What this category’s data looks like
Data sources for this category include design BOM, silicon steel sheet and insulating material data, vacuum and drying equipment, oil testing, and test stations. Update frequency is per individual product and per material batch. 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. Data fields include insulation distance, moisture content, test values, and assembly process information. Units include megapascals, microamps, millimeters, and other standard units. Full end-to-end data for a single product must be fully linked for traceability.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Exclusive full-process data for each product requires retrieval results to be accurately bound to the corresponding product number, to avoid mixing cross-batch data. Frequently updated material and test data requires retrieval to match the latest batch of document content in real time. In scenarios where multiple types of professional documents are stored together, retrieval must support filtering and recalling by document type or process. Professional data such as partial discharge and withstand voltage from high-voltage tests has strong semantic professionalism, requiring the retrieval model to accurately match semantic associations of professional terms. The core requirement of single-product traceability requires recalled results to be linked to specific source documents such as test reports and assembly records.
How to set the configuration
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches professional semantic similarity for transformer insulation tests and assembly records, filters low-correlation recall results |
limit | Top 6 entries | Adapts to the effective data volume for single-product traceability, avoids context overload that affects analysis efficiency |
searchMode | embedding | For long texts such as professional test reports and assembly records, prioritizes matching semantic associations to meet query needs for professional terms |
usingReRank | Enabled | Performs secondary ranking on initial recall results, improves ranking accuracy for professional content such as partial discharge anomalies and insulation parameters |
datasetSearchUsingExtensionQuery | Enabled | Generates extended query terms based on user questions, covers variant queries for professional terms such as insulation distance and moisture content |
collectionIds | Knowledge base collection IDs bound to product batches | Limits the retrieval scope to the corresponding batch of test and assembly documents, avoids mixing cross-batch data |
Three common misconfigurations
- Results returned after calling the
/api/v1/chat/completionsinterface are not linked to source documents. Cause:datasetSearchUsingExtensionQueryis not configured in the request parameters, orcollectionIdsis not correctly bound, leading to the retrieval scope not being limited to the target knowledge base collection. - The number of retrieval results does not match the configured
limit, with excess or insufficient results. Cause: The recall upper limit parameter for retrieval is not set correctly, or mixed multi-batch document data causes deviation in the initial recall volume. - Uploaded documents such as insulating oil reports and withstand voltage test reports cannot be retrieved. Cause: Documents are not categorized into the knowledge base collection corresponding to
collectionIds, or the file format is not correctly recognized by the parsing module of FastGPT v4.17.0.
How to confirm the configuration is correct
- Initiate a test query, enter "the cause of partial discharge anomalies in a certain batch of transformers", and check whether the returned results include the file name or paragraph information of the source document.
- View the retrieval log, confirm that the recalled text blocks all come from the knowledge base collection corresponding to the bound
collectionIds, with no cross-batch irrelevant data. - Adjust the
similarityparameter value, observe the change in retrieval result relevance, and verify that the parameter configuration takes effect. - Call the knowledge base retrieval interface, check that the number of returned results matches the configured
limit.
The values provided here are common starting points, and should be measured against the reader's own samples.