What the data for this category looks like
Shipbuilding ledger data comes from material requisition records for each construction process, block quality inspection reports, welding hour ledgers, main engine operation logs, and similar sources. Updates follow construction milestones, with changes completed when each block is finished or batch materials are placed in storage. Most documents are structured tables, containing fields such as material code, batch number, inspection date, qualification status, usage location, and hour count. Units include tons, meters, hours, pieces, and other standard measurements. The length of a single ledger document ranges from hundreds to thousands of characters. Some full-ship general ledger documents can reach tens of thousands of characters.
What constraints these characteristics impose on knowledge base retrieval and recall
Shipbuilding ledgers have numerous specialized fields. Retrieval requires precise matching of key information such as material codes and inspection results. A large share of documents are long, and direct upload will exceed the input length limit of the vector model. Chunking processing must be completed in advance. Ledger update frequencies fluctuate with construction progress. Ledgers for some urgent batches need real-time synchronization to the knowledge base, so incremental synchronization frequency must align with construction milestones. Ledger formats vary slightly across different batches. Retrieval must balance structured fields and unstructured process descriptions. A single search mode cannot cover all recall requirements.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Similarity threshold | 0.75–0.82 | Shipbuilding ledgers have many specialized terms, so low-correlation recall results need to be filtered. This range balances recall coverage and precision |
| Single search limit | Top 8 entries | Single batch ledger data is concentrated, excessive recall increases token consumption, and meets the narrow-domain retrieval needs of the shipbuilding scenario |
| Search mode | Hybrid vector + keyword | Balances structured field matching and semantic recall, adapting to the characteristics of ledgers that have both precise coded fields and process descriptions |
| Reranking | Enabled | Performs secondary sorting on initial recall results, raising the display priority of highly relevant results, adapting to multi-field retrieval scenarios |
| Query optimization | Enabled | Expands colloquial user queries to match specialized terms in ledgers, reducing false recalls from casual conversation content |
| Dataset ID set | Fill in according to the actually bound shipbuilding ledger dataset | Limits the retrieval scope to avoid recalling irrelevant knowledge base content |
Three common mistakes
- Phenomenon: The number of retrieval results far exceeds the set single search limit, causing reply token overflow. Cause: The single search limit parameter is not configured correctly, or the threshold is not adjusted according to the narrow-domain characteristics of shipbuilding ledgers, resulting in redundant data being recalled.
- Phenomenon: Casual conversation queries still recall some shipbuilding ledger content, and irrelevant results cannot be effectively filtered. Cause: The query optimization configuration is not enabled, or the similarity threshold is set too low, resulting in low-correlation semantic matching results being retained.
- Phenomenon: Vector generation fails, and the interface displays an input length limit exceeded error. Cause: Single ledger documents are not chunked to the length supported by the vector model, and direct upload of full-ship general ledgers causes the limit to be exceeded.
How to confirm the configuration is correct
- Upload a single shipbuilding material ledger document, trigger retrieval, and check that the number of recall results does not exceed the set single search limit.
- Enter a casual conversation query unrelated to shipbuilding, and check that the retrieval result is empty or contains no relevant ledger content.
- View the vector generation log, and confirm that the length of chunked single documents complies with the input limit of the vector model used by FastGPT v4.17.0.
- Test adjusting the similarity threshold, and check that the relevance of recall results fluctuates reasonably with changes in the threshold.
The values given are common starting points and should be measured 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.