Domain graph search enhancement generation method and system

Abstract

The application discloses a field graph retrieval enhancement generation method and system, comprising the following steps: collecting multi-source heterogeneous documents of a vertical field to generate triplets; writing the standardized triplets into a graph database to construct a field knowledge graph, saving the source, timestamp and expert review state in the node attribute, and simultaneously performing field vectorization coding on the standardized entity and establishing an approximate nearest neighbor index; analyzing a user query to obtain a candidate entity, generating a graph query statement after matching the candidate entity to a standardized entity through a vector index, executing multi-hop path tracking in the field knowledge graph, and deleting paths that do not conform to industrial logic according to a field rule library; obtaining and retaining original data associated with the path, reordering the associated original data and outputting a credible data set from an expert review model; and generating a response with a source identifier by using a fact constraint prompt word, and performing illusion detection and correction. The application can improve the accuracy, professionalism and traceability of vertical field retrieval generation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence information retrieval and knowledge graph technology, specifically to a method and system for enhancing the generation of domain graph retrieval. Background Technology

[0002] Retrieval enhancement generation technology combines information retrieval modules with large language models to extract context related to user questions from document libraries. This context, along with the question, is then input into the large language model to generate answers, improving the accuracy and factual consistency of question-and-answer responses to some extent. With the rapid growth of patent documents, academic papers, technical information, and industry reports in vertical industries, knowledge resources are becoming increasingly massive, heterogeneous, and highly specialized. Traditional retrieval methods based on text block similarity are struggling to reliably meet the demands of specialized retrieval needs.

[0003] In fields such as advanced materials and intelligent manufacturing, the same technical concept often exists in multiple forms, such as full names, abbreviations, English acronyms, and different translations. Traditional vector matching struggles to identify their semantic equivalence, easily leading to retrieval omissions or redundancy. Domain knowledge also exhibits strong correlation and hierarchy. For example, material composition affects microstructure, microstructure affects macroscopic performance, and macroscopic performance is related to process parameters and device reliability. Flat text blocks struggle to express these structured relationships, easily leading to biased path finding in multi-hop reasoning problems.

[0004] Existing solutions attempt to incorporate knowledge graphs or graph neural networks to enhance reasoning capabilities. However, the definitions of entities and relationships in general knowledge graphs are too broad, making it difficult to accurately depict the technical concept systems unique to vertical industries. Furthermore, updating and expanding the graph structure after its establishment is challenging, and reasoning paths often rely on predefined rules, resulting in insufficient flexibility. On the other hand, large language models may still produce illusions when generating responses. Existing systems generally lack effective mechanisms for knowledge tracing, industry logic verification, and expert review, making it difficult for users to confirm whether the generated content matches the original data.

[0005] Therefore, there is a need for a vertical retrieval enhancement generation scheme that combines domain ontology, knowledge graph, vector index, path tracing, expert review, fact constraint generation, and incremental updates to improve the accuracy, interpretability, and credibility of professional question answering. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for enhancing the generation of domain map retrieval, in order to solve the problems of insufficient concept standardization, insufficient understanding of structured knowledge associations, high risk of generating illusionary content, lack of industry logic verification, weak knowledge traceability, and insufficient knowledge updating capability in the generation of enhanced retrieval in vertical industries.

[0007] This invention provides a method for enhancing the generation of domain map retrieval, comprising the following steps: Step S1: Collect multi-source heterogeneous documents from vertical domains. After cleaning, deduplication, and semantic segmentation, use a large language model to extract entity relation triples by domain extraction templates. Then, use the domain ontology library to perform synonym unification and hierarchical generalization on the entities in the triples. Step S2: Write the standardized triples into the graph database to construct the domain knowledge graph. Save the source, timestamp and expert review status in the node attributes. At the same time, perform domain vectorization encoding on the standardized entities and establish an approximate nearest neighbor index. Step S3: Parse the user query to obtain candidate entities, match standardized entities through vector index to generate graph query statements, perform multi-hop path tracing in the domain knowledge graph, and delete paths that do not conform to industry logic according to the domain rule base; Step S4: Obtain the original data associated with the retention path, reorder it according to timeliness, source authority, technical relevance and industry influence, and then use the expert review model trained with historical review data to identify material and process mismatch, parameter anomalies and causal errors, and output a reliable dataset. Step S5: Construct fact constraint prompts based on trusted datasets, preserved paths, and user queries; control the large language model to generate responses with source identifiers; and correct unsubstantiated content after comparing the responses with the original data.

[0008] Preferably, the multi-source heterogeneous documents in step S1 include patent documents, academic papers, technical information, and industry reports; the cleaning includes removing noise content, unifying encoding and text format; the deduplication uses the SimHash algorithm to identify duplicate documents; the semantic segmentation is divided into paragraphs or sentences while preserving contextual integrity; the domain extraction template limits entity types and relationship types, with entity types including at least material entities, device entities, process entities, equipment entities, parameter entities, organization entities, and person entities; and relationship types including at least material-process relationships, process-equipment relationships, equipment-parameter relationships, organization-technology relationships, technology-application relationships, and person-organization relationships; the extraction results are filtered by confidence to form an initial set of triples.

[0009] Preferably, the domain ontology library is constructed for the fields of advanced materials and intelligent manufacturing, including six sub-ontologies: material system ontology, device structure ontology, process technology ontology, characterization and testing ontology, equipment and instrument ontology, and industrial application ontology. In step S1, the initial triples are manually verified through an expert review interface to correct entity ambiguity, relationship labeling errors, and missing key attributes. The synonym unification maps multiple expressions of the same technical concept to standard entities, and the hierarchical generalization extends specific technical concepts to their parent concepts to adapt to queries of different granularities.

[0010] Preferably, the node attributes in the graph database in step S2 further include entity standard name, entity type, synonym list, original data source and data type, and the edge attributes include relation type, relation confidence, data source, expert review status and frequency of occurrence; the domain vectorization encoding adopts a language model that has been pre-trained on vertical domain corpus to map standardized entities into semantic vectors; the approximate nearest neighbor index is a vector index built based on the HNSW algorithm, used to perform similarity retrieval on entity semantic vectors.

[0011] Preferably, in step S3, when parsing the user query, keywords are first extracted to form candidate entities, and then the candidate entities are input into the domain vectorization model and retrieved in the near nearest neighbor index to obtain a standardized entity set; the graph query statement is generated by the large language model based on the standardized entity set and the query intent, and a multi-hop path query is performed in the graph database; the domain rule base includes at least material and process compatibility rules, reasonable range rules for process parameters, and causal association rules for devices, which are used to determine whether the relationship between adjacent entities in the path conforms to the vertical domain technical logic.

[0012] Preferably, the comprehensive score S used for re-ranking in step S4 satisfies: S=α·T+β·A+γ·R+δ·I, where T is the data timeliness score, A is the source authority score, R is the technology relevance score, I is the industry influence score, and α, β, γ, and δ are weight coefficients and α+β+γ+δ=1; the technology relevance score is calculated based on word frequency-inverse document frequency and entity co-occurrence frequency, the industry influence score is calculated based on citation frequency, patent family size, and market application breadth, and the weight coefficients are dynamically adjusted according to the query type.

[0013] Preferably, the expert review model is trained based on historical expert review records, which are marked with three types of industry logic errors: material and process mismatch, abnormal parameter range, and causal relationship error. In S4, the reordered original data is subjected to secondary verification. Data that passes both the industry correctness verification and the secondary verification is marked as trustworthy, while data that fails the verification is marked as abnormal and excluded. The review results output by the expert review model are also used to update the prompt rules of the domain rule base and the expert review interface.

[0014] Preferably, the fact constraint prompts in step S5 are at least limited to: summarizing only based on trusted datasets and retained paths, without introducing external knowledge, marking the source number of the cited content, and writing the graph path in the prompts in the form of entity-relationship-entity; the consistency comparison breaks down the factual statements in the response into units to be verified, matches each unit with trusted datasets and retained paths, deletes or rewrites the units to be verified that have no corresponding source, and retains the source identifier and reasoning path identifier in the final response.

[0015] Preferably, the method also includes an incremental update step for the knowledge base: continuously monitoring new patents, new papers, new technical information, and new industry reports; regularly performing data collection, triple extraction, standardization processing, and graph index updates; aligning and disambiguating new entities with existing entities in the domain knowledge graph; merging entities when the semantic vector similarity and attribute matching degree of the new entity and the existing entity both meet preset conditions; otherwise, creating new entity nodes; and dynamically adjusting the domain ontology, domain rule base, and expert review model based on user error correction feedback, satisfaction feedback, and expert review results.

[0016] Another aspect of the present invention provides a domain graph retrieval enhancement generation system, which is applied to a domain graph retrieval enhancement generation method. The domain graph retrieval enhancement generation system includes: The data construction module is used to collect multi-source heterogeneous documents in the vertical domain, and perform cleaning, deduplication, semantic segmentation, triple extraction, and synonym normalization and hierarchical generalization based on the domain ontology library. The graph indexing module is used to write standardized triples into the graph database to form a domain knowledge graph, save the source, timestamp and expert review status, and perform domain vectorization encoding and index construction on standardized entities; The query and retrieval module is used to parse user queries, match standardized entities, generate graph query statements, and perform multi-hop path tracing, and delete paths that do not conform to industry logic based on the domain rule base; The reordering review module is used to reorder the original data with preserved path associations according to multidimensional indicators, and the expert review model identifies material and process mismatches, parameter anomalies and causal errors to output a reliable dataset. The generation and verification module is used to construct fact constraint prompts based on trusted datasets, preserved paths, and user queries, control the large language model to generate responses with source identifiers, and perform consistency comparisons and corrections.

[0017] Compared with the prior art, the domain map retrieval enhancement generation method of the present invention has the following significant improvements: This invention standardizes and semantically generalizes entity relationships through a domain ontology library and expert review mechanism, which can identify multiple expressions of the same technical concept, alleviate retrieval omissions caused by the mixed use of abbreviations, full names, English abbreviations and hierarchical concepts, and improve the accuracy and recall of vertical domain concept matching.

[0018] This invention combines domain knowledge graphs with vector indexes. During retrieval, it first locates standardized entities, then performs multi-hop path tracing along the graph, and uses a domain rule base to exclude pseudo-association paths that do not conform to industry logic. It can express the hierarchical relationships between materials, processes, equipment, parameters, devices, and applications, and enhance the structured reasoning capabilities of complex queries.

[0019] This invention introduces fact-constraining prompts, source identifiers, and illusion detection into the response generation process, restricting the generated content to the scope of trusted datasets and retained paths, and correcting or deleting unsubstantiated content. This can reduce the risk of illusion in large language models and establish a traceable link between the response content and the original data and graph paths.

[0020] This invention establishes a collaborative review mechanism of manual verification and model verification through an expert review interface, a domain rule base, and an expert review model. This mechanism can identify material and process mismatches, abnormal parameter ranges, and causal relationship errors, thereby improving the professionalism and credibility of the generated content. At the same time, it supports continuous access to new knowledge, entity alignment and disambiguation, and dynamic optimization of the ontology library, adapting to the rapidly evolving knowledge service needs of vertical industries. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the overall process of a domain map retrieval enhancement generation method according to one embodiment of the present invention. Figure 2 This is a block diagram of a domain map retrieval enhancement generation system according to one embodiment of the present invention; Figure 3 This is a flowchart of domain knowledge graph construction and indexing in one embodiment of the present invention; Figure 4 This is a flowchart of query parsing, path tracing, and expert review in one embodiment of the present invention; Figure 5 This is a flowchart illustrating the response generation, hallucination detection, and incremental knowledge base update process in one embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0024] like Figure 1As shown, the method of this implementation includes five main stages: data construction, graph indexing, query retrieval, rearrangement review, and generation verification, and can further perform incremental updates to the knowledge base. The data construction stage completes multi-source heterogeneous data collection, preprocessing, entity relation triple extraction, standardization, and generalization; the graph indexing stage completes graph database storage, domain vectorization encoding, and the construction of an approximate nearest neighbor index; the query retrieval stage completes query parsing, entity matching, graph querying, and multi-hop path tracing; the rearrangement review stage completes multi-dimensional scoring of the original data, secondary verification of the expert review model, and credibility labeling; and the generation verification stage completes the construction of fact constraint prompts, structured response generation, and illusion detection.

[0025] During the data construction phase, the system collects data from vertical sectors through interfaces with patent databases, academic databases, industry news websites, standardization organization websites, and market research institutions. Taking the advanced materials and intelligent manufacturing field as an example, the collected data includes full-text patents, claims, specifications, journal articles, conference papers, dissertations, technical news, standards and specifications, technology roadmaps, market analysis reports, and industry chain maps. The collected data is uniformly converted to structured text format, undergoes text cleaning to remove advertisements, navigation bars, and header / footer noise, is uniformly converted to UTF-8 encoding, and uses the SimHash algorithm to identify and remove duplicate documents. Then, it is semantically segmented according to paragraphs and sentences, preserving the semantic integrity of the context during segmentation to avoid breaking the same technical facts into unrelated fragments.

[0026] The preprocessed text input is extracted using a large language model framework based on prompting engineering. The extraction template predefines the entity and relation types to be identified. Entity types include material entities, device entities, process entities, equipment entities, parameter entities, organization entities, and person entities; relation types include material-process relationships, process-equipment relationships, equipment-parameter relationships, organization-technology relationships, technology-application relationships, and person-organization relationships. For the input text "Sumitomo Electric Industries uses HVPE process to epitaxially grow GaN thin films on 4H-SiC substrates at a growth temperature of 1050°C", the model can extract triples: (Sumitomo Electric Industries, uses, HVPE process), (HVPE process, applied to, GaN epitaxial growth), (4H-SiC, as, substrate material), (growth temperature, parameter is, 1050°C). The extraction results are filtered by confidence level, retaining only triples with a confidence level higher than a preset threshold to form an initial set of triples.

[0027] The initial set of triples undergoes standardization and generalization. The domain expert ontology, designed with the participation of experts possessing experience in vertical industries, comprises six sub-ontologies: Materials System Ontology, Device Structure Ontology, Process Technology Ontology, Characterization and Testing Ontology, Equipment and Instrumentation Ontology, and Industrial Application Ontology. The Materials System Ontology covers wide-bandgap semiconductor materials and their crystal forms, doping types, and defect types; the Device Structure Ontology covers the structural types and key parameters of power devices, RF devices, and optoelectronic devices; the Process Technology Ontology covers processes and parameter ranges such as epitaxial growth, ion implantation, etching, oxidation, annealing, and packaging; the Characterization and Testing Ontology covers methods and applicable scenarios such as XRD, TEM, SEM, AFM, PL, and Hall testing; the Equipment and Instrumentation Ontology covers the models, manufacturers, and technical parameters of process equipment and testing equipment; and the Industrial Application Ontology covers application areas and technical requirements such as new energy vehicles, photovoltaic inverters, 5G base stations, radar, and consumer electronics.

[0028] The expert review interface provides domain experts with an interactive interface for manual review and annotation, used to correct entity ambiguities, relational errors, and missing attributes. For example, "GaN" may refer to materials, devices, or technologies in different contexts. The system provides candidate standard entities based on the context, which are then confirmed by experts and added to the review status. The semantic generalization engine performs synonym unification, hierarchical expansion, and cross-contextual adaptation based on the expert ontology library, uniformly mapping "Metal-Organic Chemical Vapor Deposition," "Organic Metal Chemical Vapor Deposition," "MOCVD," and "Metal-Organic Chemical Vapor Deposition" to the standard entity "MOCVD." When the query granularity is coarse, "SiC MOSFET" can be generalized to "SiC power device" to improve recall capability.

[0029] like Figure 3 As shown, the standardized triples are written into the graph database to form a domain knowledge graph. Each entity serves as a graph node, and each relation as a graph edge. Node attributes store the entity's standard name, entity type, synonym list, original data source, data type, timestamp, and expert review status; edge attributes store the relation type, relation confidence level, data source, expert review status, and frequency of occurrence. Simultaneously, the system uses a language model pre-trained on vertical domain corpora to vectorize the standardized entities, obtaining entity semantic vectors, and constructs a vector index based on the HNSW algorithm to support fast similarity retrieval of large-scale entity vectors.

[0030] like Figure 4As shown, after a user inputs a query, the query parsing module first extracts keywords and identifies candidate entities. For example, if a user queries "Please introduce the latest process parameters for MOCVD epitaxial growth of GaN on 4H-SiC substrates and their impact on device performance," the system identifies candidate entities such as "4H-SiC," "MOCVD," "GaN epitaxial growth," and "device performance." After the candidate entities are encoded using a domain vectorization model, an approximate nearest neighbor search is performed in the HNSW index to obtain a normalized entity set; among them, "GaN epitaxial growth" can be matched with normalized entities such as "GaN thin film epitaxy," "gallium nitride epitaxy," and "GaN epitaxy."

[0031] The graph retrieval module generates query statements conforming to graph database syntax based on standardized entity sets and query intents. For example, it generates the Cypher statement: MATCH path=(a:Material{name:'4H-SiC'})-[:Substrate used for]->(b:Process{name:'MOCVD'})-[:Preparation]->(c:Material{name:'GaN'})-[:Impact]->(d:Device{name:'HEMT'}) RETURN path. The system performs multi-hop path queries in the graph database to obtain the associated paths between candidate entities and performs industry correctness verification based on the domain rule base. The domain rule base includes material and process compatibility rules, reasonable range rules for process parameters, and device causal association rules. If a path contains relationships that do not conform to the domain's technical logic, such as "MBE process is used for SiC epitaxy," "GaN growth temperature 2000°C," or "GaN substrate quality affects SiC device performance," the path is marked as a pseudo-associated path and deleted.

[0032] The rearrangement and examination module acquires a subset of the original data associated with the retention path. This subset includes patent paragraphs, paper abstracts, report chapters, and technical information snippets. The system scores the data across multiple dimensions: timeliness, source authority, technological relevance, and industry influence. Timeliness is prioritized based on data from the past three years; source authority prioritizes high-impact factor journals, patents from leading companies, and reports from authoritative institutions; technological relevance is calculated based on TF-IDF and entity co-occurrence frequency; and industry influence is assessed based on citation frequency, patent family size, and market application breadth. The comprehensive score S can be expressed as S = α·T + β·A + γ·R + δ·I, with weighting coefficients dynamically adjusted according to the query type. For process parameter queries, the weight of technological relevance is increased; for industry trend queries, the weights of timeliness and industry influence are increased.

[0033] The expert review model is trained on historical expert review data, which is labeled with three types of industry logic errors: material and process mismatch, abnormal parameter ranges, and incorrect causal relationships. The model performs secondary validation on the reordered original data, identifying and eliminating candidate data containing industry logic errors; data that passes validation is marked "credible," while data that fails validation is marked "abnormal" and excluded. The review results are also fed back to the expert review interface and the domain rule base to supplement new error examples and rule conditions, resulting in continuous enhancement of review capabilities.

[0034] like Figure 5 As shown, this implementation includes two independently operating functional flows. The left side represents the response generation and hallucination detection flow. The response generation module constructs a prompt word template containing factual constraints. This template explicitly limits the large language model to summarizing only based on trusted datasets and preserved paths, prohibiting the introduction of external knowledge, and requiring the source number to be marked for cited content. The template may include original data fragments and their source numbers, graph paths, user questions, and output format requirements. For example, graph paths are written in the prompt word format as entity-relationship-entity, enabling the model to explicitly perceive the reasoning link during generation. The large language model generates a structured response based on the prompt words, embedding data source citation identifiers within the response. The hallucination detection unit breaks down the factual statements in the response into units to be verified, comparing each unit against the trusted dataset and preserved paths for consistency. If a unit to be verified does not have a corresponding source or graph path, the unit is deleted or rewritten. The final output response retains the source identifier and reasoning path identifier. Users can trace back to the original document paragraph along the source number or view the entity association link along the path identifier.

[0035] The right side shows the incremental update process of the knowledge base, which operates independently of the response generation and illusion detection processes. The knowledge update module continuously monitors new patents, new papers, new technical information, and new industry reports, and periodically performs data collection, preprocessing, triple extraction, standardization, and graph index updates. For new entities, the system performs entity alignment and disambiguation based on entity semantic vector similarity and attribute matching. When a new entity meets the merging conditions with an existing entity, the new synonyms, sources, and relationships are merged into the existing node; otherwise, a new entity node is created and corresponding edges are established. The system also dynamically optimizes the expert ontology, domain rule base, and expert review model based on user error correction tags, satisfaction scores, and expert review results, enabling the knowledge graph to be continuously updated with technological evolution. This optimization result flows back to the graph index and rule base, continuously enhancing the knowledge base construction capability, but does not directly interfere with the output process of a single response.

[0036] In a specific application example, a user queries the latest process parameters for epitaxial growth of another material on a specific substrate using a particular process, and their impact on device performance. After identifying candidate entities and matching them with standardized entities, the system generates a graph query statement and performs multi-hop path tracing to obtain paths that conform to material compatibility and device physical relationships. The system further acquires patent paragraphs and paper abstracts associated with this path, reorders them according to multi-dimensional indicators, and uses an expert review model to exclude data containing anomalous parameters. The response generation module generates a response based on reliable data, annotating the source number of each process parameter and device performance conclusion in the response. The illusion detection unit verifies content without a source and outputs the final result. Simultaneously, the knowledge update module runs independently in the background, continuously collecting newly published patents and papers in the field, periodically extracting new triples and performing entity alignment, incorporating newly emerging process parameters and device structure information increments into the domain knowledge graph, and updating the error identification capability of the review model based on recent expert review records to ensure the timeliness and accuracy of the knowledge base for subsequent queries. The large language model generates a structured response based on the prompt words. The response embeds a data source reference identifier, such as "According to the patent CN2023XXXXXX, the optimized growth temperature for MOCVD epitaxial growth of GaN on a 4H-SiC substrate is 1050-1100°C[1]". The illusion detection unit breaks down the factual statements in the response into units to be verified, and compares each unit with the trusted dataset and the retained path for consistency. If a unit to be verified does not have a corresponding source or a corresponding map path, the unit is deleted or rewritten. The final output response retains the source identifier and the inference path identifier. Users can trace back to the original document paragraph along the source number, or view the entity association link along the path identifier.

[0037] The knowledge update module continuously monitors new patents, papers, technical information, and industry reports, and regularly performs data collection, preprocessing, triple extraction, standardization, and graph index updates. For new entities, the system performs entity alignment and disambiguation based on entity semantic vector similarity and attribute matching. When a new entity meets the merging conditions with an existing entity, the new synonyms, sources, and relationships are merged into the existing node; otherwise, a new entity node is created and corresponding edges are established. The system also dynamically optimizes the expert ontology, domain rule base, and expert review model based on user error correction tags, satisfaction scores, and expert review results, enabling the knowledge graph to be continuously updated with technological evolution.

[0038] In a specific application example, a user queries "the latest process parameters for MOCVD epitaxial growth of GaN on a 4H-SiC substrate and their impact on device performance." After identifying candidate entities and matching them with standardized entities, the system generates a query statement and performs multi-hop path tracing, obtaining the path "4H-SiC---substrate used for---MOCVD---preparation---GaN---impact---HEMT". The domain rule base confirms that this path conforms to material compatibility and device physical correlation. The system further obtains patent paragraphs and paper abstracts associated with this path, reorders them according to multi-dimensional indicators, and has an expert review model exclude data containing anomalous parameters. The response generation module generates a response based on reliable data, marking the source number of each process parameter and device performance conclusion in the response. The illusion detection unit verifies the absence of source content and outputs the final result.

[0039] In another verification example, to verify the effectiveness of the method of the present invention, the field of advanced materials was selected as the test scenario, and a comparative experiment was conducted using standard RAG, general GraphRAG, and the method of the present invention.

[0040] During the data preparation phase, 1200 patent documents and 800 journal articles in the field of advanced materials (wide bandgap semiconductor materials) were collected from the State Intellectual Property Office's patent search system, Web of Science, IEEE Xplore, and CNKI, totaling 2200 documents, spanning from January 2020 to June 2025. These were divided into a training set (1760 documents), a validation set (220 documents), and a test set (220 documents) in an 8:1:1 ratio. A test set containing 150 question-answer pairs was constructed using manual annotation by three experts with over five years of research experience in advanced materials. The test set questions covered the following types: 30 single-entity fact queries, 45 multi-hop reasoning queries, 25 concept equivalence queries, 30 industry logic verification queries, and 20 cross-entity association queries. Each question-answer pair was annotated with the standard answer, key entities, expected reasoning path, data source document number, and industry logic constraints.

[0041] During the system deployment phase, the standard RAG employs vector retrieval based on Dense Passage Retrieval, with chunk size=512 and overlap=128, combined with GPT-4 for answer generation, top-k=5, and temperature=0.1. The general GraphRAG constructs a general knowledge graph based on Wikidata, with entity types covering common concepts and relation types representing common semantic relationships. It combines vector similarity retrieval with graph traversal (maximum hop count=3) and uses GPT-4 to generate answers. The method of this invention is deployed as follows: The domain knowledge graph is constructed based on 2200 documents, containing approximately 18500 entity nodes and approximately 42000 relation edges. Six sub-ontologies cover material system ontology (3200 entities), device structure ontology (2800 entities), process technology ontology (4500 entities), characterization and testing ontology (1600 entities), equipment and instrument ontology (2400 entities), and industrial application ontology (4000 entities). The expert review model is trained based on historical review data. The training set contains 8500 review records annotated by experts, covering three main error types: material-process mismatch, abnormal parameter range, and causal relationship error. The vector index uses the HNSW index based on a domain pre-trained BERT model, with ef_construction=200 and M=16. The large language model uses GPT-4 with temperature=0.1.

[0042] During the experiment, 150 test question-answer pairs were run three times with each of the three methods, and the average value was taken. The average value was then independently scored by three domain experts. Evaluation metrics included accuracy (BERTScore F1 ≥ 0.85), recall (coverage of key information points), and illusion rate (percentage of unverifiable information).

[0043] Table 1. Performance Comparison of Three Methods on Advanced Materials Test Sets

[0044] As shown in Table 1, the method of this invention outperforms the standard RAG and general GraphRAG in all three dimensions: accuracy, recall, and hallucination rate. The accuracy is improved from 62.7% of the standard RAG to 81.6%, the recall is improved from 58.4% to 78.2%, and the hallucination rate is reduced from 28.3% to 15.8%.

[0045] Further analysis of the performance of various query types reveals the following: In multi-hop reasoning queries, the accuracy of the method presented in this invention significantly improved from 54.2% in the standard RAG to 79.3%, primarily due to the structured associations of the domain knowledge graph and the expert review mechanism ensuring path correctness. In concept equivalence queries, the accuracy increased from 61.3% in the standard RAG to 85.7% due to semantic generalization processing's ability to identify different expressions of the same technical concept. In cross-entity association queries, the recall rate improved from 48.6% in the standard RAG to 74.5%, mainly thanks to the graph path tracing's ability to capture implicit associations that text block similarity matching cannot detect. The reduction in the illusion rate is mainly attributed to three aspects: the fact-constrained prompt word template limits the large language model to generate answers based solely on the provided raw data; the illusion detection module performs consistency comparisons between the generated content and a subset of the raw data; and the expert review model filters candidate data containing industry logic errors.

[0046] To verify the effectiveness of the domain ontology in covering vertical industry concepts, a comparative experiment on the coverage of six sub-ontologies was conducted. All key entities were extracted from 150 question-answer pairs in the test set. After deduplication, a test entity set of 487 entities was obtained. Three domain experts categorized each entity into one of the following categories: material systems, device structures, process technologies, characterization and testing, equipment and instruments, industrial applications, and others (entities not belonging to the above six categories). The coverage of the test entity set by the six sub-ontologies of this invention and the general knowledge graph (Wikidata) was queried. Coverage was categorized as exact coverage (the entity has a completely matching standard entry in the ontology, including synonym mapping), semantic coverage (the entity can be mapped to related entries in the ontology through hierarchical relationships or semantic generalization), and uncovered (the entity has no corresponding entry in the ontology and cannot be associated through semantic generalization).

[0047] Table 2. Comparison of Entity Coverage of the Six Sub-Ontologies and the General Knowledge Graph in the Field of Advanced Materials

[0048] As shown in Table 2, the six sub-ontologies of this invention achieve a total coverage of 97.3% for the test entity set in the field of advanced materials, with 90.6% precise coverage and 6.8% semantic coverage. The coverage of the general knowledge graph (Wikidata) is only 33.5%, meaning this invention's coverage is approximately 2.9 times that of the general knowledge graph. The materials system ontology has the highest coverage (98.4%) because wide-bandgap semiconductor materials and their sub-concepts such as crystal form, doping type, and defect type are precisely defined in the materials system ontology, while the general knowledge graph only contains general material entries such as "silicon carbide" and "gallium nitride," lacking detailed concepts such as crystal form and doping type. The process technology ontology has a coverage of 97.3%, covering specific process flows and parameter ranges such as MOCVD, HVPE, MBE, ion implantation, etching, oxidation, annealing, and packaging. The general knowledge graph only contains general process concepts such as "chemical vapor deposition," lacking specific process entries such as MOCVD and HVPE and their parameter associations. The equipment and instrument coverage is 96.8%, encompassing specific equipment models and manufacturer information such as planetary reactors, vertical Bridgeman furnaces, and MOCVD equipment, which are lacking in the general knowledge graph. Thirteen entities, accounting for 2.7%, are not covered, primarily representing emerging technological concepts from 2024-2025, such as "gallium oxide (Ga2O3) power devices." These concepts will be gradually incorporated into the ontology through an incremental knowledge base update mechanism.

[0049] To verify the contribution of the expert review model and fact constraint prompts to the system performance, further ablation experiments were conducted. Based on the complete method of this invention, the following modules were removed and evaluated on the same 150 test question-answer pairs: (1) The expert review model was removed, while the knowledge graph, semantic generalization, fact constraint prompts and illusion detection modules were retained, but the expert review model trained based on the domain rule base and historical review data was removed, and no industry logic error filtering was performed; (2) The fact constraint prompts were removed, while the knowledge graph, semantic generalization, expert review model and illusion detection modules were retained, but the fact constraint prompts were replaced with general open prompts, and the large language model was not restricted to generating answers based solely on the provided raw data; (3) The expert review model and fact constraint prompts were removed at the same time, while only the knowledge graph and semantic generalization modules were retained, and the standard RAG generation method was adopted. The ablation experiment results are shown in Table 3.

[0050] Table 3. Comparison of Ablation Experiment Results

[0051] As shown in Table 3, removing the expert review model reduced the accuracy from 81.6% to 76.8%, while the illusion rate increased from 15.8% to 18.9%. This module primarily functions in the data reordering stage, filtering low-quality candidate data by identifying industry logic errors (such as material-process mismatches and abnormal parameter ranges). In the industry logic verification query, removing this module reduced the accuracy from 83.2% to 72.5%, a decrease of 10.7 percentage points, indicating that the expert review model plays a crucial role in ensuring the correctness of industry logic.

[0052] After removing fact constraint prompts, the accuracy dropped from 81.6% to 74.5%, while the illusion rate increased from 15.8% to 20.3%. This module effectively suppressed the illusion caused by the introduction of external knowledge by explicitly constraining the large language model to summarize only based on the provided raw data and graph path information. In single-entity fact queries, removing this module increased the illusion rate from 12.3% to 19.6%, an increase of 7.3 percentage points.

[0053] At the same time, after removing the expert review model and fact constraint prompts, the accuracy rate dropped to 69.8% and the hallucination rate rose to 24.1%. The performance degradation was greater than the simple sum of removing either module alone, indicating that the two modules have a synergistic enhancement effect in suppressing hallucinations.

[0054] like Figure 2 As shown, the system in this embodiment includes a data construction module, a standardization processing module, a graph indexing module, a query retrieval module, a rearrangement review module, a generation verification module, and a knowledge update module. The data construction module collects multi-source heterogeneous data from vertical domains and performs preprocessing and triple extraction; the standardization processing module includes a domain ontology library submodule, an expert review interface submodule, and a semantic generalization engine submodule, used to perform synonym unification, hierarchical expansion, and manual verification; the graph indexing module constructs a domain knowledge graph, completes entity vectorization encoding, and index construction; the query retrieval module performs query parsing, entity matching, graph query statement generation, multi-hop path query, and industry correctness verification; the rearrangement review module includes a multi-dimensional scoring submodule, an expert review model submodule, and a credibility labeling submodule; the generation verification module integrates raw data, graph paths, and user queries to generate traceable responses and perform illusion detection; and the knowledge update module continuously collects new data and incrementally updates the domain knowledge graph.

[0055] In a specific application example, a user queries "the latest process parameters for MOCVD epitaxial growth of GaN on a 4H-SiC substrate and their impact on device performance." After identifying candidate entities and matching them with standardized entities, the system generates a graph query statement and performs multi-hop path tracing, obtaining the path "4H-SiC—substrate used for—MOCVD—preparation—GaN—impact—HEMT". The domain rule base confirms that this path conforms to material compatibility and device physical correlation. The system further obtains patent paragraphs and paper abstracts associated with this path, reorders them according to multi-dimensional indicators, and has an expert review model exclude data containing anomalous parameters. The response generation module generates a response based on credible data, marking the source number of each process parameter and device performance conclusion in the response. The illusion detection unit verifies the absence of source content and outputs the final result.

[0056] In another validation example, the standard RAG, general GraphRAG, and the method of this invention can be compared. Standard RAG generates answers based on dense paragraph retrieval and a large language model; general GraphRAG performs graph traversal and answer generation based on a general knowledge graph; the method of this invention employs a domain ontology library, expert review mechanism, multi-hop path tracing, data reordering, fact constraint prompts, and illusion detection. The comparison selects questions of various types, including single-entity fact queries, multi-hop inference queries, concept equivalence queries, industry logic verification queries, and cross-entity association queries, which are evaluated by domain experts from three dimensions: precision, recall, and illusion rate. The method of this invention performs more stably in concept equivalence queries and multi-hop inference queries, mainly because semantic generalization processing can identify different expressions of the same technical concept, graph path tracing can discover implicit associations that are difficult to cover with text block similarity, and the expert review model and fact constraint prompts can filter industry logic errors and limit unfounded generation.

[0057] Furthermore, the effects of the expert review model and fact constraint cue words can be verified through ablation. Removing the expert review model decreases the accuracy of industry logic verification questions, making it easier for outliers and erroneous causal relationships to enter the generation stage. Removing the fact constraint cue words makes the large language model more susceptible to incorporating external knowledge, leading to an increase in unsourced expressions. Simultaneously removing both results degenerates the system into a generation method close to standard RAG, significantly increasing the risk of illusion. These results demonstrate that the expert review model and fact constraint cue words have a synergistic effect in suppressing illusions and enhancing professionalism.

[0058] Those skilled in the art should understand that the database type, language model type, vector indexing algorithm, scoring weight, and threshold in the above embodiments can all be adjusted according to the actual deployment environment. For example, the graph database can be Neo4j or other graph databases that support attribute graphs, the vector index can be HNSW or other approximate nearest neighbor indexes, and the large language model can be a general large language model or a model fine-tuned with vertical domain corpus. Any technical solution combining domain graph construction, expert review, path verification, reliable reordering, and fact constraint generation as defined in this invention should fall within the protection scope of this invention.

Claims

1. A method for enhancing the generation of domain map retrieval, characterized in that, Includes the following steps: Step S1: Collect multi-source heterogeneous documents from vertical domains. After cleaning, deduplication, and semantic segmentation, use a large language model to extract entity relation triples by domain extraction templates. Then, use the domain ontology library to perform synonym unification and hierarchical generalization on the entities in the triples. Step S2: Write the standardized triples into the graph database to construct the domain knowledge graph. Save the source, timestamp and expert review status in the node attributes. At the same time, perform domain vectorization encoding on the standardized entities and establish an approximate nearest neighbor index. Step S3: Parse the user query to obtain candidate entities, match standardized entities through vector index to generate graph query statements, perform multi-hop path tracing in the domain knowledge graph, and delete paths that do not conform to industry logic according to the domain rule base; Step S4: Obtain the original data associated with the retention path, reorder it according to timeliness, source authority, technical relevance and industry influence, and then use the expert review model trained with historical review data to identify material and process mismatch, parameter anomalies and causal errors, and output a reliable dataset. Step S5: Construct fact constraint prompts based on trusted datasets, preserved paths, and user queries; control the large language model to generate responses with source identifiers; and correct unsubstantiated content after comparing the responses with the original data.

2. The domain map retrieval enhancement generation method according to claim 1, characterized in that, The multi-source heterogeneous documents mentioned in step S1 include patent documents, academic papers, technical information, and industry reports; the cleaning includes removing noisy content, unifying encoding and text format; the deduplication uses the SimHash algorithm to identify duplicate documents; the semantic segmentation is divided into paragraphs or sentences while preserving contextual integrity; the domain extraction template limits entity types and relationship types, with entity types including at least material entities, device entities, process entities, equipment entities, parameter entities, organization entities, and person entities; and relationship types including at least material-process relationships, process-equipment relationships, equipment-parameter relationships, organization-technology relationships, technology-application relationships, and person-organization relationships. The extraction results are filtered by confidence level to form an initial set of triples.

3. The domain map retrieval enhancement generation method according to claim 1, characterized in that, The domain ontology library is built for the fields of advanced materials and intelligent manufacturing, and includes six sub-ontologies: material system ontology, device structure ontology, process technology ontology, characterization and testing ontology, equipment and instrument ontology, and industrial application ontology. In step S1, the initial triples are manually verified through an expert review interface to correct entity ambiguity, relationship labeling errors, and missing key attributes. The synonym unification maps multiple expressions of the same technical concept to standard entities, and the hierarchical generalization extends specific technical concepts to their parent concepts to adapt to queries of different granularities.

4. The domain map retrieval enhancement generation method according to claim 1, characterized in that, The node attributes in the graph database mentioned in step S2 also include entity standard name, entity type, synonym list, original data source and data type, and edge attributes include relation type, relation confidence, data source, expert review status and frequency of occurrence; the domain vectorization encoding adopts a language model that has been pre-trained on vertical domain corpus to map standardized entities into semantic vectors; the approximate nearest neighbor index is a vector index built based on the HNSW algorithm, used to perform similarity retrieval on entity semantic vectors.

5. The domain map retrieval enhancement generation method according to claim 1, characterized in that, In step S3, when parsing the user query, keywords are first extracted to form candidate entities. Then, the candidate entities are input into the domain vectorization model and retrieved in the near nearest neighbor index to obtain a standardized entity set. The graph query statement is generated by the large language model based on the standardized entity set and the query intent, and a multi-hop path query is performed in the graph database. The domain rule base includes at least material and process compatibility rules, reasonable range rules for process parameters, and causal association rules for devices, which are used to determine whether the relationship between adjacent entities in the path conforms to the vertical domain technical logic.

6. The domain map retrieval enhancement generation method according to claim 1, characterized in that, In step S4, the comprehensive score S used for re-ranking satisfies: S = α·T + β·A + γ·R + δ·I, where T is the data timeliness score, A is the source authority score, R is the technology relevance score, I is the industry influence score, and α, β, γ, and δ are weight coefficients, and α + β + γ + δ = 1. The technology relevance score is calculated based on word frequency-inverse document frequency and entity co-occurrence frequency, and the industry influence score is calculated based on citation frequency, patent family size, and market application breadth. The weight coefficients are dynamically adjusted according to the query type.

7. The enhanced generation method for domain map retrieval according to claim 1, characterized in that, The expert review model is trained based on historical expert review records, which are marked with three types of industry logic errors: material and process mismatch, abnormal parameter range, and causal relationship error. In S4, the reordered original data is verified a second time. Data that passes both the industry correctness verification and the second verification is marked as reliable, while data that fails the verification is marked as abnormal and excluded. The review results output by the expert review model are also used to update the prompt rules of the domain rule base and the expert review interface.

8. The enhanced generation method for domain map retrieval according to claim 1, characterized in that, The fact constraint prompts in step S5 are at least limited to: summarizing only based on trusted datasets and retained paths, without introducing external knowledge, marking the source number of the cited content, and writing the graph path in the prompt in the form of entity-relationship-entity; the consistency comparison breaks down the factual statements in the response into units to be verified, matches each unit with trusted datasets and retained paths, deletes or rewrites units to be verified that have no corresponding source, and retains the source identifier and reasoning path identifier in the final response.

9. The domain map retrieval enhancement generation method according to any one of claims 1 to 8, characterized in that, It also includes incremental update steps for the knowledge base: continuously monitoring new patents, new papers, new technical information, and new industry reports; regularly performing data collection, triple extraction, standardization processing, and graph index updates; aligning and disambiguating new entities with existing entities in the domain knowledge graph; merging entities when the semantic vector similarity and attribute matching degree of new entities and existing entities both meet preset conditions; otherwise, creating new entity nodes; and dynamically adjusting the domain ontology, domain rule base, and expert review model based on user error correction feedback, satisfaction feedback, and expert review results.

10. A domain map retrieval enhancement generation system, wherein the domain map retrieval enhancement generation system is applied to the domain map retrieval enhancement generation method according to any one of claims 1 to 9, characterized in that, The enhanced generation system for domain map retrieval includes: The data construction module is used to collect multi-source heterogeneous documents in the vertical domain, and perform cleaning, deduplication, semantic segmentation, triple extraction, and synonym normalization and hierarchical generalization based on the domain ontology library. The graph indexing module is used to write standardized triples into the graph database to form a domain knowledge graph, save the source, timestamp and expert review status, and perform domain vectorization encoding and index construction on standardized entities; The query and retrieval module is used to parse user queries, match standardized entities, generate graph query statements, and perform multi-hop path tracing, and delete paths that do not conform to industry logic based on the domain rule base; The reordering review module is used to reorder the original data with preserved path associations according to multidimensional indicators, and the expert review model identifies material and process mismatches, parameter anomalies and causal errors to output a reliable dataset. The generation and verification module is used to construct fact constraint prompts based on trusted datasets, preserved paths, and user queries, control the large language model to generate responses with source identifiers, and perform consistency comparisons and corrections.

Images