What the data for this category looks like
Internal office data for shipbuilding primarily comes from ship design drawings, production process specifications, equipment maintenance manuals, bills of materials, employee training materials, and classification society compliance documents. Update rhythms vary: design drawings are only updated during major redesigns or quarterly reviews, production process specifications are adjusted with each production batch, maintenance manuals are updated alongside equipment maintenance cycles, and bills of materials are updated monthly alongside procurement plans. Documents include structured fields such as material codes, main engine models, and rated power, as well as unstructured text such as operation steps and troubleshooting workflows. Professional units include newton meters, kilowatts, work hours, tons, and similar units.
Constraints Imposed on Knowledge Base Retrieval and Recall by These Characteristics
A large number of specialized structured fields require prioritizing matching specified fields during retrieval, to prevent irrelevant content from being included in generalized recall results. Long document lengths require retaining field integrity during chunking, to avoid splitting critical parameter information. Uneven update frequencies require configuring differentiated incremental update tasks, separating update cycles for static and dynamic documents. Dense specialized terminology, combined with employee questions often using colloquial language, requires professional query expansion to improve recall accuracy.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Shipbuilding documents contain many specialized terms. A threshold that is too low will introduce irrelevant content, while a threshold that is too high will miss relevant specialized documents |
limit | Top 8–12 results, single chunk length 300–800 characters | Single retrieved document chunk length fits the context window of most large models. 8–12 results ensure complete content without exceeding limits |
searchMode | Structured priority | Internal office documents include structured fields such as material codes and classification society certification numbers. Prioritizing matching these fields improves recall accuracy |
usingReRank | Enabled | Initial similarity calculations are easily disrupted by synonymous terminology. Reranking corrects the initial recall ranking results |
datasetSearchUsingExtensionQuery | Enabled | Employee questions often use colloquial language. Expanding specialized terms improves recall relevance |
collectionIds | Bind shipbuilding-specific collections | Isolate knowledge base data from other categories to avoid retrieving irrelevant content |
Three Common Configuration Errors
- Phenomenon: Retrieved answers do not use the content of the document with the highest initial similarity score. Cause: When
usingReRankis enabled, the reranking model reorders results based on global semantic relevance, and does not retain the highest similarity entry from the initial recall list. - Phenomenon: Irrelevant document fragments unrelated to shipbuilding appear in retrieval results. Cause: Dedicated
collectionIdsare not bound, resulting in recall of knowledge base content from other categories. - Phenomenon: Retrieval results fail to match precise parameter fields such as rated torque. Cause: Structured fields were split during chunking, preventing retrieval from matching complete parameter information.
How to Confirm the Configuration Is Correct
- Enter the FastGPT knowledge base test page, input professional questions related to shipbuilding, and check whether retrieval results include document content from the corresponding category.
- Check the reranking switch status in the retrieval configuration, confirm that
usingReRankis enabled, and compare the initial recall list with the final displayed ranking results. - Adjust the value of
similarity, observe changes in the number of retrieval results, and confirm that the current threshold meets business requirements for relevance. - Check the binding configuration of the knowledge base collection, confirm that only document collections related to shipbuilding are associated, and no data from other categories is included.
- After enabling
datasetSearchUsingExtensionQuery, input colloquial questions and check whether retrieval results include more matching specialized documents.
The values provided 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.