Intelligent processing method for standard documents in ship industry

Abstract

The embodiment of the invention provides a ship industry standard document intelligent processing method, which comprises the following steps: S1, document analysis: carrying out structured analysis and restoration on a ship industry standard document to generate a structured text supporting vectorization storage and semantic retrieval; s2, database construction: constructing a document vector index database based on the structured text to form a semantic retrieval task-oriented efficient data structure; and S3, content retrieval: performing user-oriented query, constructing a semantic retrieval mechanism, and completing high-precision matching from a natural language problem to a structured document content and result return. According to the embodiment of the invention, the document structure can be accurately restored, the content segmentation is more reasonable, the retrieval matching is more accurate, the retrieval result has more content depth, the technical path is clear, and the project landing is easy.

Description

Technical Field

[0001] This invention relates to the field of document structuring and semantic information management technology in the shipbuilding industry, and in particular to an intelligent processing method for standard documents in the shipbuilding industry. Background Technology

[0002] In the shipbuilding industry, standard documents play a crucial role in standardizing production and design processes, ensuring product quality requirements, and promoting product inspection standards. As shipbuilding enterprises transform and upgrade towards digitalization and intelligentization, the needs for the use, management, and retrieval of standard documents in ship-related fields are constantly evolving, presenting the following typical current situation: 1. Standard documents come from diverse sources and cover a wide range of business areas. National standards, industry standards, and enterprise technical documents are widely used in various stages such as product research and development, manufacturing, and testing. These documents are usually developed by authoritative institutions, industry leaders, or the companies themselves, and are in line with industry development trends. They are highly standardized and binding, and serve as important guidelines for the industry's development path.

[0003] 2. Standardized document format but opaque structure. Most standard documents are published in PDF format, with rigorous content and consistent layout, but lack semantic structural information that can be recognized by programs. Logical levels such as chapters, clauses, and appendices in these documents rely primarily on visual formatting and lack structural tags that directly support machine parsing.

[0004] 3. Customization is common among enterprises, resulting in significant version differences. In actual production, shipbuilding companies often adopt customized practices, such as adding clauses, deleting content, and optimizing expression based on general standards, according to their current situation and cutting-edge industry norms. This results in multiple versions of internal enterprise standards. While inheriting the original document structure, these enterprise standards also add personalized clauses, leading to multiple versions of documents with different structures.

[0005] 4. The number of standard documents is growing rapidly, increasing management pressure. As the industry continues to develop and improve, the number and update frequency of standard documents that shipbuilding companies need to manage are also increasing. Relying on traditional methods of manual classification, manual naming, and keyword retrieval is no longer sufficient to meet the current practical needs for high-frequency access, rapid location, and version traceability of standard documents.

[0006] 5. Diverse information usage needs and increasingly refined retrieval granularity. When using standard documents, shipbuilding companies need to browse a category of specifications as a whole, and often need to quickly locate specific clauses, terminology definitions, technical parameters, or appendices. With the diversification of business scenarios, users' document retrieval needs have expanded from "finding documents" to "finding content," "reading context," and "comparing versions."

[0007] 6. Structured management and semantic utilization are becoming trends. Faced with the increasing number and complexity of standard documents, more and more shipping companies are beginning to focus on the structured storage, semantic organization, and intelligent retrieval capabilities of standard documents to support business scenarios such as internal system integration, knowledge management platform construction, and intelligent question answering.

[0008] To address the issues of structural inconsistencies and inaccurate search results in the digitization and intelligent retrieval of standard PDF documents, existing technologies face the following five core technical challenges: 1. Difficulty in recognizing PDF document structure. Standard PDF documents lack semantic markup, making it difficult for traditional OCR or template methods to accurately reconstruct the hierarchical structure. In some cases, there are even "customized developments," resulting in a lack of universality and robustness.

[0009] 2. The converted content is disconnected from page numbers, and structural information is missing. Although Markdown files generated by open-source PDF conversion tools have the advantage of being editable, they lose meta-information such as page number positions, resulting in a mismatch between the converted content and the original document, which hinders manual retrieval and location of the original document.

[0010] 3. Lack of deep semantic understanding. Traditional search is based on keywords and cannot identify the deep semantics in paragraphs, often resulting in matching errors or omissions during queries.

[0011] 4. The reliance on a single retrieval mechanism makes it difficult to balance accuracy and efficiency. Standard documents are concise and terminologically dense, requiring both precise semantic matching and comprehensive keyword coverage. A single dense or sparse vector model cannot meet these dual requirements, making it difficult to simultaneously achieve both accuracy and efficiency.

[0012] 5. Fragmented search results lack contextual readability: Existing search results are mostly returned as independent text blocks, lacking structural aggregation and parent-child hierarchical relationships within their context. Users find it difficult to grasp the position and logical belonging of terms in the document, which can easily lead to "out-of-context" interpretation errors and reduce the accuracy of use. Summary of the Invention

[0013] In view of the above-mentioned problems existing in the prior art, the present invention provides an intelligent processing method for standard documents in the shipbuilding industry to solve the technical problems existing in the prior art, such as difficulty in recognizing the structure of PDF documents, disconnect between converted content and page numbers, missing structural information, lack of deep semantic understanding ability, single retrieval mechanism, difficulty in balancing accuracy and efficiency, fragmented retrieval results, and lack of contextual readability.

[0014] This invention provides an intelligent processing method for shipbuilding industry standard documents, comprising the following steps: Step S1, document parsing, involves performing structured parsing and restoration of the shipbuilding industry standard documents to generate structured text that supports vectorized storage and semantic retrieval; Step S2, database construction: build a document vector index library based on structured text to form an efficient data structure for semantic retrieval tasks; Step S3, content retrieval, is based on user queries. A semantic retrieval mechanism is built to achieve high-precision matching and result return from natural language questions to structured document content.

[0015] In one embodiment, step S1 includes, Step S11, OCR document recognition, performs preliminary structured parsing of standard PDF documents, obtains a chapter hierarchy structure file that initially reflects the document's chapter hierarchy, a page layout structure file that records text content and text block position information, a content block structure file containing structured text and page number information, and a chart screenshot file; Step S12, document structure restoration: use a large language model to optimize the structure of the headings in the initially generated chapter hierarchy file in order to reconstruct the accurate document parent-child hierarchy. Step S13, Content Segmentation and Metadata Annotation: Based on the optimized chapter hierarchy file, the document content is segmented into semantic units, and metadata information related to the original document structure is annotated to construct structured text that supports vectorized storage and semantic retrieval.

[0016] In one embodiment, in step S11, OCR document recognition is performed based on the MinerU tool.

[0017] In one embodiment, in step S12, the DeepSeek-Chat large language model is used to restore the document structure.

[0018] In one embodiment, step S12 includes, Step S121: Use regular expressions to extract all initial header lines; Step S122: Using a large language model and combining it with preset prompt word templates, understand the logical relationships between titles; Step S123: Output standardized objects, each object containing the original entry and the modified entry; Step S124: Generate a chapter hierarchy file containing heading levels based on the output standardized objects.

[0019] In one embodiment, step S13 includes, Step S131, title level segmentation; Step S132, page number mapping and original text restoration; Step S133: Content segmentation and fragment generation; Step S134: Structured processing of chart information; Step S135: Metadata integration and output structure construction.

[0020] In one embodiment, step S2 includes, Step S21, Vector generation and semantic data structure construction, performs semantic vectorization encoding on each structured segment in the structured text, generates dense vector fields to represent deep semantics and sparse vector fields to capture the importance of keywords, and generates a unique identifier field for each structured segment, forming a semantic data structure that integrates semantic content, hierarchical structure and contextual association. Step S22, Vector database construction and field mapping; Step S23, Index structure configuration and hybrid retrieval support: Configure the index structure for different types of vector features respectively; Step S24: Write vectors and metadata. After writing the data, the data is successfully written to disk.

[0021] In one embodiment, step S3 includes, Step S31, hybrid vector retrieval; Step S32, Reverse sorting fusion: The reverse sorting fusion strategy is used to merge and sort the mixed vector retrieval results to select the final Top-K candidate results; Step S33: Expand the context sliding window. Based on the core semantic fragment hit, and using the size of the sliding window as the range, extract several semantically continuous fragments to form a context content window. Step S34: Top-K semantic fragment backhaul, returning the merged and sorted Top-K fragments and their corresponding context window content as search results.

[0022] Compared with existing technologies, the beneficial effects of the intelligent processing method for shipbuilding industry standard documents provided in this embodiment of the invention are as follows: 1. This invention accurately restores the document structure. By combining a large language model with prompt words, it semantically identifies and modifies the heading hierarchy. Compared to traditional rule-based or template-based methods, it more accurately restores the chapter structure and clause hierarchy in standard documents, and is suitable for complex situations such as diverse heading formats and nested numbering.

[0023] 2. The content segmentation in this embodiment of the invention is more reasonable. Based on structured Markdown and text segmentation algorithms, multi-level semantic blocks are achieved, and metadata such as page numbers and parent-child level headings are integrated to form text units suitable for semantic vectorization embedding, solving the problem of difficulty in traditional paragraph splitting.

[0024] 3. The embodiments of this invention offer more accurate retrieval and matching. By introducing the BGE-M3 model to generate dense and sparse dual-channel embeddings, and combining it with the Milvus vector database to achieve a hybrid retrieval mechanism, the output results have the advantages of precise positioning and similar semantic content, overcoming the limitations of traditional keyword retrieval's weak semantic understanding capabilities.

[0025] 4. The search results in this embodiment of the invention have greater content depth. By leveraging the chunk_id chain structure and sliding window strategy, a structured semantic window containing the context of the hit fragment can be returned, enhancing content continuity and logical aggregation, avoiding "taking things out of context," and improving the user's understanding efficiency.

[0026] 5. The technical path of this invention is clear and easy to implement in engineering. The entire process uses open-source toolchains (MinerU, Milvus, etc.) and mainstream models. Specifically, the semantic vectorization part uses the BGE-M3 model; the semantic structure recognition and language understanding tasks are completed by the DeepSeek-Chat model. The overall architecture supports modular deployment and has good scalability and practical engineering value. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating an intelligent processing method for shipbuilding industry standard documents, provided as an embodiment of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0030] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0031] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application, which have the features described in the claims and are therefore all within the scope of protection defined herein.

[0032] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0033] Specific embodiments of this application are described below with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to ascertain the true intent based on the user's historical operations, and to avoid unnecessary or redundant details that would obscure this application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely serve as the basis and representative basis for the claims to teach those skilled in the art to use this application in various ways with substantially any suitable detailed structure.

[0034] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0035] The principles and features of the present invention are described below with reference to the accompanying drawings. The embodiments described are for illustrative purposes only and are not intended to limit the scope of the invention. The following description, in conjunction with... Figure 1 The preferred embodiments of the present invention will be described in further detail below: This invention belongs to the field of document structuring and semantic information management technology in the shipbuilding industry, specifically involving key technologies such as document structure recognition and parsing, semantic vectorization and embedding, text meta-information annotation, vector index construction, and database integration. The embodiments of this invention are mainly applied to scenarios such as automatic structure extraction and semantic content querying of text materials such as standard documents, normative documents, and instruction manuals, belonging to the interdisciplinary technology direction of intelligent document processing and AI information retrieval.

[0036] like Figure 1 As shown, this invention provides an intelligent processing method for standard documents in the shipbuilding industry. Specifically, it is an intelligent processing method for standard documents that integrates large language models, structural recognition, semantic vectorization, and content retrieval. The overall process includes five stages: OCR recognition, structural restoration, content segmentation, vector generation and storage, and semantic retrieval. The above process can be categorized into three main modules: document parsing, database construction, and content retrieval, forming a closed-loop processing system from unstructured documents to structured document question answering. Specific steps include... 1. Document parsing S1. OCR Document Recognition. This step uses the MinerU tool to perform preliminary structured parsing of standard PDF documents. MinerU integrates an OCR engine and layout analysis module, capable of recognizing structural elements such as text areas, header lines, tables, and appendices on a page. Therefore, the open-source MinerU tool is used to convert the PDF document format. The main output after conversion includes: • Markdown file (full.md): Initially shows the document's chapter structure; • Page layout structure file (layout.json): Records text content and text block position information, etc.; • Content block structure file (content_list.json): Contains structured text, page number information, etc.; • Chart screenshot file (JPG): Captures tables and image areas from the original PDF file.

[0037] Since the original PDF often lacks explicit semantic hierarchy markers, the conversion results at this stage still have problems such as disordered title hierarchy, misaligned page numbers, and page breaks. Therefore, this result is only used as the input basis for subsequent structural optimization and semantic understanding. To restore the complete document structure and assign metadata to text blocks, it is necessary to use the content_list.json file.

[0038] S2. Document Structure Restoration. This step uses the DeepSeek-Chat large language model to optimize the structure of the headings in the initially generated Markdown file in order to reconstruct an accurate document parent-child hierarchy. The specific steps are as follows: • Use regular expressions (such as ^#{1,6}) to extract all initial header lines; • Utilize the DeepSeek large language model and combine it with preset prompt word templates to understand the logical relationships between titles (such as chapter numbers, chart titles, etc.); The model outputs a standardized JSON array, with each object containing the original entry and the modified entry; Generates a Markdown file containing heading levels from a JSON array.

[0039] Although the generated Markdown files have a hierarchical heading structure, the format itself still has certain limitations, mainly including: • Lacking structured fields, it is not suitable for direct use in database management and semantic retrieval; • Page numbers and location information from the original document are not retained, making precise location impossible; • Non-text information such as charts and graphs is not structured.

[0040] The above issues will be addressed in the next phase of content segmentation and metadata annotation.

[0041] S3. Content Segmentation and Metadata Annotation. Based on the Markdown file with optimized heading structure, this step performs semantic unit segmentation of the document content and annotates metadata information related to the original document structure (such as page numbers, parent level heading paths, document type, document path, etc.) to construct structured text that can be used for vectorized storage and semantic retrieval. The specific processing flow is as follows: 1) Heading hierarchy segmentation. Using the MarkdownHeaderTextSplitter function from the LangChain package, the improved Markdown file is segmented into several content units according to the heading hierarchy. Each unit's page_content contains the complete paragraph text, and metadata records the corresponding parent heading path information, thus constructing a complete hierarchical structure information stored in JSON format; 2) Page number mapping and original text restoration. Align the above segmentation results with the content block structure JSON file output by MinerU. Use the BM25 algorithm to calculate the semantic similarity between each content unit and the JSON text block. Successfully matched content blocks will be assigned corresponding page number information in the metadata. Non-heading text blocks within the same page are merged, and the above JSON file containing metadata is reconstructed. 3) Content Segmentation and Fragment Generation. To adapt to subsequent semantic embedding and vector retrieval requirements, the `RecursiveCharacterTextSplitter` function in the `LangChain` package is further called to perform secondary segmentation of the merged content based on the number of characters. The length of each fragment is controlled within 500 characters, and a 50-character overlap range is set to enhance semantic continuity and retrieval robustness. The segmentation results replace the paragraph text in the original JSON, while inheriting the corresponding metadata, as the final semantic fragment set; 4) Structured processing of chart information. For structured table areas, a custom script is used to parse the table screenshots generated by MinerU and generate table data in standardized JSON format; for image areas, semantic information of the images is extracted based on the results of manual inspection, and caption description text is generated and associated with the original image. 5) Metadata Integration and Output Structure Construction. All table structures and image information are embedded into a JSON content structure containing metadata, at the placeholder for the chart / figure encoding. This JSON structure supports flexible expansion of metadata fields, including text content, parent level title path, page number information, document type, document path, and whether it is a figure / table, etc., which can be directly used for subsequent semantic vector generation and database storage.

[0042] 2. Database Construction S1. Semantic Vector Generation and Storage. After completing content segmentation and metadata annotation, this step builds a document vector index library based on structured JSON files, realizing dual-channel vector encoding of semantic fragments, context association identification, index structure configuration, and vector database writing, forming an efficient data structure for semantic retrieval tasks.

[0043] 1) Vector generation and UUID identification construction. The BGE-M3 model is invoked to perform semantic vectorization encoding on the `page_content` field of each structured fragment, generating: • Dense vector fields (dense_vector) are used to represent deep semantics; • Sparse vector field (sparse_vector) is used to capture keyword importance.

[0044] To achieve traceability and sequential management of document fragments, the UUID library in Python is used to generate a unique identifier field (chunk_id) for each structured fragment. At the same time, each fragment also has prev_chunk_id and next_chunk_id fields, which are used to point to the chunk_id field of the context block, thereby establishing a context connection relationship between text blocks and generating a bidirectional context chain structure.

[0045] All generated vectors and UUID information will be directly added to the original JSON file, forming a semantic data structure that integrates semantic content, hierarchical structure, and contextual association.

[0046] 2) Milvus Vector Database Construction and Field Mapping. Based on the updated JSON file, a vector database Collection is constructed using Milvus, with the following fields defined: • Primary key and structure control fields: id (primary key), chunk_id, prev_chunk_id, next_chunk_id; • Content and metadata fields: page_content (text content); header[1-6] (title name and corresponding level), page_num (page number), file_type (document type), file_path (document location); • Vector fields: dense_vector (dense vector), sparse_vector (sparse vector).

[0047] This collection supports flexible field filtering, hierarchical navigation, and pagination.

[0048] 3) Index structure configuration and hybrid search support. Index structures can be configured separately for different types of vector features to improve search efficiency and coverage: • Dense Vector Index: Given that the actual application scenarios targeted by this embodiment of the invention are small to medium-sized, semantically dense document sets such as industrial standard documents and internal enterprise specifications, and the query targets are mainly high-precision clause matching and contextual consistency, the IVF_FLAT (inverted file index + full-precision scan) index structure provided by Milvus is adopted. The parameter setting nlist represents the number of cluster centers. During retrieval, full-precision query is only performed on the few subclusters with the closest vector distance. Combined with the inner product (IP) similarity measure, the query efficiency and accuracy are balanced. • Sparse Vector Index: Given that standard documents contain a large number of structured terms and technical terms, retrieval relies not only on semantic understanding but also on precise keyword targeting. Therefore, Milvus's SPARSE_INVERTED_INDEX sparse inverted index structure, combined with the DAAT_MAXSCORE query strategy, is used to focus on retrieving technical terms. Compared to traditional dense retrieval, this indexing mechanism can significantly improve recall when handling sparse embeddings, achieving efficient keyword matching. • Text block unique identifier index: After accurately locating the standard document, it is necessary to combine a large amount of context to view the relationship between the clauses. Therefore, the Trie dictionary tree index provided by Milvus is configured to support block-level accurate search and debugging location. Combined with the doubly linked list structure built by the prev_chunk_id and next_chunk_id fields, the context relationship tracing and structural navigation between semantic fragments can be realized.

[0049] 4) Vector and metadata writing. Each structured JSON record is written as a data object to the Milvus database, including: • Embedded vectors (dense + sparse); • Information on the links before and after the primary key and chunk_id; • Text fragments and their structured meta-information.

[0050] After the data is written, a flush operation is performed to complete the data persistence to disk, ensuring that the system can load and use it immediately in the subsequent semantic retrieval process.

[0051] 3. Content retrieval S1. Database Retrieval and Matching. This step addresses user queries by constructing a semantic retrieval mechanism that integrates dense and sparse vectors, supports sliding windows and contextual aggregation, and implements multi-dimensional ranking. This mechanism enables high-precision matching and result return from natural language questions to structured document content. The overall process includes four core steps: 1) Hybrid Vector Search. After the user inputs a query, the system first uses the BGE-M3 model to generate two types of embedding representations for the query: dense vectors and sparse vectors, capturing deep semantics and keyword weights respectively. Then, a hybrid vector search is performed through the Milvus vector database, using these two types of vectors as search conditions for a joint query. Based on the set number of returned results, the system recalls the most relevant candidate sets of semantic fragments in the original document, preparing for subsequent output of high-precision results and aggregation of contextual content. This process is performed simultaneously based on the IVF_FLAT index and the SPARSE_INVERTED_INDEX index, effectively improving semantic coverage and hit rate.

[0052] 2) Reciprocal Rank Fusion Strategy (RRF Ranking Mechanism). To improve the relevance and stability of the ranking results, an RRF (Reciprocal Rank Fusion) strategy is used to fuse and rank the results of dense and sparse vector retrieval. This strategy generates a comprehensive score based on the position of the fragments retrieved in each channel within its local ranking, using a reciprocal weighting method, and then selects the Top-K final candidate results. This fusion ranking avoids single-channel ranking bias and improves the overall semantic retrieval quality.

[0053] 3) Contextual Sliding Window Expansion. To enhance the contextual completeness and readability of search results, based on the core semantic fragments hit, and using the sliding window size as the range, several semantically contiguous fragments are extracted using the fragment identifier `chunk_id` and its `prev_chunk_id` and `next_chunk_id` fields maintained in the database to form a contextual content window. This sliding window mechanism enables the aggregated display of logically structured fragments, facilitating user understanding of contextual semantics, avoiding misinterpretations, and providing corpus support for further search terms and cited content.

[0054] 4) Top-K semantic fragment return. Finally, the merged and sorted Top-K fragments and their corresponding context window content are returned as search results. Each result contains complete metadata information such as text content, parent and child level titles, original page number, file address, and file type, making it easy for users to locate the original PDF standard document.

[0055] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its spirit and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.

Claims

1. A method for intelligent processing of shipbuilding industry standard documents, characterized in that, Includes the following steps: Step S1, document parsing, involves performing structured parsing and restoration of the shipbuilding industry standard documents to generate structured text that supports vectorized storage and semantic retrieval; Step S2, database construction: build a document vector index library based on structured text to form an efficient data structure for semantic retrieval tasks; Step S3, content retrieval, is based on user queries. A semantic retrieval mechanism is built to achieve high-precision matching and result return from natural language questions to structured document content.

2. The intelligent processing method for shipbuilding industry standard documents according to claim 1, characterized in that: Step S1 includes, Step S11, OCR document recognition, performs preliminary structured parsing of standard PDF documents, obtains a chapter hierarchy structure file that initially reflects the document's chapter hierarchy, a page layout structure file that records text content and text block position information, a content block structure file containing structured text and page number information, and a chart screenshot file; Step S12, document structure restoration: use a large language model to optimize the structure of the headings in the initially generated chapter hierarchy file in order to reconstruct the accurate document parent-child hierarchy. Step S13, Content Segmentation and Metadata Annotation: Based on the optimized chapter hierarchy file, the document content is segmented into semantic units, and metadata information related to the original document structure is annotated to construct structured text that supports vectorized storage and semantic retrieval.

3. The intelligent processing method for shipbuilding industry standard documents according to claim 2, characterized in that: In step S11, OCR document recognition is performed based on the MinerU tool.

4. The intelligent processing method for shipbuilding industry standard documents according to claim 2, characterized in that: In step S12, the DeepSeek-Chat large language model is used to restore the document structure.

5. The intelligent processing method for shipbuilding industry standard documents according to claim 2, characterized in that: Step S12 includes, Step S121: Use regular expressions to extract all initial header lines; Step S122: Using a large language model and combining it with preset prompt word templates, understand the logical relationships between titles; Step S123: Output standardized objects, each object containing the original entry and the modified entry; Step S124: Generate a chapter hierarchy file containing heading levels based on the output standardized objects.

6. The intelligent processing method for shipbuilding industry standard documents according to claim 2, characterized in that: Step S13 includes, Step S131, title level segmentation; Step S132, page number mapping and original text restoration; Step S133: Content segmentation and fragment generation; Step S134: Structured processing of chart information; Step S135: Metadata integration and output structure construction.

7. The intelligent processing method for shipbuilding industry standard documents according to claim 1, characterized in that: Step S2 includes, Step S21, Vector generation and semantic data structure construction, performs semantic vectorization encoding on each structured segment in the structured text, generates dense vector fields to represent deep semantics and sparse vector fields to capture the importance of keywords, and generates a unique identifier field for each structured segment, forming a semantic data structure that integrates semantic content, hierarchical structure and contextual association. Step S22, Vector database construction and field mapping; Step S23, Index structure configuration and hybrid retrieval support: Configure the index structure for different types of vector features respectively; Step S24: Write vectors and metadata. After writing the data, the data is successfully written to disk.

8. The intelligent processing method for shipbuilding industry standard documents according to claim 1, characterized in that: Step S3 includes, Step S31, hybrid vector retrieval; Step S32, Reverse sorting fusion: The reverse sorting fusion strategy is used to merge and sort the mixed vector retrieval results to select the final Top-K candidate results; Step S33: Expand the context sliding window. Based on the core semantic fragment hit, and using the size of the sliding window as the range, extract several semantically continuous fragments to form a context content window. Step S34: Top-K semantic fragment backhaul, returning the merged and sorted Top-K fragments and their corresponding context window content as search results.

Images