Data Profile for This Category
Business process data for shipbuilding originates primarily from general shipbuilding assembly process specifications, classification society inspection work instructions, workshop station operation ledgers, and material requisition and inspection templates. Data updates align with new hull type project launches and quarterly compliance audits, with no requirement for high-frequency updates. Individual documents are long-form process descriptions, containing fields such as operation node numbers, technical parameters, material codes, inspection standards, and responsible personnel. Units used include professional engineering units such as millimeters, tons, work hours, and workdays.
Constraints on Knowledge Base Retrieval and Recall
The long-text structure of shipbuilding business documents requires limiting the number of single-batch recalled entries during retrieval. This prevents context overflow that reduces the readability of process guidance. Strongly structured fields such as material codes and classification society numbers in documents may not be accurately matched using only semantic recall. A combination of semantic and keyword matching must be used. Data updates occur on a long cycle, so real-time index updates are not necessary. Regular full synchronization is a suitable approach. Business process workflows have clear node hierarchies. Recalled results must match the priority of the target business scenario, to avoid disrupting the correct processing order.
The following configurations are optimized for FastGPT v4.17.0.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Shipbuilding business documents contain a large number of technical terms and precise parameters. A threshold that is too low will introduce irrelevant technical documents, while a threshold that is too high will miss matching process steps. This range is calibrated based on actual testing. |
limit | Top 3–5 entries | Individual business documents have long content. Too many recalled entries will occupy excessive context space. Combined with the node density of business processes, 3 to 5 entries are sufficient to cover core processing steps. |
searchMode | Semantic + keyword hybrid retrieval | Business documents include structured fields such as material codes. Hybrid retrieval can match both semantic meaning and precise identifiers, adapting to the dual matching needs of the business. |
usingReRank | Enabled | Initial recalled results may include entries with similar semantics but different business priorities. Reranking can adjust the result order according to the business scenario, aligning with the logic of process guidance. |
datasetSearchUsingExtensionQuery | Disabled | Shipbuilding business questions mostly focus on specific process nodes. No additional question optimization is needed, to avoid introducing redundant semantic extensions that cause retrieval bias. |
collectionIds | Bind to a dedicated business process knowledge base collection | Shipbuilding business data is isolated from data of other categories. Binding to a dedicated collection avoids irrelevant documents interfering with retrieval results. |
Three Common Configuration Errors
- Phenomenon: Response time exceeds 30 seconds after enabling
usingReRankanddatasetSearchUsingExtensionQuery. Cause: Shipbuilding business documents have long content. Question optimization generates redundant extended queries, and the reranking model needs to process more recalled entries, leading to increased calculation time. - Phenomenon: Retrieval results do not match documents with specified material codes. Cause: Only semantic retrieval mode is enabled, and hybrid retrieval is not used. Precise matching of structured fields in documents cannot be achieved.
- Phenomenon: Recalled results include non-business process documents. Cause: No dedicated
collectionIdsis bound, so the retrieval scope covers knowledge base content from other categories.
How to Verify Proper Configuration
- Submit a business query that includes a specific material code, and confirm that retrieval results include documents with matching fields.
- Review interface response logs to confirm that latency after enabling
usingReRankfalls within expected ranges. - Check the count of retrieved results, and confirm it does not exceed the configured
limitvalue. - Validate that after disabling
datasetSearchUsingExtensionQuery, retrieval results more closely align with the original wording of the query.
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.