Manufacturing knowledge augmented retrieval method and system fusing knowledge graph and large model

Abstract

The application provides a manufacturing industry knowledge enhanced retrieval method and system fusing a knowledge graph and a large model, and relates to the technical field of information retrieval.The manufacturing industry original document set is divided into original text segments in the application, entity information and relationship information are extracted by using a large language model, a manufacturing industry knowledge graph, a graph structure index and a vector index are constructed, the user query is analyzed for intention, low-level keywords and high-level keywords are formed, the trigger result of the graph retrieval process is determined accordingly, the corresponding graph retrieval mode is executed to form a graph retrieval result, the vector retrieval result is formed based on the vector index, and the vector retrieval result and the graph retrieval result are processed in layers to obtain a candidate retrieval result, and the candidate retrieval result is taken as a structured input to constrain the large language model to generate a retrieval answer.The application is beneficial to improving the coverage, semantic consistency and answer reliability of manufacturing industry complex knowledge retrieval.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, and in particular to a method and system for enhancing knowledge retrieval in the manufacturing industry by integrating knowledge graphs and large models. Background Technology

[0002] In the context of knowledge management in manufacturing, technical data is often distributed across different business systems, document repositories, and experience records, with inconsistencies in the time of data creation, expression methods, and content granularity. The same technical object may appear in different documents under different names, abbreviations, or descriptive statements, and there are implicit connections between different pieces of knowledge, such as process adaptation, parameter constraints, and application scenarios. Existing retrieval methods mostly rely on keyword matching or manually maintained knowledge bases, and the search results tend to remain at the level of single text fragments, making it difficult to present the continuous connections between knowledge.

[0003] With the development of search-enhanced generation technologies, document retrieval based on semantic similarity has been used in professional question-answering systems. However, in knowledge-intensive scenarios such as manufacturing, semantic retrieval tends to weaken the precise targeting of model numbers, parameters, and proprietary names; keyword retrieval, on the other hand, struggles to cover implicit relationships across documents. When users raise questions involving a mix of specific objects and abstract technical requirements, existing methods are prone to issues such as fragmented retrieval content, semantic misalignment, and unclear basis for answers.

[0004] Therefore, in the process of knowledge retrieval in the manufacturing industry, how to improve the retrieval coverage, semantic consistency, and credibility of answers for complex questions, given the dispersed nature of professional knowledge, significant differences in expression, and deep layers of association, has become a technical problem that needs to be solved in the intelligent application of knowledge in the manufacturing industry. Summary of the Invention

[0005] To address the challenges of fragmented knowledge resources, deep hierarchical relationships between knowledge objects in the manufacturing industry, and the tendency of existing retrieval methods to suffer from scattered recall content, semantic misalignment, and unclear answer basis when dealing with the intertwining of concrete entities and abstract topics, this invention proposes a manufacturing knowledge enhancement retrieval method and system that integrates knowledge graphs and large-scale models. This method divides the original manufacturing document set into original text fragments and establishes a correspondence between these fragments and their source document IDs and original text fragment IDs. It transforms the original manufacturing document set into a manufacturing knowledge graph, graph structure index, and vector index. By combining low-level and high-level keywords from user queries, the method determines the triggering results for the graph retrieval process. Furthermore, it performs hierarchical collaborative processing of graph retrieval results and vector retrieval results, enabling candidate retrieval results to simultaneously reflect entity relationships, semantically similar content, keyword matching content, and source-related content. This, in turn, constrains the large-scale language model to generate retrieval answers.

[0006] First, a knowledge-enhanced retrieval method for the manufacturing industry that integrates knowledge graphs and large-scale models is proposed, which includes:

[0007] S1. Divide the original document set of the manufacturing industry into original text fragments, and establish the correspondence between the original text fragments and their source document IDs and original text fragment IDs. Use a large language model to extract entity information and relational information from the original text fragments, and construct a manufacturing industry knowledge graph, graph structure index and vector index.

[0008] S2. Perform intent parsing on user queries, identify low-level keywords pointing to specific entities and high-level keywords pointing to abstract topics, and generate keyword parsing results;

[0009] S3. Based on the keyword parsing results, determine the triggering result of the graph retrieval process, and execute the corresponding graph retrieval mode based on the graph structure index according to the triggering result of the graph retrieval process to form the graph retrieval result;

[0010] S4. Based on the vector index, vector retrieval results are generated, and according to the triggering results of the graph retrieval process, the vector retrieval results and the graph retrieval results are processed in a hierarchical and collaborative manner to obtain candidate retrieval results;

[0011] S5. Use candidate search results as structured input to constrain the large language model to generate search answers.

[0012] Optionally, in step S1, the entity information includes entity ID, entity type, entity name, and semantic description. The entity type includes product, enterprise, technology, industry, and personnel. The relationship information includes relationship ID, source entity, target entity, relationship keywords, and relationship semantic description. Entity nodes are generated based on the entity information, relationship edges are generated based on the relationship information, and a graph structure index is established from the entity nodes and relationship edges.

[0013] Optionally, in step S1, the extracted entity names are used as the mentions to be aligned. A dual encoder architecture is used to calculate the semantic similarity between the mentions to be aligned and the existing entities in the manufacturing knowledge graph, and standard entity candidates are selected based on the semantic similarity. The attribute similarity between the mentions to be aligned and the standard entity candidates is calculated based on the enterprise background and technical specifications. When the attribute similarity meets the preset aggregation threshold, the mentions to be aligned are merged into the entity nodes corresponding to the standard entity candidates. When there are conflicting values ​​for the same attribute, the conflicting values ​​are converted into complementary semantic descriptions using a large language model.

[0014] Optionally, in step S2, low-level keywords are determined based on the equipment model, company name, equipment number, or specific entity name found in the user query; high-level keywords are determined based on the technical solutions, industry standards, applicable scenarios, or abstract concept categories found in the user query; the keyword parsing results include keyword text, keyword type, and query intent.

[0015] Optionally, in step S3, if the keyword parsing results contain only low-level keywords, the local search mode is triggered; if the keyword parsing results contain only high-level keywords, the global search mode is triggered; if the keyword parsing results contain both low-level and high-level keywords, the local search mode and the global search mode are triggered sequentially.

[0016] Optionally, in step S3, when the local retrieval mode is triggered, low-level keywords are mapped and matched to core entity nodes in the graph structure index, and a breadth-first search with a maximum depth of two layers is performed starting from the core entity nodes to obtain the corresponding neighbor nodes and relation edges. The graph retrieval results in the local retrieval mode are formed by the core entity nodes, neighbor nodes, and relation edges. When the global retrieval mode is triggered, a multi-hop path exploration with a maximum number of three hops is performed with the abstract concept corresponding to the high-level keywords as the retrieval topic to form the graph retrieval results in the global retrieval mode.

[0017] Optionally, in step S4, forming vector retrieval results based on vector indexing includes: forming a semantically similar candidate set using dense vector retrieval, forming a keyword matching candidate set using BM25-based sparse vector retrieval, and fusing them using a Round-Robin alternating strategy to form a vector fusion candidate set; when the local retrieval mode is triggered, original text fragments that meet at least one association condition are selected from the semantically similar candidate set. The association condition includes: the entity information extracted from the original text fragment contains the same entity ID as the core entity node or neighbor node; the relationship path corresponding to the relationship information extracted from the original text fragment is the same as the relationship path corresponding to the relationship edge; or the original text fragment has the same source document ID or the same original text fragment ID as the graph retrieval result in the local retrieval mode. The selected original text fragments are then compared with the local retrieval mode. The graph search results are fused to form a local fusion candidate set. When the global search mode is triggered, the graph search results under the global search mode are used as the global candidate set. When the keyword parsing results contain only low-level keywords and trigger the local search mode, the local fusion candidate set and the vector fusion candidate set are determined as candidate sets to be merged. When the keyword parsing results contain only high-level keywords and trigger the global search mode, the global candidate set and the vector fusion candidate set are determined as candidate sets to be merged. When the keyword parsing results contain both low-level and high-level keywords and trigger the local search mode and the global search mode in sequence, the local fusion candidate set, the global candidate set, and the vector fusion candidate set are determined as candidate sets to be merged. The original text fragments, entities, and relations in the candidate sets to be merged are deduplicated and then reordered to obtain candidate search results.

[0018] Optionally, in step S4, the dense vector retrieval to form a semantically similar candidate set includes cross-granularity association recall; the vector index includes the original text fragment granularity index and the source document granularity index; the cross-granularity association recall includes first obtaining fine-grained candidates based on the original text fragment granularity index, and then obtaining coarse-grained candidates based on the source document ID corresponding to the fine-grained candidates, or first obtaining coarse-grained candidates based on the source document granularity index, and then obtaining fine-grained candidates based on the source document ID corresponding to the coarse-grained candidates; and merging the fine-grained candidates and coarse-grained candidates into the semantically similar candidate set.

[0019] Optionally, in step S5, the candidate search results include entities, relations, original text fragments and their corresponding source information; the source information includes entity ID, source document ID, original text fragment ID and relation path; the domain-specific prompt template constrains the large language model to annotate the source document ID in the technical parameter table item, and annotate the entity ID or relation path in the key discussion, generating a search answer with source annotation.

[0020] Secondly, a knowledge-enhanced retrieval system for manufacturing that integrates knowledge graphs and large models is proposed, which includes a graph construction module, an intent parsing module, a graph retrieval module, a hierarchical collaborative processing module, and an answer generation module.

[0021] The graph construction module is used to divide the original document set of the manufacturing industry into original text fragments, establish the correspondence between the original text fragments and their source document IDs and original text fragment IDs, and use a large language model to extract entity information and relational information from the original text fragments to construct a manufacturing industry knowledge graph, graph structure index and vector index.

[0022] The intent parsing module is used to parse the intent of user queries, identify low-level keywords that point to specific entities and high-level keywords that point to abstract topics, and form keyword parsing results;

[0023] The graph retrieval module is used to determine the triggering result of the graph retrieval process based on the keyword parsing results, and execute the corresponding graph retrieval mode based on the graph structure index according to the triggering result of the graph retrieval process to form the graph retrieval results;

[0024] The hierarchical collaborative processing module is used to generate vector retrieval results based on vector indexes, and to perform hierarchical collaborative processing on vector retrieval results and graph retrieval results according to the triggering results of the graph retrieval process to obtain candidate retrieval results;

[0025] The answer generation module is used to take candidate search results as structured input and constrain the large language model to generate search answers.

[0026] The beneficial effects of this invention compared to the prior art are:

[0027] This invention divides the original document set of the manufacturing industry into original text fragments, establishes the correspondence between the original text fragments and their source document IDs and original text fragment IDs, and then extracts entity information and relationship information from the original text fragments. This can transform the scattered manufacturing industry knowledge data into a searchable, associative and traceable graph structure and vector index, which is beneficial to reducing the entity relationship break caused by the dispersion of manufacturing industry knowledge data.

[0028] This invention identifies low-level and high-level keywords and determines the triggering results of the graph retrieval process based on the keyword parsing results. It can adapt to specific entity queries and abstract topic queries respectively, so that the system can form corresponding retrieval paths when processing different types of queries such as equipment models, company names, technical solutions and applicable scenarios, which is conducive to improving the retrieval coverage of complex manufacturing problems.

[0029] This invention generates vector retrieval results based on vector indexes and performs hierarchical collaborative processing of vector retrieval results and graph retrieval results according to the triggering results of the graph retrieval process. It can incorporate graph relationships, semantically similar content, keyword matching content and cross-granularity related content into the same candidate retrieval result formation process, which helps to reduce semantic bias and long-tail parameter missed recall caused by single-path retrieval.

[0030] This invention uses candidate retrieval results as structured input to constrain a large language model to generate retrieval answers. This enables the large language model to form answer content within the scope of candidate entities, relations, original text fragments, and source information, which is beneficial to improving the verifiability and credibility of knowledge-based question answering in the manufacturing industry. Attached Figure Description

[0031] Figure 1 A schematic diagram of the overall process for a knowledge-enhanced retrieval method in the manufacturing industry that integrates knowledge graphs and large models.

[0032] Figure 2 This is an example diagram of the manufacturing industry. Detailed Implementation

[0033] This embodiment provides a manufacturing knowledge enhancement retrieval method and system that integrates knowledge graphs and large language models. It is used for structured management, related retrieval, fusion recall, re-ranking optimization, and retrieval answer generation of manufacturing knowledge documents. Manufacturing knowledge documents include product manuals, equipment manuals, process specifications, industry standards, company information, technical articles, application cases, and expert experience records. Manufacturing knowledge documents typically contain both concrete entity information such as equipment models, company names, and product parameters, and abstract thematic information such as process plans, industry standards, applicable scenarios, and technical routes. Ordinary keyword retrieval mainly relies on literal matching, vector retrieval mainly relies on semantic similarity, and large language models require stable candidate criteria to generate answers. This embodiment utilizes the collaborative work of a manufacturing knowledge graph, graph structure index, vector index, and large language model to ensure that the retrieval process covers concrete entity neighborhood relationships, long-distance paths of abstract themes, semantically similar fragments, and keyword matching fragments, and organizes the candidate retrieval results into structured input that can be read by the large language model.

[0034] like Figure 1 As shown, the processing flow of this embodiment includes knowledge source access, multi-dimensional attribute extraction of entities and relationships, semantic disambiguation and entity alignment, heterogeneous storage, local retrieval, global retrieval, hybrid retrieval, re-ranking, and large language model to generate answers. After document parsing, the knowledge source yields original text fragments, and a correspondence is established between the original text fragments and their source document IDs and original text fragment IDs. After the original text fragments enter the large language model, entity and relationship information is output. The entity and relationship information are written into the manufacturing knowledge graph after semantic disambiguation and entity alignment. The manufacturing knowledge graph forms a graph structure index, and the original text fragments form a vector index. After a user query enters the system, the system first performs keyword parsing, then determines the triggering result of the graph retrieval process based on low-level and high-level keywords, and triggers local retrieval mode, global retrieval mode, or a combination of local and global retrieval modes sequentially. Subsequently, the graph retrieval results and vector retrieval results are processed hierarchically and collaboratively, and after processing by the re-ranking model, candidate retrieval results are formed, and the large language model generates the retrieval answer.

[0035] The processing flow of this embodiment can also be summarized as follows: Divide the original document set of the manufacturing industry into original text fragments and establish a correspondence between the original text fragments and their source document IDs and original text fragment IDs; extract entity information and relational information from the original text fragments using a large language model; construct a manufacturing industry knowledge graph, graph structure index, and vector index based on the entity information and relational information; perform intent parsing on user queries to identify low-level keywords pointing to specific entities and high-level keywords pointing to abstract topics; determine the triggering result of the graph retrieval process based on the keyword parsing results; execute the corresponding graph retrieval mode according to the triggering result; generate vector retrieval results based on the vector index; perform hierarchical collaborative processing of the vector retrieval results and graph retrieval results based on the triggering result of the graph retrieval process; and use the obtained candidate retrieval results as structured input to constrain the large language model to generate retrieval answers.

[0036] The original document set of the manufacturing industry is denoted as:

[0037] ;

[0038] in, Represents the original document set of the manufacturing industry. Indicates the first A fine-grained block of text or a fragment of the original text.

[0039] The original documents in the manufacturing industry are first numbered according to their source, with each original document corresponding to a source document ID. Then, they are divided into text fragments based on heading level, paragraph boundaries, semantic integrity, and length window, with each text fragment corresponding to an original text fragment ID. For product parameter tables, process steps, and industry standard clauses, the system prioritizes keeping the content of the same parameter table or clause within the same text fragment to reduce retrieval bias caused by the separation of parameter fields from their explanatory content.

[0040] In step S1, the original manufacturing document set is divided into text fragments, and a correspondence is established between the text fragments and their source document IDs and original text fragment IDs. Then, a large language model is used to extract entity information and relational information from the text fragments. Entity information is used to represent node objects in the manufacturing knowledge graph, and relational information is used to represent the connection semantics between node objects. Through this processing, unstructured or semi-structured manufacturing documents are converted into structured data required for graph construction.

[0041] Entity information includes entity ID, entity type, entity name, and semantic description. Entity ID is used to distinguish different entities; entity type is used to represent the category of the entity in the manufacturing knowledge graph; entity name comes from product name, company name, technology name, industry name, or personnel name in the original text fragment; semantic description is used to record the entity's function, purpose, technical background, or application scenario.

[0042] For example, as shown in Table 1, entity types include products, enterprises, technologies, industries, and people. Products represent equipment, components, materials, or industrial products in a manufacturing system; enterprises represent manufacturers, suppliers, integrators, or service providers; technologies represent process methods, control algorithms, processing schemes, or testing methods; industries represent application areas, industry backgrounds, or usage scenarios; and people represent technical roles related to technical documentation, project implementation, or professional experience. Through these five entity types, object information in manufacturing documents can be entered into a knowledge graph according to a unified entity framework.

[0043] Table 1. Five Core Entity Relationship Models for Manufacturing Knowledge Graphs

[0044] Entity type meaning Typical examples Key attribute examples Map Function product Physical objects in a manufacturing system Industrial robots, CNC machine tools, bearings Functional description, technical specifications, and application areas Core object node enterprise Manufacturer or supplier entity Siemens, Sany Heavy Industry Enterprise attributes and business scope Industry Relationship Modeling technology Process method or technical solution Laser cutting technology, intelligent control algorithms Technology type, principle, and scope of application Expressing manufacturing process knowledge industry Application areas or industry background Aerospace manufacturing, automotive industry Industry description and application scenarios Constructing semantic context personnel Key technical roles Engineers, experts Position and research direction Building key industry objects

[0045] Relationship information includes relationship ID, source entity, target entity, relationship keywords, and relationship semantic description. The relationship ID distinguishes different relationships; the source and target entities represent the start and end points of the relationship edge, respectively; the relationship keywords indicate the semantic connection type between entities; and the relationship semantic description records the contextual meaning of the relationship within the original text. For example, a product entity and a company entity can form a "production" or "supply" relationship; a product entity and a technology entity can form a "applicable" or "support" relationship; and a technology entity and an industry entity can form a "applied to" or "serves" relationship.

[0046] Entity nodes are generated based on entity information, and relationship edges are generated based on relationship information. A graph structure index is then built from the entity nodes and relationship edges. The graph structure index supports entity localization, neighborhood traversal, and multi-hop path exploration. Simultaneously, the system builds a vector index based on original text fragments. This vector index supports semantic similarity retrieval and sparse vector retrieval. The graph structure index focuses on entity relationship paths, while the vector index focuses on original text fragment recall. Together, they form candidate retrieval results in subsequent hierarchical collaborative processing.

[0047] When the large language model extracts entity and relation information, it uses a preset extraction prompt template. The prompt template includes entity type ranges, relation field formats, source retention rules, and output format constraints. After inputting the original text fragment into the large language model, it outputs entity candidates and relation candidates. In this embodiment, the process of the large language model performing entity recognition, relation candidate recognition, source entity and target entity matching, and relation semantic description generation on the original text fragment is denoted as the entity recognition and relation extraction function. For relation candidates, source entities and target entities are matched based on the current entity candidates or the existing entity set. If a match is successful, the relation candidate is written into the manufacturing knowledge graph to form relation edges with endpoints and semantic descriptions.

[0048] In this embodiment, a large language model is used to extract multi-dimensional attribute information of entities and relationships from the original manufacturing document set. Divided into fine-grained text blocks Utilizing large language models to perform entity recognition and relation extraction functions Generate a knowledge graph for the manufacturing industry Manufacturing Knowledge Graph It can be represented as:

[0049] ;

[0050] ;

[0051] in, and These represent the node set and edge set after preliminary deduplication and normalization, respectively. and These represent the initial node set and initial edge set extracted from each original text segment, respectively. This indicates deduplication. This indicates the processing of attribute normalization and semantic description reconstruction. Indicates from the first The corresponding initial entity nodes and initial relation edges were extracted from the original text fragments. This represents the union operation, used to combine the initial entity nodes and initial relation edges corresponding to each original text segment into a single set. and .

[0052] Entity nodes are generated based on entity information, and relationship edges are generated based on relationship information. A graph structure index is then built using these entity nodes and relationship edges. The graph structure index supports entity localization, neighborhood traversal, and multi-hop path exploration in graph databases. A vector index is built based on the semantic vectors of the original text fragments. The vector index supports semantic similarity retrieval and sparse vector retrieval. The graph structure index focuses on entity relationship paths, while the vector index focuses on original text fragment recall; the two work together in subsequent retrieval stages.

[0053] like Figure 2 As shown, the manufacturing knowledge graph example includes multiple different categories of entity nodes and relationship connections. Figure 2 The nodes in the system correspond to entities such as products, enterprises, technologies, industries, or personnel, and the connections between nodes correspond to relationships such as applicable, production, association, belonging, and support. Figure 2 It also illustrates a radial association structure formed around entities, which supports neighborhood traversal centered on an entity and cross-entity retrieval along relational paths. Node attribute information includes node ID, entity description, source field, and entity identifier.

[0054] When constructing a manufacturing knowledge graph, different original text fragments may contain different spellings of the same entity, such as full model number, abbreviation, company alias, or technical abbreviation. To reduce duplicate nodes, step S1 also includes entity alignment and conflict handling. The system uses the extracted entity names as mentions to be aligned, employs a dual encoder architecture to calculate the semantic similarity between the mentions to be aligned and existing entities in the manufacturing knowledge graph, and selects standard entity candidates from the existing entities based on semantic similarity.

[0055] The dual-encoder architecture includes a mention encoder to be aligned and an entity encoder. The mention encoder maps the mentions to be aligned and their context to a high-dimensional vector space; the entity encoder maps the entity names, entity types, and semantic descriptions of existing entities in the manufacturing knowledge graph to the same high-dimensional vector space; semantic similarity... It can be represented as:

[0056] ;

[0057] in, Indicates the mentions to be aligned. Representing the entities in the graph, and These represent the feature extraction operators for the mentions to be aligned and the entities in the graph, respectively.

[0058] The dot product of the two entities in the vector space is calculated, and standard entity candidates are selected from existing entities based on semantic similarity, thereby improving the entity consistency of complex industrial models, abbreviations, and aliases. Standard entity candidates are used as merging targets for mentions to be aligned, and their entity consistency with the mentions to be aligned is further verified in subsequent attribute similarity judgments.

[0059] After identifying standard entity candidates, the attribute similarity between the mentions to be aligned and the standard entity candidates is further calculated. Attribute similarity is calculated based on corporate background and technical specifications; corporate background includes manufacturer, supplier, business scope, industry sector, and industry relationships; technical specifications include product model, functional description, performance parameters, applicable processes, application scenarios, and technical principles; the attribute set is denoted as... , This represents the set of attribute parameters from the company background and technical specifications corresponding to the mention to be aligned, or the set of attribute parameters from the company background and technical specifications corresponding to the standard entity candidate. For the attribute set corresponding to the mention to be aligned... The set of attributes corresponding to standard entity candidates Using the cosine similarity function Calculate attribute similarity.

[0060] Node merging decision function It can be defined as:

[0061] ;

[0062] in, This indicates the candidate entities corresponding to the mentions to be aligned. This indicates that there are already entities in the manufacturing knowledge graph. and These represent the sets of multidimensional attributes corresponding to the two. This represents the function for calculating cosine similarity. This indicates the preset aggregation threshold.

[0063] In one implementation, The similarity is calculated by comparing the corresponding attribute fields separately and then obtaining the comprehensive similarity. Specifically, the corresponding fields such as manufacturer, supplier, business scope, industry field, industrial relationship, product model, functional description, performance parameters, applicable process, application scenario and technical principle are first vectorized and cosine similarity calculated. Then, the similarity of each field involved in the calculation is weighted according to the preset field weight or the same field weight and the average is calculated to obtain the attribute similarity between the attribute sets.

[0064] In one embodiment, ;when When the time comes, the mentions to be aligned are merged into the entity nodes corresponding to the standard entity candidates, and description reconstruction is performed; when When this happens, the mentions to be aligned are treated as new entity candidates or enter the verification process; this determination method can merge mentions of similar entities with similar names and consistent corporate backgrounds and technical specifications into the same entity node.

[0065] When conflicting values ​​exist for the same attribute, a large language model is used to convert these conflicting values ​​into complementary semantic descriptions. When documents from different sources provide different descriptions of the applicable scenarios, processing ranges, or technical specifications for the same equipment, the system converts the differences from different sources into complementary semantic descriptions under different operating conditions, materials, or applicable boundaries. This process reduces contradictory coverage in the manufacturing knowledge graph while preserving the differences in operating conditions across different sources.

[0066] In step S2, the user query is analyzed to determine intent, identifying low-level and high-level keywords to generate keyword analysis results. Low-level keywords point to specific entities, while high-level keywords point to abstract topics. Low-level keywords are determined based on equipment models, company names, equipment serial numbers, or specific entity names found in the user query. High-level keywords are determined based on technical solutions, industry standards, applicable scenarios, or abstract concept categories found in the user query. The keyword analysis results include keyword text, keyword type, and query intent.

[0067] In one example, a user query is "Which process solutions are suitable for 828D in automotive parts processing?". The system identifies "828D" as a low-level keyword because it corresponds to a specific equipment model. The system identifies "automotive parts processing" and "process solutions" as high-level keywords because they correspond to application scenarios and abstract technical solutions, respectively. The keyword parsing results include the keyword text, keyword type, and query intent. The query intent is to retrieve process solutions for a specific product in a specific industry scenario.

[0068] In step S3, the triggering result of the graph retrieval process is determined based on the keyword parsing results. Then, the corresponding graph retrieval mode is executed based on the graph structure index according to the triggering result, forming the graph retrieval results. Graph retrieval modes include local retrieval mode and global retrieval mode. Local retrieval mode is used for neighborhood retrieval around specific entities, while global retrieval mode is used for path exploration around abstract topics.

[0069] When the keyword parsing results only contain low-level keywords, the local retrieval mode is triggered. This trigger indicates that the user query mainly targets a specific product, company, equipment number, or entity name. The system needs to locate the core entity node in the manufacturing knowledge graph and obtain the neighborhood relationships around the core entity node.

[0070] When the keyword parsing results contain only high-level keywords, the global search mode is triggered. This trigger indicates that the user's query mainly targets technical solutions, industry standards, applicable scenarios, or abstract concepts, requiring multi-hop path exploration within the manufacturing knowledge graph to obtain cross-entity and cross-document related knowledge.

[0071] When the keyword parsing results contain both low-level and high-level keywords, the local and global search modes are triggered sequentially. This triggering indicates that the user's query requires both locating specific entities and expanding the search around an abstract theme. A local search is first performed around the core entity nodes corresponding to the low-level keywords, followed by a global search around the abstract concepts corresponding to the high-level keywords. This allows both specific object information and theme-related information to simultaneously enter subsequent hierarchical collaborative processing.

[0072] When the local search mode is triggered, low-level keywords are mapped and matched to core entity nodes in the graph structure index. Mapping matching includes exact entity name matching, model alias matching, and semantic similarity matching. If multiple nodes simultaneously match low-level keywords, the core entity node is selected based on entity type, query intent, and contextual semantics. After determining the core entity node, the system uses it as a starting point and performs a breadth-first search to traverse the neighborhood with a maximum depth of two levels.

[0073] In a neighborhood traversal with a maximum depth of two layers, the first layer includes enterprise nodes, technology nodes, industry nodes, product nodes, or personnel nodes directly connected to the core entity node, as well as corresponding relationship edges. The second layer includes nodes and relationship edges further connected to the first-layer neighbor nodes. The core entity node, neighbor nodes, and relationship edges are obtained, and these are used to form the graph retrieval results in the local retrieval mode. The graph retrieval results in the local retrieval mode are used to answer questions about the attributes, associated objects, and applicable scope of a specific model, specific equipment, specific enterprise, or specific technology.

[0074] The global search mode is used for multi-hop path exploration around abstract topics. Using the abstract concepts corresponding to high-level keywords as the search topics, it performs multi-hop path exploration with a maximum of three hops within the manufacturing knowledge graph. The abstract concepts corresponding to high-level keywords can be located using graph fields such as entity names, relational keywords, semantic descriptions, industry scenarios, and technical solutions to pinpoint relevant path entry points. Long-distance relationships across documents and entities are mined along semantic relationship paths to form graph search results in the global search mode.

[0075] For example, for the abstract topic of "automotive parts processing solutions," exploration can be conducted along either the path of "industry scenario—processing difficulties—technical solutions—adaptive equipment" or the path of "material properties—process requirements—control technology—application cases." The graph retrieval results in the global retrieval mode are used to form cross-document knowledge chains related to the abstract topic, thereby supporting systematic solution responses and long-distance semantic association retrieval.

[0076] In step S4, vector retrieval results are generated based on the vector index. Then, according to the triggering results of the graph retrieval process, the vector retrieval results and the graph retrieval results undergo hierarchical collaborative processing to obtain candidate retrieval results. The vector retrieval results are used to supplement the graph retrieval process in recalling textual evidence and semantically similar content, while the graph retrieval results are used to provide entity relationships and path structures. The hierarchical collaborative processing combines, deduplicates, and reorders the two results according to the triggering results.

[0077] Vector retrieval results based on vector indexing include dense vector retrieval, sparse vector retrieval, and vector-side fusion processing. Dense vector retrieval is used to form a semantically similar candidate set. Dense vector retrieval is used to capture the deep semantic similarity between user queries and original text fragments. The user query is converted into a query vector, and original text fragment vectors with high similarity to the query vector are retrieved from the vector index to obtain the semantically similar candidate set. The semantically similar candidate set is suitable for recalling original text fragments with different expressions but similar semantics.

[0078] A sparse vector retrieval based on BM25 is used to form a keyword matching candidate set. BM25 retrieval is used to capture precise entities, models, proper nouns, and parameter fields in the manufacturing industry. For short model numbers or proper nouns such as "828D" and "CM390", sparse vector retrieval can improve recall stability based on term matching relationships and reduce missed recalls caused by the semantic generalization of short model numbers by dense vector retrieval.

[0079] The semantically similar candidate set and the keyword matching candidate set are fused using a Round-Robin alternating strategy to form a vector fusion candidate set as the vector retrieval result. The Round-Robin alternating strategy means that candidate content is selected from different candidate sets in turn, and duplicates are removed based on the original text fragment identifier or the source document identifier. If the same original text fragment exists in both the semantically similar candidate set and the keyword matching candidate set, the candidate with the better overall ranking is retained, and its origin from multiple recall paths is recorded.

[0080] When the local retrieval mode is triggered, original text fragments that satisfy at least one association condition are selected from the semantic similarity candidate set. Association conditions include: the entity information extracted from the original text fragment contains the same entity ID as the core entity node or neighboring nodes; the relationship path corresponding to the relationship information extracted from the original text fragment is the same as the relationship path corresponding to the relationship edge; the original text fragment has the same source document ID or the same original text fragment ID as the graph retrieval result in the local retrieval mode. The selected original text fragments are then fused with the graph retrieval result in the local retrieval mode to form a local fusion candidate set. The graph retrieval result in the local retrieval mode provides core entities, neighboring nodes, and relationship edges, while the selected original text fragments provide textual evidence related to entities and relationships. After fusion, the local fusion candidate set can simultaneously reflect the entity relationship structure and the original text evidence.

[0081] When the global search mode is triggered, the graph search results in the global search mode are used as the global candidate set. The global candidate set includes semantic paths, path entities, and relational information related to the abstract concepts corresponding to high-level keywords. This global candidate set is used to support long-distance association analysis across documents and entities.

[0082] When the keyword parsing results contain only low-level keywords, triggering the local retrieval mode, the local fusion candidate set and the vector fusion candidate set are identified as the candidate set to be merged. This candidate set to be merged includes local graph relationships formed around core entity nodes, associated original text fragments, and semantically similar content and keyword matching content recalled by the vector side.

[0083] When the global search mode is triggered because the keyword parsing results only contain high-level keywords, the global candidate set and the vector fusion candidate set are determined as the candidate set to be merged. This candidate set to be merged includes multi-hop path information, path entities, relationship information formed around the abstract topic, as well as semantically similar content and keyword matching content recalled by the vector side.

[0084] When the keyword parsing results contain both low-level and high-level keywords, triggering both local and global search modes sequentially, the local fusion candidate set, global candidate set, and vector fusion candidate set are determined as the candidate set to be merged. This candidate set to be merged simultaneously covers specific entity neighborhood relationships, abstract topic path relationships, and vector search results, and is used to support the generation of comprehensive answers in scenarios where specific objects and abstract topics are intertwined.

[0085] The system deduplicates original text fragments, entities, and relations separately within the candidate set to be merged. For original text fragments, deduplication is performed based on the source document ID or original text fragment ID corresponding to the fragment; for entities, deduplication is performed based on the entity ID corresponding to the entity; and for relations, deduplication is performed based on the relation path corresponding to the relation. Through categorized deduplication, the system can retain the different roles of entities, relations, and original text fragments in the candidate search results, forming the complete candidate data required for structured input.

[0086] After deduplication, a re-ranking model is used to rank the candidate sets to be merged based on relevance, yielding candidate retrieval results. The re-ranking model can be a cross-encoder model. The system pairs the user query with the candidate objects to be ranked and inputs this pairing into the re-ranking model. It calculates the deep matching score between the query and the candidate objects and determines the output order of the candidate retrieval results based on the matching score. Candidate retrieval results include entities, relations, original text fragments, source information, and ranking information.

[0087] In step S4, the dense vector retrieval forming of the semantic similarity candidate set also includes cross-granularity association recall. The vector index includes the original text fragment granularity index and the source document granularity index. The original text fragment granularity index is built based on the original text fragments and is used to locate fine-grained textual evidence; the source document granularity index is built based on the set of original text fragments corresponding to the source document and is used to locate the coarse-grained semantic background of the document.

[0088] Cross-granularity association-based recall includes two recall paths. The first recall path first obtains fine-grained candidates based on the original text fragment granularity index, and then obtains coarse-grained candidates based on the source document ID corresponding to the fine-grained candidates. This path is suitable for scenarios where user queries contain specific model numbers, parameters, or component names. The system first locates the relevant original text fragments, and then traces back to the source document level to obtain the complete background.

[0089] The second recall path involves first obtaining coarse-grained candidates based on the source document granularity index, and then obtaining fine-grained candidates based on the source document IDs corresponding to the coarse-grained candidates. This path is suitable for scenarios where user queries contain technical solutions, application scenarios, or industry standards. The system first locates the relevant source documents, and then retrieves the corresponding original text fragments under those source documents. The system merges the fine-grained and coarse-grained candidates into a semantic similarity candidate set, ensuring that the semantic similarity candidate set simultaneously covers text fragment evidence and document-level background information.

[0090] In step S5, the system uses candidate search results as structured input to constrain the large language model to generate search answers. The structured input includes query fields, candidate entity fields, candidate relation fields, candidate original text fragment fields, source fields, and ranking fields. The query fields record user queries and keyword parsing results; the candidate entity fields record entity ID, entity name, entity type, and semantic description; the candidate relation fields record source entity, target entity, relation keywords, and relation semantic description; the candidate original text fragment fields record original text fragment content, source document ID, and original text fragment ID; the source fields record the source binding relationships between entities, relations, original text fragments, and relation paths; and the ranking fields record the ranking of candidate objects in the candidate search results.

[0091] Candidate search results include entities, relationships, original text fragments, and their corresponding source information. Source information includes entity ID, source document ID, original text fragment ID, and relationship path. Entity ID identifies entity nodes in the manufacturing knowledge graph; source document ID identifies original manufacturing documents; original text fragment ID identifies original text fragments; and relationship path represents the connection sequence from source entities to target entities in the manufacturing knowledge graph.

[0092] Domain-specific prompt templates constrain the generation of search answers by the large language model. These templates include role constraints, answer structure constraints, source annotation constraints, and basis scope constraints. Role constraints enable the large language model to organize its answers according to the manufacturing knowledge question-and-answer scenario; answer structure constraints ensure that the answers are output according to parameters, applicable scenarios, process suggestions, relationship paths, or basis descriptions; source annotation constraints enable the large language model to annotate the source document ID in the technical parameter table and the entity ID or relationship path in the key arguments; and basis scope constraints enable the large language model to generate answer content based on candidate search results.

[0093] In a series of validation tests on manufacturing knowledge retrieval, based on the same set of manufacturing knowledge documents and the same query samples, the retrieval process using manufacturing knowledge graphs, local retrieval modes, global retrieval modes, and vector fusion candidate sets was evaluated. The validation results show that the context entity recall rate was improved by more than 56.4%, which can reduce semantic bias caused by document slicing and semantic generalization in general retrieval. In the answer generation stage, by constraining the large language model with source information, the answer quality was improved by more than 12.83%, and the fidelity index was improved by more than 37.70%. These results demonstrate that this embodiment can improve recall coverage and answer verifiability in complex manufacturing knowledge retrieval.

[0094] This embodiment also provides a knowledge-enhanced retrieval system for the manufacturing industry that integrates knowledge graphs and large-scale models. The system includes a knowledge graph construction module, an intent parsing module, a knowledge graph retrieval module, a hierarchical collaborative processing module, and an answer generation module. Each module can be deployed on a server, cloud computing platform, enterprise knowledge management platform, or industrial intelligent question-and-answer platform, or it can be integrated into a manufacturing enterprise knowledge base system as a software service.

[0095] The graph construction module is used to divide the original document set of the manufacturing industry into original text fragments and establish the correspondence between the original text fragments and their source document IDs and original text fragment IDs. It uses a large language model to extract entity and relation information from the original text fragments, constructing a manufacturing industry knowledge graph, graph structure index, and vector index. The graph construction module can include a document segmentation unit, an entity and relation extraction unit, an entity alignment unit, a conflict resolution unit, a graph database writing unit, and a vector index generation unit. The document segmentation unit divides the original document set of the manufacturing industry into original text fragments and establishes the correspondence between the original text fragments and their source document IDs and original text fragment IDs; the entity and relation extraction unit calls the large language model to output entity and relation information; the entity alignment unit performs dual encoder semantic similarity calculation and attribute similarity judgment; the conflict resolution unit converts conflict values ​​into complementary semantic descriptions; the graph database writing unit generates entity nodes and relation edges; and the vector index generation unit establishes original text fragment granular indexes and source document granular indexes.

[0096] The intent parsing module is used to parse user queries, identifying low-level keywords pointing to specific entities and high-level keywords pointing to abstract topics, thus generating keyword parsing results. The intent parsing module may include a query cleaning unit, a keyword recognition unit, and an intent classification unit. The query cleaning unit standardizes user queries; the keyword recognition unit identifies equipment models, company names, equipment serial numbers, specific entity names, technical solutions, industry standards, applicable scenarios, and abstract concept categories; the intent classification unit outputs keyword text, keyword types, and query intent.

[0097] The graph retrieval module determines the triggering result of the graph retrieval process based on the keyword parsing results, and executes the corresponding graph retrieval mode based on the graph structure index according to the triggering result, forming the graph retrieval results. The graph retrieval module can include an entity mapping unit, a local traversal unit, and a global path exploration unit. The entity mapping unit maps and matches low-level keywords to core entity nodes; the local traversal unit performs a breadth-first search with a maximum depth of two layers to obtain the corresponding neighbor nodes and relation edges, and forms the graph retrieval results in the local retrieval mode from the core entity nodes, neighbor nodes, and relation edges; the global path exploration unit performs multi-hop path exploration with a maximum number of hops of three.

[0098] The hierarchical collaborative processing module is used to generate vector retrieval results based on vector indexes, and, according to the triggering results of the graph retrieval process, to perform hierarchical collaborative processing on the vector retrieval results and the graph retrieval results to obtain candidate retrieval results. The hierarchical collaborative processing module may include a dense vector recall unit, a sparse vector recall unit, a cross-granularity association recall unit, a vector fusion unit, a local fusion unit, a global candidate unit, a deduplication unit, and a re-ranking unit. The dense vector recall unit forms a semantically similar candidate set; the sparse vector recall unit forms a keyword matching candidate set; the cross-granularity association recall unit forms fine-grained and coarse-grained candidates; the vector fusion unit uses a round-robin alternating strategy to fuse the semantically similar candidate set and the keyword matching candidate set to form a vector fusion candidate set as the vector retrieval result; the local fusion unit selects original text fragments from the semantically similar candidate set that meet at least one association condition, and fuses the selected original text fragments with the graph retrieval results in the local retrieval mode to form a local fusion candidate set. The association conditions include: the entity information extracted from the original text fragment contains the same entity ID as the core entity node or neighbor node; the relationship path corresponding to the relationship information extracted from the original text fragment is the same as the relationship path corresponding to the relationship edge; the original text fragment has the same source document ID or the same original text fragment ID as the graph retrieval result in the local retrieval mode; the global candidate unit forms a global candidate set; the deduplication unit deduplicates the original text fragments, entities, and relationships respectively; and the re-ranking unit outputs the candidate retrieval results.

[0099] The answer generation module takes candidate search results as structured input and constrains the large language model to generate search answers. The answer generation module can include a structured input generation unit, a prompt template invocation unit, and an answer output unit. The structured input generation unit organizes entities, relations, original text fragments, source information, and ranking information from the candidate search results into a unified input format; the prompt template invocation unit provides domain-specific prompt templates to the large language model; and the answer output unit outputs the search answer, and, in scenarios requiring source annotation, outputs a search answer with source annotation.

[0100] Through this embodiment, unstructured knowledge in original manufacturing documents can be transformed into searchable objects such as entities, relationships, and original text fragments; low-level keywords and high-level keywords can drive specific entity neighborhood retrieval and abstract topic path retrieval, respectively; local retrieval mode can improve the recall completeness of specific entity attributes and neighborhood relationships; global retrieval mode can improve the recall coverage of cross-document relationship chains under abstract topics; vector retrieval can improve the recall stability of semantically similar content and keyword matching content; hierarchical collaborative processing can improve the relevance and completeness of candidate retrieval results; and structured input can constrain the large language model to generate retrieval answers based on candidate retrieval results.

[0101] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A knowledge-enhanced retrieval method for manufacturing industry that integrates knowledge graphs and large-scale models, characterized in that, include: S1. Divide the original document set of the manufacturing industry into original text fragments, and establish the correspondence between the original text fragments and their source document IDs and original text fragment IDs. Use a large language model to extract entity information and relational information from the original text fragments, and construct a manufacturing industry knowledge graph, graph structure index and vector index. S2. Perform intent parsing on user queries, identify low-level keywords pointing to specific entities and high-level keywords pointing to abstract topics, and generate keyword parsing results; S3. Based on the keyword parsing results, determine the triggering result of the graph retrieval process, and execute the corresponding graph retrieval mode based on the graph structure index according to the triggering result of the graph retrieval process to form the graph retrieval result; S4. Based on the vector index, vector retrieval results are generated, and according to the triggering results of the graph retrieval process, the vector retrieval results and the graph retrieval results are processed in a hierarchical and collaborative manner to obtain candidate retrieval results; S5. Use candidate search results as structured input to constrain the large language model to generate search answers.

2. The method according to claim 1, characterized in that, In step S1, entity information includes entity ID, entity type, entity name, and semantic description. Entity type includes product, enterprise, technology, industry, and personnel. Relationship information includes relationship ID, source entity, target entity, relationship keywords, and relationship semantic description. Entity nodes are generated based on entity information, relationship edges are generated based on relationship information, and a graph structure index is established from entity nodes and relationship edges.

3. The method according to claim 2, characterized in that, In step S1, the extracted entity names are used as the mentions to be aligned. A dual encoder architecture is used to calculate the semantic similarity between the mentions to be aligned and the existing entities in the manufacturing knowledge graph, and standard entity candidates are selected based on the semantic similarity. The attribute similarity between the mentions to be aligned and the standard entity candidates is calculated based on the enterprise background and technical specifications. When the attribute similarity meets the preset aggregation threshold, the mentions to be aligned are merged into the entity nodes corresponding to the standard entity candidates. When there are conflicting values ​​for the same attribute, the conflicting values ​​are converted into complementary semantic descriptions using a large language model.

4. The method according to claim 1, characterized in that, In step S2, low-level keywords are determined based on the equipment model, company name, equipment number, or specific entity name found in the user's query; high-level keywords are determined based on the technical solutions, industry standards, applicable scenarios, or abstract concept categories found in the user's query; the keyword parsing results include keyword text, keyword type, and query intent.

5. The method according to claim 4, characterized in that, In step S3, if the keyword parsing results contain only low-level keywords, the local search mode is triggered; if the keyword parsing results contain only high-level keywords, the global search mode is triggered; if the keyword parsing results contain both low-level and high-level keywords, the local search mode and the global search mode are triggered sequentially.

6. The method according to claim 5, characterized in that, In step S3, when the local retrieval mode is triggered, low-level keywords are mapped and matched to core entity nodes in the graph structure index. A breadth-first search with a maximum depth of two layers is performed starting from the core entity node to obtain the corresponding neighbor nodes and relation edges. The graph retrieval results in the local retrieval mode are formed by the core entity nodes, neighbor nodes, and relation edges. When the global retrieval mode is triggered, a multi-hop path exploration with a maximum number of three hops is performed with the abstract concept corresponding to the high-level keywords as the retrieval topic to form the graph retrieval results in the global retrieval mode.

7. The method according to claim 6, characterized in that, In step S4, forming vector retrieval results based on vector indexing includes: forming a semantically similar candidate set using dense vector retrieval, forming a keyword matching candidate set using BM25-based sparse vector retrieval, and fusing them using a Round-Robin alternating strategy to form a vector fusion candidate set; when the local retrieval mode is triggered, original text fragments that meet at least one association condition are selected from the semantically similar candidate set. The association conditions include: the entity information extracted from the original text fragment contains the same entity ID as the core entity node or neighbor node; the relationship path corresponding to the relationship information extracted from the original text fragment is the same as the relationship path corresponding to the relationship edge; or the original text fragment has the same source document ID or the same original text fragment ID as the graph retrieval result in the local retrieval mode. The selected original text fragments are then compared with the results in the local retrieval mode. The graph search results are merged to form a local fusion candidate set. When the global search mode is triggered, the graph search results in the global search mode are used as the global candidate set. When the keyword parsing results contain only low-level keywords and trigger the local search mode, the local fusion candidate set and the vector fusion candidate set are determined as candidate sets to be merged. When the keyword parsing results contain only high-level keywords and trigger the global search mode, the global candidate set and the vector fusion candidate set are determined as candidate sets to be merged. When the keyword parsing results contain both low-level and high-level keywords and trigger the local search mode and the global search mode in sequence, the local fusion candidate set, the global candidate set, and the vector fusion candidate set are determined as candidate sets to be merged. The original text fragments, entities, and relations in the candidate sets to be merged are deduplicated and then reordered to obtain the candidate search results.

8. The method according to claim 7, characterized in that, In step S4, dense vector retrieval forms a semantically similar candidate set, including cross-granularity association recall; the vector index includes original text fragment granularity index and source document granularity index; Cross-granularity association recall includes first obtaining fine-grained candidates based on the original text fragment granularity index, and then obtaining coarse-grained candidates based on the source document ID corresponding to the fine-grained candidates, or first obtaining coarse-grained candidates based on the source document granularity index, and then obtaining fine-grained candidates based on the source document ID corresponding to the coarse-grained candidates; and merging the fine-grained candidates and coarse-grained candidates into a semantic similarity candidate set.

9. The method according to claim 1, characterized in that, In step S5, the candidate search results include entities, relations, original text fragments and their corresponding source information; the source information includes entity ID, source document ID, original text fragment ID and relation path; by constraining the large language model with domain-specific prompt templates, the source document ID is marked in the technical parameter table items, and the entity ID or relation path is marked in the key arguments, generating a search answer with source annotation.

10. A knowledge-enhanced retrieval system for manufacturing that integrates knowledge graphs and large-scale models, characterized in that, It includes a graph construction module, an intent parsing module, a graph retrieval module, a hierarchical collaborative processing module, and an answer generation module; The graph construction module is used to divide the original document set of the manufacturing industry into original text fragments, establish the correspondence between the original text fragments and their source document IDs and original text fragment IDs, and use a large language model to extract entity information and relational information from the original text fragments to construct a manufacturing industry knowledge graph, graph structure index and vector index. The intent parsing module is used to parse the intent of user queries, identify low-level keywords that point to specific entities and high-level keywords that point to abstract topics, and form keyword parsing results; The graph retrieval module is used to determine the triggering result of the graph retrieval process based on the keyword parsing results, and execute the corresponding graph retrieval mode based on the graph structure index according to the triggering result of the graph retrieval process to form the graph retrieval results; The hierarchical collaborative processing module is used to generate vector retrieval results based on vector indexes, and to perform hierarchical collaborative processing on vector retrieval results and graph retrieval results according to the triggering results of the graph retrieval process to obtain candidate retrieval results; The answer generation module is used to take candidate search results as structured input and constrain the large language model to generate search answers.

Images