Structured information retrieval method based on large language model

Abstract

The invention relates to the technical field of language information processing, and provides a large language model-based structured information retrieval method, which comprises the following steps of: deploying an adapter, accessing operation and maintenance data, carrying out timestamp synchronization on an original entry and identifying a source, mapping equipment, parts and personnel based on an asset directory and recording mapping confidence, extracting structural elements from the records; injecting three types of metadata into candidate evidences, establishing corresponding nodes and edges in a graph database, analyzing query into a structured retrieval intention, expanding a candidate evidence set along three chains to form candidate sub-graphs, and calculating comprehensive scores for sorting; the candidate evidences should be verified and constrained, meanwhile, LLM is called for inference, feasibility inference confidence is output, actual measurement results, the inference confidence and source confidence are fused according to preset weights, feasibility scores of the candidate evidences are obtained, and for the candidate evidences passing feasibility verification, accurate anchor points are calibrated for each evidence along a source chain; and generating a structured answer.

Description

Technical Field

[0001] This invention relates to the field of language information processing technology, specifically a structured information retrieval method based on a large language model. Background Technology

[0002] Information retrieval is the process and activity of finding and extracting the information needed by users from a massive collection of information through a series of methods and technologies. It involves the organization and storage of information, as well as the efficient location of content that is relevant to user needs.

[0003] The Chinese patent application CN202411411905.8 discloses a structured information retrieval method and system based on a large language model. This method includes: in speech recognition information retrieval based on a large language model, to achieve personalized recognition model iteration effects and a high-accuracy language understanding goal, obtaining correctly matched keyword groups through retrieval feedback based on a conventional big data language model, and obtaining individual pronunciation features based on the deviation between the user's speech of these word groups and the standard retrieval features provided by the large language model. These features are then used for personalized correction of the big data language model, achieving a high-accuracy language retrieval scheme based on the individual.

[0004] In the field of language information processing technology, although there are methods to obtain correctly matched keyword groups based on retrieval feedback from conventional big data language models, and to obtain individual pronunciation features based on the deviation between the user's speech and the standard retrieval features provided by the big data language model, thereby achieving a high-recognition-rate language retrieval solution based on individuals, existing technologies have the following problems in the operation and maintenance and troubleshooting scenarios of process industry sites:

[0005] 1. Current technologies in process industry site operation and maintenance and troubleshooting scenarios still rely on work order systems and keyword searches, lacking unified alignment and parsing of equipment identification and timelines in operation and maintenance data, making it difficult to identify and extract key points that can be structured into fields;

[0006] 2. Due to the lack of a verifiable backlink mechanism for evidence location, there is a lack of corresponding evidence anchors when outputting conclusions, resulting in high manual review costs and uncontrollable alarm location time. Summary of the Invention

[0007] This invention provides a structured information retrieval method based on a large language model, aiming to solve the problems of existing technologies that lack unified alignment and parsing of equipment and time in operation and maintenance data, and lack of anchor points in the evidence verification backlink mechanism.

[0008] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: A structured information retrieval method based on a large language model is provided, comprising:

[0009] Collect and clean operation and maintenance data, identify structural elements and extract fields, deploy adapters, access multi-source operation and maintenance data in read-only mode, synchronize original entries with timestamps and identify sources, map equipment, components and personnel based on asset catalog and record mapping confidence, and extract structural elements from records using a hybrid strategy of rule priority and model assistance.

[0010] Construct a three-chain integrated evidence graph, parse and transform retrieval tasks, inject three types of metadata into candidate evidence, establish corresponding nodes and edges in the graph database, parse the query into a structured retrieval intent, expand the candidate evidence set along the feasibility chain, timeliness chain, and source chain to form a candidate subgraph, and calculate the comprehensive score for ranking.

[0011] The system verifies and filters feasibility constraints, identifies sources to generate structured answers, applies a set of rules to candidate evidence to verify constraints, calls LLM to perform feasibility inference and outputs the confidence level of feasibility inference, queries on-site resources in real time and integrates the measured results with the inference confidence level and source credibility according to preset weights to obtain the feasibility score of candidate evidence. For candidate evidence that passes the feasibility verification, it identifies precise anchor points for each piece of evidence along the source chain and generates structured answers.

[0012] As a preferred implementation, the specific steps for collecting and cleaning operation and maintenance data, identifying structural elements and extracting fields are as follows: deploying adapters on gateway devices and edge servers in the industrial network, accessing multi-source data in parallel through the adapters, establishing connections in read-only mode, uniformly aligning the timestamps of the collected raw entries, and extracting image metadata through optical character recognition.

[0013] The asset catalog maps equipment, components, and personnel to unique identifiers, unifies numerical values ​​of different units and formats in the source data to standard units while retaining the original values ​​and units, normalizes different naming conventions to standard field names according to the predefined field mapping table in the asset catalog, removes noise from natural language text, and identifies and corrects syntax errors.

[0014] Structural elements are extracted from metadata by combining a predefined set of deterministic rules with a trained LLM. After field extraction, vectorization and similarity deduplication are performed. The deterministic rule set is used to extract and verify fields and security information. If the rules are not covered, the LLM is called to supplement and infer. The LLM outputs the mapping confidence score. The strategy of extracting structural elements by combining the two is to first perform the extraction of the rule layer. When the rule hit rate is higher than 85%, a high mapping confidence is given. Otherwise, the LLM is called to complete the extraction.

[0015] As a preferred implementation, the specific steps for constructing a three-chain integrated evidence graph and parsing and transforming the retrieval task are as follows: the fields of the extracted structural elements are combined with confidence levels as candidate evidence, and three types of metadata are injected into each candidate evidence: source metadata, version timeliness metadata, and process feasibility metadata, forming an evidence set combining metadata and confidence levels. A graph database is constructed based on the evidence set, and evidence, assets, work orders, and events are defined as node types, and source and dependency are defined as edge types. The evidence set is traversed and read, and an evidence node is created in the graph database for each piece of evidence and its attributes are written. At the same time, the asset node corresponding to the evidence is created and its attributes are written. The work orders associated with the evidence are used to create work order nodes and an edge is established between the two using the evidence source.

[0016] Based on the source metadata of the evidence, source edges are established between evidence nodes and source nodes. The edges record the jumpable location information, collection method and initial source credibility. The source chain is used to trace back the original source and calculate the source credibility. By parsing the version information of documents and evidence, the version relationship between documents and clauses is identified. Revision edges are established between document nodes. At the same time, edges are established between evidence nodes and corresponding document nodes. The release time and version number are recorded on the edges. According to the data of the rule engine, asset catalog and real-time resources, the feasibility metadata of each piece of evidence is parsed and dependency edges are established between evidence nodes and asset, personnel and work order nodes. The required resources, qualifications and edge confidence are recorded on the edges, forming an evidence graph that combines process feasibility chain, version timeliness chain and source coordinate chain.

[0017] The user's natural language query is parsed into a structured search intent, which is automatically identified and extracted by the query parsing module in the retrieval system. For queries that are ambiguous or unclear, deterministic rules are invoked to complete them. The search intent is mapped into a hybrid search task, and candidate evidence nodes are retrieved from the vector index using semantic vector retrieval. At the same time, the filtering conditions of the fields are translated and filtered on the graph. The candidate set is expanded along the feasibility chain, timeliness chain, and source chain to construct a candidate subgraph. A comprehensive score is calculated for each evidence node in the candidate subgraph, using the following formula:

[0018] Score e =α1S sem +α2S fea +α3S time +α4S src ,

[0019] Among them, Score e The overall score of evidence node e is represented by α1, α2, α3, α4, and S. sem S represents the semantic score. fea S represents the feasibility score. timeS represents the timeliness score. src Indicates the source score;

[0020] The evidence nodes are sorted according to the comprehensive score. Based on the sorting results, high-scoring evidence is aggregated into a step sequence, with each step accompanied by its supporting evidence set and comprehensive score.

[0021] Source metadata is the original records obtained from various systems. The original text location is directly located by adding timestamps and unique device identifiers. Version timeliness metadata records the version information of documents associated with evidence through the version number of the database and the version time of work orders and reports. Process feasibility metadata extracts feasibility from work order records and accident reviews, sorts the priority of solutions, and finds other alternative operations in the graph.

[0022] As a preferred implementation, the specific steps for verifying and filtering feasibility constraints and identifying the source to generate the structured answer are as follows: Feasibility verification and filtering of candidate evidence are performed using a deterministic rule set combined with LLM; based on the process feasibility metadata injected into the evidence record, the current resource status is queried in real time; the necessary resources are estimated according to the predefined rule set; for cases where the rules are not covered, LLM and regression models are called to complete the data and provide a confidence level; and the confidence level of the model, the source credibility, and the feasibility inference confidence level are fused according to preset weights to obtain a feasibility score.

[0023] Candidate evidence is filtered based on a predefined confidence threshold, and the formula for calculating the feasibility score is as follows:

[0024] feas=β1c model +β2c src +β3c feas ,

[0025] Where feas represents the feasibility score, β1 represents the model weight, β2 represents the source weight, β3 represents the feasibility weight, and c model c represents the model confidence level. src Indicates the credibility of the source, c feas Indicates the confidence level of the feasibility inference;

[0026] Based on the three-chain integrated evidence graph, the evidence that has passed the feasibility verification is anchored along the source chain, the timeliness is verified along the version chain, and the dependency relationship is confirmed along the feasibility chain. The comprehensive score of the evidence nodes is ranked first, and a structured answer is generated according to the predetermined output template. The structured answer is presented as a step sequence. For answers whose comprehensive score is in the range of manual review, they are automatically marked as needing review and the relevant context, evidence anchor points and recommended review items are pushed. The entire generation process records the audit items at the same time, and the manual review results and execution feedback are fed back to the rule base and model training set to complete the closed-loop improvement.

[0027] The beneficial effects of this invention are as follows:

[0028] 1. This invention achieves operational decision support from relevant retrieval to executable, verifiable, and traceable operations and maintenance decisions by integrating a three-chain evidence graph, rule-based and model-assisted extraction and verification, and a closed-loop mechanism of evidence anchors and audit feedback, thereby improving on-site investigation efficiency and compliance audit capabilities.

[0029] 2. This invention improves extraction accuracy by unifying and aligning assets and timelines. It uses a standardized asset catalog to uniquely identify and map equipment, components, and personnel, avoiding information silos caused by inconsistent source naming and time misalignment. This improves the accuracy and coverage of identifying structured fields from maintenance texts and forms. Attached Figure Description

[0031] Figure 1 This is a flowchart of a structured information retrieval method based on a large language model.

[0032] Figure 2 This is a comparison chart of the performance of a structured information retrieval method based on a large language model. Detailed Implementation

[0033] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0034] Example 1, as Figure 1 It is a structured information retrieval method based on a large language model, which includes the following steps:

[0035] Collect and clean operation and maintenance data, identify structural elements and extract fields;

[0036] Construct a three-chain integrated evidence map to analyze and transform retrieval tasks;

[0037] Verify the feasibility constraints of the filtering and calibrate the source to generate a structured answer.

[0038] The specific implementation steps are as follows: A structured information retrieval method based on a large language model, wherein the specific steps for collecting and cleaning operation and maintenance data, identifying structural elements, and extracting fields are as follows:

[0039] In industrial networks, adapters are deployed on gateway devices and edge servers to access multi-source data in parallel through the adapters. Connections are established in read-only mode. The timestamps of the collected raw items, such as inspection forms, alarm streams, accident debriefings, and multimedia attachments, are uniformly aligned. Image metadata is extracted through optical character recognition. Specifically, optical character recognition is a technology that automatically recognizes the text in an image and converts it into editable and searchable electronic text.

[0040] The asset catalog maps equipment, components, and personnel to unique identifiers, unifies numerical values ​​of different units and formats in the source data to standard units while retaining the original values ​​and units, normalizes different naming conventions to standard field names according to the predefined field mapping table in the asset catalog, removes noise from natural language text, and identifies and corrects syntax errors.

[0041] Specifically, the asset catalog collects information from different data sources such as maintenance records and equipment manuals, cleans and standardizes it, maps assets, components and personnel to unique identifiers through a rule engine, and verifies their consistency and accuracy to form a structured data table.

[0042] By combining a predefined set of deterministic rules with a trained large language model (LLM), structural elements such as action steps, preconditions, required resources, and key parameters are extracted from metadata. After field extraction, vectorization and similarity deduplication are performed.

[0043] Specifically, the predefined deterministic rules combined with LLM for extracting structural elements employ a deterministic set of rules to prioritize and strongly validate key fields and security-related information. If a rule is not covered, LLM is invoked for supplementation and inference. Rules include formatted templates and form column mapping rules, regular expressions and pattern matching, aliases and synonym mapping tables, etc. LLM includes a text classifier, semantic vectors, an instruction-based generative large language model, and a regression model for confidence scores. LLM outputs a mapped confidence score, specifically by summarizing features such as the log probability and output entropy of the model's answers and mapping them to a confidence score of [0,1] using a calibration method. The strategy for extracting structural elements by combining these two approaches involves first performing rule-based extraction. When the rule hit rate is higher than 85% and the field is a security-critical item, a high mapping confidence score is assigned; otherwise, LLM is invoked for completion.

[0044] The specific steps for constructing a unified three-chain evidence graph and analyzing and transforming the retrieval task are as follows:

[0045] The extracted structural elements are combined with confidence scores to form candidate evidence. Three types of metadata are injected into each candidate evidence: source metadata, version validity metadata, and process feasibility metadata, forming an evidence set that combines metadata and confidence scores. A graph database is constructed based on the evidence set. Evidence, assets, work orders, and events are defined as node types, and source and dependency are defined as edge types. The evidence set is traversed and read. For each piece of evidence, an evidence node is created in the graph database and its attributes are written. At the same time, the asset node corresponding to the evidence is created and its attributes are written. The work orders associated with the evidence are used to create work order nodes and establish an edge between the two using the evidence source.

[0046] Specifically, the aforementioned source metadata, version timeliness metadata, and process feasibility metadata include source metadata which is the original record obtained from each system, and the original text location is directly located by adding timestamps and unique device identifiers; version timeliness metadata records the version information of documents associated with evidence through the version number of the database and the version time of work orders and reports; and process feasibility metadata extracts feasibility from work order records and accident reviews, sorts the priority of solutions, and searches for other alternative operations in the graph.

[0047] Based on the source metadata of the evidence, source edges are established between evidence nodes and source nodes. The edges record the jumpable location information, collection method, and initial source credibility. The source chain is used to trace back the original source and calculate the source credibility. By parsing the version information of documents and evidence, the version relationship between documents and clauses is identified, and revision edges are established between document nodes. At the same time, edges are established between evidence nodes and corresponding document nodes, and the release time and version number are recorded on the edges. Based on the data of the rule engine, asset catalog, and real-time resources, the feasibility metadata of each piece of evidence is parsed, and dependency edges are established between evidence nodes and asset, personnel, and work order nodes. The edges record the required resources, qualifications, and edge confidence, forming an evidence graph that integrates the process feasibility chain, version timeliness chain, and source coordinate chain.

[0048] The user's natural language query is parsed into a structured search intent. Specifically, the query parsing module in the retrieval system automatically identifies and extracts the target asset, time window, constraints, expected return format, and keyword slots. For ambiguous or unclear queries, deterministic rules are invoked to complete the query. The search intent is mapped into a hybrid search task, using semantic vector retrieval to retrieve candidate evidence nodes from the vector index. Simultaneously, the filtering conditions of the fields are translated and filtered on the graph. The candidate set is expanded along the feasibility chain, timeliness chain, and source chain to construct a candidate subgraph. A comprehensive score is calculated for each evidence node in the candidate subgraph, using the following formula:

[0049] Score e =α1S sem +α2S fea +α3Stime +α4S src ,

[0050] Among them, Score e The overall score of evidence node e is represented by α1, α2, α3, α4, and S. sem S represents the semantic score. fea S represents the feasibility score. time S represents the timeliness score. src Indicates the source score;

[0051] The evidence nodes are sorted according to the comprehensive score. For example, the weights can be α1 = 0.45, α2 = 0.25, α3 = 0.15, and α4 = 0.15. Based on the sorting results, high-scoring evidence is aggregated into a step sequence, with each step accompanied by its supporting evidence set and comprehensive score.

[0052] The specific steps for verifying the feasibility constraints of filtering and identifying the source-generated structured answer are as follows:

[0053] Feasibility verification and filtering of candidate evidence are performed using a deterministic rule set combined with LLM (Limited Ledger Model). Based on the process feasibility metadata injected into the evidence record, the current resource status is queried in real time. The required resources are estimated according to the predefined rule set. For cases not covered by the rules, LLM and regression models are invoked to complete the data and provide confidence scores. The feasibility score is calculated by fusing the confidence scores of the model, source credibility, and feasibility inference confidence scores according to preset weights. Source credibility is obtained by weighting the basic trust value of the source type with the version timeliness of the record, the number of independent supports from multiple sources, the reliability of the collection method, and historical verification results according to a mapping model with preset weights. Feasibility inference confidence scores are obtained by fusing and normalizing the initial probabilistic output of the rule engine on the feasibility of the operation with real-time resource verification, historical execution success rate, and manual confirmation signals according to a calibration model with preset weights. Candidate evidence is filtered according to a predefined confidence threshold. The formula for calculating the feasibility score is as follows:

[0054] feas=β1c model +β2c src +β3c feas ,

[0055] Where feas represents the feasibility score, β1 represents the model weight, β2 represents the source weight, β3 represents the feasibility weight, and c model c represents the model confidence level. src Indicates the credibility of the source, c feas Indicates the confidence level of the feasibility inference;

[0056] Based on the three-chain integrated evidence graph, the evidence that has passed the feasibility verification is anchored along the source chain, such as work order number, media coordinates, etc.; the timeliness is verified along the version chain; and the dependency relationship is confirmed along the feasibility chain. The comprehensive score of the evidence node is ranked first, and a structured answer is generated according to the predetermined output template. The structured answer is presented as a step sequence. For answers with a comprehensive score in the range of 0.60-0.80 that requires manual review, the relevant context, evidence anchor points and recommended review items are pushed. The entire generation process records the audit items at the same time, and the manual review results and execution feedback are fed back to the rule base and model training set to complete the closed-loop improvement.

[0057] like Figure 2 The graph shows the performance comparison of a structured information retrieval method based on a large language model. The horizontal axis lists key performance indicators, and the vertical axis represents the exemplified performance scores, ranging from 0 to 100. Higher values ​​indicate better performance. The aim is to intuitively demonstrate the expected improvement of the present invention in key capabilities compared to typical existing technologies.

[0058] Example 2, based on Example 1 above, presents a structured information retrieval method based on a large language model for operation and maintenance and troubleshooting scenarios in process industries. The specific solution is as follows:

[0059] Step 1: Deploy multimodal data collectors on the factory gateway and edge server to collect data in streaming and batch parallel modes. For items containing images, use OCR technology to convert them into text, and use speech-to-text technology for speech items. Summarize and unify the obtained text and fields, fuse and calibrate the confidence level, and use auditable deterministic rules and LLM sets to extract data. Combine the rule hit confidence level, LLM output probability and source credibility according to preset weights and map them through a regression model to generate candidate evidence for high-confidence results.

[0060] Step 2: After cleaning, candidate evidence combined with confidence and location anchor points is written into a temporary evidence set. Based on a predefined graph pattern, node types and multiple edge types are established. Nodes include evidence, assets, procedural terms, work orders, and events. Edges include at least source chains, version validity chains, and process feasibility chains. Confidence, timestamps, and other meta-attributes are maintained for each node and edge. The retrieval and parsing uses a rule-first, model-assisted query parser to convert natural language queries into structured retrieval intents, which are then mapped to graph retrieval templates. Semantic scores, feasibility scores, validity scores, and source scores are calculated in parallel for each evidence node on the candidate subgraph and weighted according to preset weights to synthesize a comprehensive score.

[0061] Step 3: Prioritize constraints using an auditable set of deterministic rules. For items not covered by the rules, use LLM and regression models to complete the preconditions and resource requirements, and output the feasibility inference confidence level. Simultaneously, initiate a verification request to the resource system. Combine the actual test results with the inferred confidence level to form a feasibility score, which is propagated along the feasibility dependency edges on the evidence graph. When a node becomes infeasible, dynamically adjust the feasibility score of the relevant evidence and automatically trigger alternative path search to synthesize alternative solutions. Verify the timeliness of the documents on the version chain, anchor a precise location string for each piece of evidence on the source chain, and record the collection method and initial confidence level. For evidence that passes the feasibility verification, generate a structured answer according to a predetermined template.

[0062] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. A structured information retrieval method based on a large language model, characterized in that, include: Collect and clean operation and maintenance data, identify structural elements and extract fields, deploy adapters, access multi-source operation and maintenance data in read-only mode, synchronize original entries with timestamps and identify sources, map equipment, components and personnel based on asset catalog and record mapping confidence, and extract structural elements from records using a hybrid strategy of rule priority and model assistance. Construct a three-chain integrated evidence graph, parse and transform retrieval tasks, inject three types of metadata into candidate evidence, establish corresponding nodes and edges in the graph database, parse the query into a structured retrieval intent, expand the candidate evidence set along the feasibility chain, timeliness chain, and source chain to form a candidate subgraph, and calculate the comprehensive score for ranking. The system verifies and filters feasibility constraints, identifies sources to generate structured answers, applies a set of rules to candidate evidence to verify constraints, calls LLM to perform feasibility inference and outputs the confidence level of the feasibility inference, queries on-site resources in real time and integrates the measured results with the inference confidence level and source credibility according to preset weights to obtain the feasibility score of candidate evidence. For candidate evidence that passes the feasibility verification, it identifies precise anchor points for each piece of evidence along the source chain and generates structured answers.

2. The structured information retrieval method based on a large language model according to claim 1, characterized in that: The specific steps for collecting and cleaning operation and maintenance data, identifying structural elements, and extracting fields are as follows: In industrial networks, adapters are deployed on gateway devices and edge servers to access multi-source data in parallel through the adapters. Connections are established in read-only mode, and the timestamps of the collected raw entries are uniformly aligned. Image metadata is extracted through optical character recognition. The asset catalog maps equipment, components, and personnel to unique identifiers, unifies numerical values ​​of different units and formats in the source data to standard units while retaining the original values ​​and units, normalizes different naming conventions to standard field names according to the predefined field mapping table in the asset catalog, removes noise from natural language text, and identifies and corrects syntax errors.

3. The structured information retrieval method based on a large language model according to claim 2, characterized in that: The specific steps for collecting and cleaning operational data, identifying structural elements, and extracting fields also include: Structural elements are extracted from metadata by combining a predefined set of deterministic rules with a trained LLM. After field extraction, vectorization and similarity deduplication are performed. The deterministic rule set is used to extract and verify fields and security information. If the rules are not covered, the LLM is called to supplement and infer. The LLM outputs the mapping confidence score. The strategy of extracting structural elements by combining the two is to first perform the extraction of the rule layer. When the rule hit rate is higher than 85%, a high mapping confidence is given. Otherwise, the LLM is called to complete the extraction.

4. The structured information retrieval method based on a large language model according to claim 1, characterized in that: The specific steps for constructing a unified three-chain evidence graph and analyzing and transforming the retrieval task are as follows: The extracted structural elements are combined with confidence scores to form candidate evidence. Three types of metadata are injected into each candidate piece of evidence: source metadata, version validity metadata, and process feasibility metadata, forming an evidence set that combines metadata and confidence scores. A graph database is constructed based on the evidence set, defining evidence, assets, work orders, and events as node types, and source and dependency as edge types. The evidence set is traversed and read, and an evidence node is created in the graph database for each piece of evidence and its attributes are written. At the same time, the asset node corresponding to the evidence is created and its attributes are written. Work orders associated with the evidence are used to create work order nodes and an edge is established between the two using the evidence source.

5. The structured information retrieval method based on a large language model according to claim 4, characterized in that: The specific steps for constructing a unified three-chain evidence graph and analyzing and transforming retrieval tasks also include: Based on the source metadata of the evidence, source edges are established between evidence nodes and source nodes. The edges record the jumpable location information, collection method, and initial source credibility. The source chain is used to trace back the original source and calculate the source credibility. By parsing the version information of documents and evidence, the version relationship between documents and clauses is identified. Revision edges are established between document nodes, and edges are established between evidence nodes and corresponding document nodes. The release time and version number are recorded on the edges. Based on the data of the rule engine, asset catalog, and real-time resources, the feasibility metadata of each piece of evidence is parsed, and dependency edges are established between evidence nodes and asset, personnel, and work order nodes. The edges record the required resources, qualifications, and edge confidence, forming an evidence graph that integrates the process feasibility chain, version timeliness chain, and source coordinate chain.

6. The structured information retrieval method based on a large language model according to claim 4, characterized in that: The specific steps for constructing a unified three-chain evidence graph and analyzing and transforming retrieval tasks also include: The user's natural language query is parsed into a structured search intent, which is automatically identified and extracted by the query parsing module in the retrieval system. For queries that are ambiguous or unclear, deterministic rules are invoked to complete them. The search intent is mapped into a hybrid search task, and candidate evidence nodes are retrieved from the vector index using semantic vector retrieval. At the same time, the filtering conditions of the fields are translated and filtered on the graph. The candidate set is expanded along the feasibility chain, timeliness chain, and source chain to construct a candidate subgraph. A comprehensive score is calculated for each evidence node in the candidate subgraph, using the following formula: Score e =α1S sem +α2S fea +α3S time +α4S src , Among them, Score e The overall score of evidence node e is represented by α1, α2, α3, α4, and S. sem S represents the semantic score. fea S represents the feasibility score. time S represents the timeliness score. src Indicates the source score; The evidence nodes are sorted according to the overall score. Based on the sorting results, high-scoring evidence is aggregated into a sequence of steps, with each step accompanied by its supporting evidence set and overall score.

7. The structured information retrieval method based on a large language model according to claim 4, characterized in that: The specific steps for constructing a unified three-chain evidence graph and analyzing and transforming retrieval tasks also include: Source metadata is the original records obtained from various systems. The original text location is directly located by adding timestamps and unique device identifiers. Version timeliness metadata records the version information of documents associated with evidence through the version number of the database and the version time of work orders and reports. Process feasibility metadata extracts feasibility from work order records and accident reviews, sorts the priority of solutions, and finds other alternative operations in the graph.

8. The structured information retrieval method based on a large language model according to claim 1, characterized in that: The specific steps for verifying the feasibility constraints of filtering and identifying the source-generated structured answer are as follows: Feasibility verification and filtering of candidate evidence are performed by combining a deterministic set of rules with LLM. Based on the process feasibility metadata injected into the evidence record, the current resource status is queried in real time. The necessary resources are estimated according to the predefined set of rules. For cases not covered by the rules, LLM and regression models are called to complete the data and provide a confidence score. The confidence score of the model, the credibility of the source, and the confidence score of the feasibility inference are fused according to preset weights.

9. The structured information retrieval method based on a large language model according to claim 8, characterized in that: The specific steps for verifying the feasibility constraints of filtering and labeling the sources of structured answers also include: Candidate evidence is filtered based on a predefined confidence threshold, and the formula for calculating the feasibility score is as follows: ugly=β1c model +β2c src +β3c feas , Where feas represents the feasibility score, β1 represents the model weight, β2 represents the source weight, β3 represents the feasibility weight, and c model c represents the model confidence level. src Indicates the credibility of the source, c feas This indicates the confidence level of the feasibility inference.

10. The structured information retrieval method based on a large language model according to claim 8, characterized in that: The specific steps for verifying the feasibility constraints of filtering and labeling the sources of structured answers also include: Based on the three-chain integrated evidence graph, the evidence that has passed the feasibility verification is anchored along the source chain, the timeliness is verified along the version chain, and the dependency relationship is confirmed along the feasibility chain. The comprehensive score of the evidence nodes is ranked first, and a structured answer is generated according to the predetermined output template. The structured answer is presented as a step sequence. For answers whose comprehensive score is in the range of manual review, they are automatically marked as needing review and the relevant context, evidence anchor points and recommended review items are pushed. The entire generation process records the audit items at the same time.

Images