Preferably, the dense vectorization processing in S1 is a technology of mapping to a continuous dense vector space; the sparse indexing is a technology based on keyword matching, which is used to enhance the indexing accuracy of proper nouns or error codes.

[0010] Preferably, the coding analysis step in S2 is: identifying a material code in the unstructured document, disassembling the material code into hierarchical metadata fields through built-in analysis logic and a mapping table, and storing the hierarchical metadata fields into the structured storage system.

[0011] Preferably, in S4, the intent type is used for query routing processing using a large language model, which includes: if the large language model query is aggregate type, then an SQL statement is generated and executed in the structured storage system; if the large language model query is semantic type, then dense vector retrieval and sparse retrieval are executed in the unstructured storage system; if the large language model query is composite type, then the SQL query, dense vector retrieval and sparse retrieval are executed in parallel, and the results are fused.

[0012] Preferably, S4, which calculates the confidence score of the query results based on a preset mathematical model, also includes confidence management and conflict resolution operations: The preset mathematical model is a weighted scoring algorithm, which calculates a confidence score for unstructured documents. Based on the confidence score, it performs structured record update and filtering operations, outputs the updated structured records, and stores the updated structured records in the structured data storage system. Before the structured records are written to the structured storage system, or after semantic features are retrieved from structured and unstructured storage, inconsistent data with the same entity attribute is identified. A weighted scoring model is constructed based on the entity's timeliness, authority, and consistency. Data updates or conflict marking are performed according to the confidence score to complete the conflict resolution operation.

[0013] Preferably, the confidence score is calculated. The mathematical model is as follows: = . + . + . ; in, To regulate The hyperparameters of the weights, To regulate The hyperparameters of the weights, To regulate The hyperparameters of the weights satisfy... + + =1, For time-sensitive functions, For authoritative functions, It is a consistency function.

[0014] Preferably, S5 specifically includes: High-confidence data is input into the large language model as contextual information. The final answer is generated based on the contextual information. The large language model is required to add source identification to the contextual information. The source identification is a source link that displays the source of the data.

[0015] The application also provides a defense mobilization knowledge base question answering system based on structured retrieval enhancement, comprising: An original text processing module: acquires mobilization field data of multiple sources and heterogeneous, the mobilization field data being unstructured documents, performs text block processing on the unstructured documents to obtain original text segments, performs dense vectorization processing on the original text segments to generate dense vector data, performs sparse indexing on keywords in the original text segments, performs semantic analysis and coding analysis on the unstructured documents by using a large language model, extracts a triple data set containing entities, attributes and relationships in the unstructured documents, and dynamically generates a structured record adapted to the triple data set; A hybrid index storage system module: constructs a structured data storage system and an unstructured data storage system, stores the structured record to the structured data storage system, performs confidence calculation on the structured record based on a preset mathematical model and marks a conflict state, updates data in the structured storage system based on the data after confidence calculation and the data after marking the conflict state, and stores the original document, the dense vector and the sparse indexed keywords of the original text segment to the unstructured data storage system; An intent routing module: receives natural language query information input by a user, analyzes semantic features of the natural language query information through the structured data storage system and the unstructured storage system, outputs an intent type of the semantic features, performs query routing processing by using a large language model according to the intent type, receives query results fed back by the structured data storage system and the unstructured storage system, calculates a confidence score of the query results based on a preset mathematical model, and retains high-confidence data; An answer generation and traceability module: outputs a final answer by using a large language model enhancement based on the high-confidence data, and displays a traceability link of a data source on the final answer.

[0016] According to the technical solution, compared with the prior art, the application provides a defense mobilization knowledge base question answering method and system based on structured retrieval enhancement, and the above technical solution provided by the application has at least the following beneficial effects: 1. Greatly improves the accuracy of complex queries: by converting unstructured text into structured data that can be queried by SQL, the application significantly improves the accuracy of traditional baseline models when processing mobilization queries involving counting, summation, filtering and sorting, and solves the inherent defects of large models in mathematical logic reasoning.

[0017] 2. Breaks data silos and realizes global asset visualization: can uniformly process PDF, Excel, HTML and other formats of documents, and map scattered material data to a unified potential standard classification system, making it possible to schedule materials across departments and systems.

[0018] 3. Enhance the robustness and security of decision-making: through strict conflict resolution and traceability mechanism, effectively filter out-of-date and false information, prevent command failure caused by data illusion, meet the strict requirements of information accuracy in the field of national defense.

[0019] 4. Adapt to the rapidly changing needs of wartime: without manual reconstruction of the database, the system can automatically adjust the structured structure according to new input intelligence documents, quickly adapt to new equipment types or logistics needs, and have strong battlefield adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only a part of the embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0021] Figure 1 A user query and data processing and storage process of a national defense mobilization knowledge base question answering method based on structured retrieval enhancement provided in an embodiment of the present application; Figure 2 A module closed-loop collaborative processing sequence provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0023] As shown in Figure 1 The embodiment of the present application discloses a national defense mobilization knowledge base question answering method based on structured retrieval enhancement, which specifically comprises the following steps: S1, acquire multi-source heterogeneous mobilization field data, the mobilization field data is an unstructured document, perform text block processing on the unstructured document to obtain original text segments, perform dense vectorization processing on the original text segments to generate dense vector data, and perform sparse indexing on keywords in the original text segments; S2, use a large language model to perform semantic analysis and coding analysis on the unstructured document, extract a triple data set containing entities, attributes and relationships in the unstructured document, and dynamically generate a structured record adapted to the triple data set; S3, construct a structured data storage system and an unstructured data storage system, store structured records to the structured data storage system, calculate the confidence of the structured records based on a preset mathematical model, and mark the conflict state, update the data in the structured storage system based on the data after the confidence calculation and the data after the conflict state marking; the original document, the dense vector and the sparse index keyword of the original text segment are stored in the unstructured data storage system; S4, receive the natural language query information input by the user, analyze the semantic features of the natural language query information through the structured data storage system and the unstructured storage system, output the intent type of the semantic features, perform query routing processing using a large language model according to the intent type, receive the query results fed back by the structured data storage system and the unstructured storage system, calculate the confidence score of the query results based on a preset mathematical model, and retain high-confidence data; S5, based on the high-confidence data, output the final answer through large language model enhancement, and display the traceability link of the data source on the final answer.

[0024] Further, the dense vectorization processing is a technology of mapping to a continuous dense vector space; the sparse index is a technology based on keyword matching, used to enhance the index accuracy of proper nouns or error codes.

[0025] Further, the encoding analysis step is: identifying the material code in the unstructured document, decomposing the material code into hierarchical metadata fields through built-in analysis logic and mapping table, and storing the hierarchical metadata fields in the structured storage system.

[0026] Further, if the large language model query is aggregation type, generate a SQL statement to execute in the structured storage system; if the large language model query is semantic type, perform dense vector retrieval and sparse retrieval in the unstructured storage system; if the large language model query is composite type, perform SQL query, dense vector retrieval and sparse retrieval in parallel, and fuse the results.

[0027] Further, the high-confidence data is input to the large language model as context information, the final answer is generated based on the context information, and the large language model is required to attach a traceability identifier to the context information, the traceability identifier is a traceability link that displays the data source.

[0028] For the technical solution of the application, the application provides a specific embodiment for constructing a national defense mobilization knowledge base question answering system based on structured retrieval enhancement, which includes the following four core modules: Module 1: Dynamic structured induction based on large model and heterogeneous data structured engine This module aims to solve the core contradiction between "data unstructured" and "query needs structured". Unlike traditional ETL (Extract-Transform-Load) processes that require manual rule design, the invention uses the powerful semantic understanding capabilities of large language models to automatically complete the mapping from unstructured text to relational databases.

[0029] Entity, relationship, and attribute triplets adaptive extraction: The system first processes the input PDF, HTML, or text stream in blocks. Using the fine-tuned LLM, it identifies mobilization key entities (such as suppliers, materials, warehouses, and transportation lines) and their attributes (such as inventory, code, latitude and longitude, and credit rating).

[0030] Professional code parsing enhancement: Due to the special nature of mobilization, the engine has a built-in code parser. When encountering material codes, the system not only stores them as strings but also automatically disassembles them into independent fields and stores them in structured indexes, thereby supporting hierarchical queries based on material categories.

[0031] Module Two: Hybrid Storage System To balance precise queries and fuzzy searches, the invention designs a dual-track parallel storage architecture: Structured storage: Stores SQL-compatible data generated by the structured induction engine. These data include cleaned numerical, date, enumeration, and entity relationships, specifically for responding to aggregate queries.

[0032] Unstructured storage: Stores text blocks and their high-dimensional vector embeddings of the original documents. Utilizing dense vector retrieval techniques to capture semantic information, it is used to respond to descriptive and open-ended queries.

[0033] Sparse retrieval enhancement: Introducing BM25 or SPLADE sparse retrieval algorithms in unstructured storage, specifically for handling specific military terminology, error codes, or proper nouns, to compensate for the lack of precise keyword matching in dense vectors.

[0034] Module Three: Conflict Resolution and Confidence Scoring Mechanism For multi-source intelligence conflict problems, the invention proposes a dynamic resolution algorithm based on evidence roundtable: Conflict detection: During data ingestion or retrieval, the system automatically identifies inconsistent descriptions for the same entity attribute.

[0035] Weighted confidence scoring: Construct a mathematical model to calculate the confidence score of unstructured documents, taking into account data timeliness, source authority, and cross-validation consistency.

[0036] Verifiable determination of conflict resolution: Distinguish between resolvable conflicts and those requiring human intervention. For major conflicts that cannot be resolved automatically, the system will explicitly mark a "data conflict warning" when generating a response and list opposing evidence for the commander's assessment.

[0037] Furthermore, when the system retrieves multiple values ​​for the same entity attribute, conflict detection is triggered.

[0038] ; in, and Different values ​​for the same entity attribute and For a preset threshold, when v is a numerical value Refers to a specific numerical value; when v is text type. It is a high-dimensional dense vector of text description after being transformed by the embedding model. The function is the formula for calculating semantic similarity.

[0039] Furthermore, the mathematical model employs a weighted scoring algorithm system to calculate the confidence score for each piece of evidence. : = . + . + . ; in, To regulate The hyperparameters of the weights, To regulate The hyperparameters of the weights, To regulate The hyperparameters of the weights satisfy... + + =1, For time-sensitive functions, For authoritative functions, It is a consistency function.

[0040] Timeliness function Used to measure the degree to which data diminishes in value over time. In mobilization and logistics, the value of intelligence is often highly time-sensitive.

[0041] An exponential decay model is adopted.

[0042] = .

[0043] in, The current time t is the timestamp where the data was generated. The attenuation coefficient is a hyperparameter that is dynamically adjusted based on the business scenario.

[0044] Authority function Discrete scoring is performed based on the hierarchical credibility of the data source. This is a form of prior knowledge, and a piecewise mapping function reflects the system's pre-defined trust levels for different intelligence sources.

[0045] ( )= ; This function embodies the command chain logic in national defense mobilization. Data from official databases (Tier 1) is inherently superior to reports from third-party logistics providers (Tier 2), and even more so to open-source intelligence from the internet (Tier 3). In conflict resolution, for Tier 3 data to outperform Tier 1 data, it must achieve extremely high scores in timeliness or consistency to compensate for its lack of authority.

[0046] Consistency function This measure assesses whether a piece of evidence has "popular support." It leverages the fact that truth often resides in the majority or in clusters of mutually corroborating evidence.

[0047] Mean calculation based on semantic similarity ( , ) = ; in, This indicates the evidence currently being evaluated. This refers to a collection of evidence from other independent sources on the same topic. This represents the formula for calculating semantic similarity.

[0048] Furthermore, the conflict resolution strategy is discussed with a specific example: if the highest score If the score is significantly higher than the second-highest score, the highest score will be accepted directly, and the value will be updated in the generated SQL table. If the scores are similar, the system will mark the field as UNCERTAIN and output the following in the response: "There is a conflict regarding inventory data. The logistics department report shows 1000, while the frontline sensors show 0. Immediate verification is recommended."

[0049] Module 4: Intelligent Intent Routing and Structured Enhancement Generation The system is equipped with an intent recognition router based on a large language model, which serves as the central hub of the entire question-and-answer process. Precise routing: If the user query contains statistical or comparative logic, it is routed to the Text-to-SQL module to execute the database query; if the query involves an operation process, it is routed to the vector retrieval module.

[0050] Hybrid execution: For compound queries, the system decomposes them into SQL subqueries and vector subqueries, executes them in parallel, and then merges the results.

[0051] Source tracing generation: The final generated answer is required to include source tracing links, pointing to specific database records or original document fragments, ensuring that every sentence in the answer can be traced back to its source.

[0052] Furthermore, the routing module analyzes the user's natural language input, extracts feature vectors, and determines which of the following intervals it falls into, as shown in Table 1: Table 1. Query Intent Classification

[0053] like Figure 2 As shown, this invention, through another specific embodiment, performs closed-loop collaboration of the above four modules in a "sensing-storage-verification-execution" process to jointly achieve high-precision question answering: Data transformation flow from Module 1 to Module 2: Module 1 acts as the system's "perception layer," responsible for parsing the chaotic raw document stream into ordered information. The structured records it generates are directly injected into the "structured storage" of Module 2 to support SQL queries, while the raw text blocks and semantic vectors are injected into "unstructured storage." This collaboration solves the pain points of traditional databases lacking data sources and traditional vector libraries lacking logical structure.

[0054] Data cleaning: Before the data is written to module two, module three pre-scans for conflicts between the old and new data and uses a time-space game algorithm to update the trust_score field of the data.

[0055] Search Filtering: When Module 4 performs the search, Module 3 intervenes in real time to perform consistency checks on the retrieved candidate evidence. If multiple retrieved documents are found to be logically mutually exclusive (such as two retrieved procedures that contradict each other), Module 3 will remove low-confidence evidence to ensure that the context fed to the generative model is logically consistent.

[0056] The query execution flow from Module 4 to Module 2 to the user: After Module 4 parses the user's intent, it dynamically schedules the resources of Module 2. For requests requiring precise statistics, it generates SQL commands to invoke structured indexes; for fuzzy exploration requests, it invokes vector indexes. Through this hybrid routing mechanism, the system achieves a dynamic balance between "accuracy" and "flexibility," ultimately returning a cleaned, traceable, high-quality answer to the user.

[0057] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0058] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A question-answering method for a national defense mobilization knowledge base based on structured retrieval enhancement, characterized in that, Includes the following steps: S1. Obtain multi-source heterogeneous mobilization domain data, wherein the mobilization domain data is an unstructured document, perform text block processing on the unstructured document to obtain original text fragments, perform dense vectorization processing on the original text fragments to generate dense vector data, and perform sparse indexing on the keywords in the original text fragments. S2. Use a large language model to perform semantic analysis and encoding parsing on the unstructured document, extract the triple dataset containing entities, attributes and relations in the unstructured document, and dynamically generate structured records that are adapted to the triple dataset; S3. Construct a structured data storage system and an unstructured data storage system. The structured records are stored in the structured data storage system. Calculate the confidence level of the structured records based on a preset mathematical model and mark conflict states. Update the data in the structured storage system based on the data after the confidence level calculation and the data after marking conflict states. The original document of the original text fragment, the dense vector, and the keywords of the sparse index are stored together in the unstructured data storage system; S4. Receive natural language query information input by the user, analyze the semantic features of the natural language query information through the structured data storage system and the unstructured storage system, output the intent type of the semantic features, perform query routing processing using a large language model according to the intent type, receive the query results fed back by the structured data storage system and the unstructured storage system, calculate the confidence score of the query results based on a preset mathematical model, and retain high confidence data. S5. Based on the high-confidence data, the final answer is output through large language model enhancement, and the source link of the data source is displayed on the final answer.

2. The question-answering method for a national defense mobilization knowledge base based on structured retrieval enhancement as described in claim 1, characterized in that, The dense vectorization process described in S1 is a technique for mapping to a continuous dense vector space; the sparse index is a keyword matching technique used to enhance the indexing accuracy of proper nouns or error codes.

3. The question-answering method for a national defense mobilization knowledge base based on structured retrieval enhancement as described in claim 1, characterized in that, The encoding and parsing steps in S2 are as follows: The material codes in the unstructured document are identified, and the material codes are decomposed into hierarchical metadata fields through built-in parsing logic and mapping tables, and the hierarchical metadata fields are stored in the structured storage system.

4. The question-answering method for a national defense mobilization knowledge base based on structured retrieval enhancement as described in claim 1, characterized in that, The intent type described in S4, which utilizes a large language model for query routing, includes: if the large language model query is aggregated, generating an SQL statement and executing it in the structured storage system; if the large language model query is semantic, performing dense vector retrieval and sparse retrieval in the unstructured storage system; and if the large language model query is composite, executing the SQL query, dense vector retrieval, and sparse retrieval in parallel, and then fusing the results.

5. The question-answering method for a national defense mobilization knowledge base based on structured retrieval enhancement according to claim 1, characterized in that, S4 calculates the confidence score of the query result based on a preset mathematical model, and also includes confidence management and conflict resolution operations: The preset mathematical model is a weighted scoring algorithm, which calculates a confidence score for the unstructured document. Based on the confidence score, it performs structured record update and filtering operations, outputs the updated structured record, and stores the updated structured record in the structured data storage system. Before the structured record is written to the structured storage system, or after retrieving semantic features from the structured storage and the unstructured storage, it identifies inconsistent data with the same entity attribute, constructs a weighted scoring model based on the entity's timeliness, authority, and consistency, and performs data updates or conflict marking according to the confidence score to complete the conflict resolution operation.

6. The question-answering method for a national defense mobilization knowledge base based on structured retrieval enhancement according to claim 5, characterized in that, The calculation of confidence score The mathematical model is as follows: = . + . + . ； in, To regulate The hyperparameters of the weights, To regulate The hyperparameters of the weights, To regulate The hyperparameters of the weights satisfy... + + =1, For time-sensitive functions, For authoritative functions, It is a consistency function.

7. The question-answering method for a national defense mobilization knowledge base based on structured retrieval enhancement as described in claim 1, characterized in that, S5 specifically includes: The high-confidence data is input into the large language model as context information. The final answer is generated based on the context information. The large language model is required to add a source identifier to the context information. The source identifier is a source link that displays the source of the data.

8. A question-and-answer system for a national defense mobilization knowledge base based on structured retrieval enhancement, characterized in that, A question-answering method for a national defense mobilization knowledge base based on structured retrieval enhancement according to any one of claims 1-7, comprising: The original text processing module acquires multi-source heterogeneous mobilization domain data, which is unstructured document data. It performs text block processing on the unstructured document to obtain original text fragments, performs dense vectorization processing on the original text fragments to generate dense vector data, performs sparse indexing on the keywords in the original text fragments, uses a large language model to perform semantic analysis and encoding parsing on the unstructured document, extracts the triple dataset containing entities, attributes and relations in the unstructured document, and dynamically generates structured records adapted to the triple dataset. Hybrid Index Storage System Module: Constructs a structured data storage system and an unstructured data storage system. The structured records are stored in the structured data storage system. Confidence is calculated on the structured records based on a preset mathematical model, and conflict states are marked. The data in the structured storage system is updated based on the data after confidence calculation and the data after marking conflict states. The original document of the original text fragment, the dense vector, and the keywords of the sparse index are jointly stored in the unstructured data storage system. The intent routing module receives natural language query information input by the user, analyzes the semantic features of the natural language query information through the structured data storage system and the unstructured storage system, outputs the intent type of the semantic features, performs query routing processing using a large language model based on the intent type, receives the query results fed back by the structured data storage system and the unstructured storage system, calculates the confidence score of the query results based on a preset mathematical model, and retains high-confidence data. The answer generation and tracing module, based on the high-confidence data, enhances the output through a large language model, outputs the final answer, and displays a tracing link indicating the source of the data on the final answer.

Images