What this category’s data looks like
Shipbuilding business process-related data primarily originates from shipbuilding process specifications, classification society compliance documents, production work orders, inspection records, and equipment operation and maintenance manuals. Data is updated per individual construction project batch, and is synchronized with process adjustments and inspection progress during the project cycle. Most documents are a mix of structured and semi-structured content, including fields such as process steps, parameter thresholds, inspection nodes, and responsible departments. Units use standard engineering units such as millimeters, tons, and man-hours. Some documents also include exclusive identifiers such as classification society numbers and project numbers.
What constraints these characteristics impose on vector models and indexing
Shipbuilding business data is managed per project batch and contains a large number of engineering technical terms and structured fields. This requires indexes to support metadata filtering by project dimension, to avoid irrelevant recall across projects. Documents exist in a mix of structured and semi-structured forms, so field association information must be retained during chunking to prevent the loss of binding relationships between parameters and corresponding steps. The length of documents for a single project varies widely, with some process specifications reaching tens of thousands of characters, so long text chunking logic must be supported. Engineering units and standard specification numbers are key retrieval features, so vector models must be able to accurately encode these professional identifiers to avoid recall bias.
How to configure the settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-paragraph information volume of shipbuilding process documents, avoiding splitting that breaks complete process steps |
| Overlap Length | 100–150 characters | Retains process connection information between adjacent chunks, preventing key inspection nodes from being split across two chunks |
| Similarity Threshold | 0.72–0.80 | Filters low-match engineering term recall results, meeting the retrieval accuracy requirements of the professional field |
| Single Search Limit | Top 8 entries | Controls the number of documents returned per round of retrieval, avoiding long contexts exceeding the model’s processing limits |
defaultToken | Calibrated based on actual testing | Adapts to the token encoding length of shipbuilding technical terms, ensuring professional fields are fully encoded |
maxToken | 4096 | Matches the maximum token limit of the mainstream built-in embedding models in FastGPT v4.17.0, adapting to encoding requirements after long document chunking |
Three common configuration mistakes
- Symptom: When creating a knowledge base, progress gets stuck at the "Build Index" stage, with no clear error log or logs showing "Chunking failed". Cause: The chunking configuration is not adapted to shipbuilding long documents, and the character count per chunk exceeds the model’s encoding limit, causing index construction to interrupt.
- Symptom: After using a newly added engineering domain embedding model to complete indexing, search tests return "Model call failed" or "Vector dimension mismatch". Cause:
defaultTokenormaxTokenare not configured to parameters adapted to this embedding model, resulting in inconsistent vector encoding and storage dimensions. - Symptom: After importing CAD-exported PDFs or process Excel files for shipbuilding, retrieval results fail to cover equipment parameters or CAD annotation text within tables. Cause: The text extraction plugin for the corresponding file type is not enabled, resulting in structured tables and non-standard text content not being parsed correctly.
How to confirm the configuration is correct
- Import a single shipbuilding process specification document, run a chunking preview test, and confirm that the split text does not break key process steps or inspection nodes.
- After adding an engineering domain embedding model, configure the
defaultTokenandmaxTokenparameters, run a single-text vector encoding test, and confirm that the returned vector dimension matches the official published parameters of the model. - Retrieve inspection records for a specified project, and confirm that the returned results include metadata fields such as the project number and responsible department from the original document.
- Import multi-format files (PDF, Excel, CAD-exported text), run a batch retrieval test, and confirm that results cover professional parameter content from different file types.
The values provided here are common starting points, and should be evaluated against samples specific to the deployment use case.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.