What the data for this category looks like
The knowledge base data for shipbuilding mainly comes from classification society official specifications, shipyard process documents, shipowner custom requirement documents, and inspection records. Data updates are triggered on demand when new shipbuilding projects are initiated or industry specifications are revised, with no fixed cycle. Document structures fall into three categories: compliance documents with chapter and clause number fields, process documents with process steps, parameters, and equipment model fields, and inspection documents with batch, ship classification number, and inspection date fields. Units include industry-specific measurement standards such as millimeters, megapascals, welding current amperes, and navigation hours.
What constraints these characteristics impose on the knowledge base retrieval and recall step
The fixed clause structure of compliance documents requires retrieval to accurately match chapters and clauses to avoid mixing irrelevant results. Precise parameters in process documents require matching both semantics and numerical values; basic semantic retrieval cannot cover parameter-related needs. Industry-specific terms such as classification society abbreviations and ship type codes account for a large proportion, and expanded queries must be added to cover the full retrieval scope. Individual process documents are lengthy, and field context must be retained during chunking, otherwise the association between parameters and process steps will be lost. Since data updates have no fixed cycle, on-demand synchronization of source data must be supported to prevent retrieval results from lagging behind the latest specifications.
How to set the configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Shipbuilding documents contain a large number of precise parameters and fixed clauses. A threshold that is too low will mix irrelevant process data, while a threshold that is too high will miss matching compliance clauses |
limit | Top 8 entries | Content in individual shipborne process documents is concentrated. Too many recalled entries will exceed the context window, while too few will fail to fully cover process requirements |
searchMode | hybrid | Combines semantic retrieval and keyword retrieval to adapt to fixed terms and process parameters in specification texts |
usingReRank | Enabled | Multiple versions of expressions exist in classification society specifications; re-ranking can filter results that are semantically similar but do not meet compliance requirements |
datasetSearchUsingExtensionQuery | Enabled | Adds expanded queries for general shipbuilding terms to cover industry-specific abbreviations such as DNV, ABS |
Three common mistakes
- Phenomenon: Some documents are written directly as full text without generating question-answer pairs. Cause: The documents contain a large number of non-standardized shipborne parameter fragments, and exclusive fields are not associated during chunking, causing the question-answer pair generation logic to fail to match.
- Phenomenon: The error "insufficient_quota Current group upstream load is saturated" is returned. Cause: A reasonable value for the
limitparameter is not configured, and a large number of recall requests exceed the upper resource carrying limit of the upstream system. - Phenomenon: The model return result does not match the knowledge base content. Cause:
datasetSearchUsingExtensionQueryis not enabled, and shipborne term expanded queries are not added, causing the model to fail to associate exclusive content in the knowledge base.
How to confirm the configuration is set correctly
- Upload a typical shipborne welding process card, and check whether the retrieval results include corresponding welding current and voltage parameters.
- Trigger a batch document synchronization, and confirm that updated specification files can be retrieved within 1 hour.
- Enter a query containing a classification society abbreviation, and check whether the results preferentially return compliance documents from the corresponding classification society.
- View the retrieval logs of FastGPT v4.17.0, and confirm that the number of recalled entries matches the configured
limitvalue.
The values provided here are common starting points and should be evaluated against the reader's own samples.
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.