What the data for this category looks like
The internal office and employee service data for rail transit equipment mainly comes from equipment operation and maintenance manuals, parts BOM tables, train maintenance procedures, on-site operation logs, and employee training courseware. Data sources cover the entire R&D, production, and operation and maintenance links. The update rhythm is adjusted based on business scenarios: operation and maintenance manuals are updated quarterly in accordance with industry standards, BOM tables are updated every six months with new product development iterations, and on-site maintenance records are entered in real time. Most documents are a mix of structured and semi-structured data, including fields such as equipment number, model, maintenance cycle (unit: hours), fault code, and operation steps.
What constraints these characteristics impose on knowledge base retrieval and recall
Structured BOM tables and maintenance records contain clear fields and units. Retrieval must accurately match field dimensions to avoid irrelevant results from general semantic retrieval. Real-time updated on-site maintenance logs require incremental recall support to avoid delays from full retrieval. For long documents such as operation and maintenance manuals and training courseware, chunked content must retain chapter association information to prevent retrieved segments from being separated from their context. Internal employees often query specific equipment models. Retrieval scope must be filtered by equipment number and model dimensions to reduce the pool of invalid results.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches the precision requirements for structured fields of rail transit equipment, filters low-correlation non-equipment results |
limit | Top 8 entries | Internal employee queries only require up to 8 relevant documents per request to cover information needs for maintenance and training scenarios |
searchMode | vector | For structured and semi-structured equipment data, vector retrieval can accurately match fields such as equipment model and fault code |
usingReRank | Enabled | Retrieval results after chunking long documents require reranking to correct priority deviations from semantic matching |
datasetSearchUsingExtensionQuery | Enabled | Optimize queries from employees’ colloquial requests such as “How to maintain XX model train” by supplementing precise fields like equipment numbers |
collectionIds | Bind corresponding collection IDs by production workshop and equipment model | Limit the retrieval scope to operation and maintenance documents of the corresponding category, avoid invalid recall across models |
Three common errors
- In FastGPT v4.17.0, after
datasetSearchUsingExtensionQueryis enabled, even ifusingReRankis disabled in the final AI answer plugin, the knowledge base reranking switch remains displayed as inactive. Cause: WhendatasetSearchUsingExtensionQueryis enabled, the system automatically enables reranking logic, and the linkage configuration with the global reranking switch is not correctly identified. - After migrating the image, the login interface and knowledge base management interface switch to English, and all configuration items have no Chinese display. Cause: Localized language configuration files were not synchronized during image migration, and the system loads the English language pack by default.
- When calling knowledge base retrieval, the number of valid results returned is far lower than the preset
limitvalue, and timeouts occur frequently. Cause:collectionIdsare not bound by equipment category, and the retrieval scope covers all long documents, resulting in excessive system resource usage.
How to confirm configuration is correctly set
- View retrieval logs to confirm that returned results include precisely matched fields such as equipment number and model, to verify the effectiveness of retrieval configuration.
- Submit a colloquial query such as “Monthly maintenance steps for XX model train” to check whether precise retrieval fields are automatically supplemented, to verify the effectiveness of query optimization configuration.
- Check the sorting logic of retrieval results to confirm that the priority of results after chunking long documents aligns with business expectations, to verify the effectiveness of reranking configuration.
- Switch queries for different equipment models to confirm that the retrieval scope only covers documents of the corresponding category, to verify the effectiveness of collection binding configuration.
The values provided here are common starting points and should be validated against samples used in the target deployment.
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.