What data for this category looks like
The system sources core category data from design BOM, silicon steel sheet and insulating material procurement data, vacuum and drying equipment operation logs, oil test reports, and high-voltage test station records. Daily documents include winding and core assembly records, vacuum drying records, insulating oil reports, partial discharge and withstand voltage test reports, and nameplate parameter tables. The category updates data per individual product and per material batch. Fields include insulation distance, water content, partial discharge peak value, test voltage, and others. Units include kV, ppm, mA, and similar units.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The category updates data per single product unit and single material batch. This requirement means the retrieval system must support incremental indexing and single-unit associated recall. Document types cover structured parameter tables and unstructured test reports. Both exact matching and semantic retrieval must be supported. Core work focuses on insulation and high-voltage tests. Retrieval must prioritize recalling key documents such as insulating material data and partial discharge test records. Full-chain traceability for single products means retrieval results must strictly associate with all data for the corresponding model and batch, to avoid interference across product lines.
How to configure the system
The following configurations are optimized for FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Insulation and high-voltage test parameters have high matching confidence requirements, to avoid risks from low-match results |
limit | Top 6–8 results | The number of documents associated with a single product is limited. Too many results will exceed the model context window |
usingReRank | Enabled | Semantic matching for unstructured test reports requires reranking to improve result accuracy |
searchMode | mixedRecall | Supports both exact matching of structured nameplate parameters and semantic retrieval of unstructured process documents |
datasetSearchUsingExtensionQuery | Enabled | Supplementary context association for process specifications, automatically associates insulating material data for the corresponding batch |
collectionIds | Bind exclusive collections for corresponding product lines | Isolate transformer data for different models and batches, to avoid interference from irrelevant documents across product lines |
Three common configuration errors
- Phenomenon: Retrieval results include documents for non-target transformer models. Some structured fields cannot be matched correctly. Cause: Did not bind an exclusive dataset via
collectionIds, leading to recall of irrelevant data across collections. - Phenomenon: Retrieval results do not associate with insulating material data for the corresponding batch. Only isolated process specification text is returned. Cause: Did not enable
datasetSearchUsingExtensionQuery, so no supplementary context association information was added. - Phenomenon: Some test data in insulating oil reports cannot be retrieved. The interface displays parsing failure. Cause: Did not unify the character encoding format of documents. Documents with non-UTF-8 encoding cannot be parsed correctly.
How to confirm the configuration is correct
- Input the model keyword for a single transformer, and verify that recalled documents include full-chain data such as winding records and withstand voltage test reports for that model.
- View the matching score of retrieval results, and confirm that it meets the preset
similarityrange requirement. - Test a model that does not support function call, and confirm that retrieval results are automatically included in the conversation context window.
- Upload test documents with different character encoding formats, and verify that they can be parsed normally and participate in retrieval.
The values provided are common starting points and should be measured against the samples used in actual operations.