What the Data for This Category Looks Like
This category of data mainly comes from design BOMs, silicon steel sheet and insulating material inventory data, vacuum drying equipment logs, insulating oil test reports, and partial discharge and withstand voltage test records from high-voltage test stations. Daily documents include winding and core assembly records, vacuum drying records, insulating oil test reports, partial discharge and withstand voltage test reports, and product nameplate parameter tables. Data is updated per single product or each batch of raw materials. Most documents contain core fields such as insulation distance, water content, and test voltage, with units including professional measurement standards such as kV, ppm, and mΩ.
What Constraints These Characteristics Impose on the Knowledge Base Retrieval and Recall Link
Data for this category is updated per single product or each batch of raw materials. This requires knowledge base retrieval to accurately match identifiers such as product model and batch number, to avoid mixing irrelevant results across batches. Most daily documents contain structured test data and parameter tables. The recall logic must support accurate extraction of fields within tables. Core indicators such as insulation distance and water content have strict precision requirements. The matching degree of retrieval results must align with industrial compliance standards. Non-standardized log fragments from multiple sources require the recall link to support content parsing across multiple document formats. The full-process traceability requirement for single products means recall results must link to all relevant documents from corresponding production stages.
How to Set the Configurations
All following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the precision requirements for matching core fields such as transformer insulation parameters and test data, prevents low-matching irrelevant documents from entering the review process |
limit | Top 8 entries | The number of relevant documents for a single product typically falls between 5 and 7. Retaining 2 redundant entries ensures coverage of full-process records |
searchMode | embedding | Most documents for this category consist of structured parameters and test data. Vector retrieval can accurately match semantic associations of numeric fields |
usingReRank | Enabled | Re-ranks initial recall results by priority of test results and insulation parameters, prioritizes high-risk documents related to partial discharge abnormalities and insulation defects |
datasetSearchUsingExtensionQuery | Enabled | Automatically supplements product model and batch number as extended query terms, adapts to the accurate traceability requirements for single products |
collectionIds | Bind corresponding collection IDs by production batch | Divide knowledge base collections by single product or production batch, isolate documents from different batches to avoid confusion |
Three Common Mistakes
- Symptom: Knowledge base search tests return matching scores exceeding 4000, which falls outside the standard numerical range. Cause: No threshold limit is set for the
similarityparameter, or unnormalized matching calculation logic is used, resulting in scores not calibrated to the standard range. - Symptom: No response after operating in the agent editing page, and the page freezes. Cause: Bound
collectionIdsinclude too many redundant document collections across batches, or thelimitparameter is set too large, causing the amount of data loaded during retrieval to exceed the front-end rendering limit. - Symptom: Complex test report tables cannot be correctly recalled or parsed into retrievable content. Cause: No structured table content parsing configuration is enabled, or
searchModeis not adapted to the retrieval logic for numeric fields within tables.
How to Confirm the Configuration Is Properly Set
- Run a single retrieval test, enter the model and batch number of a single product, and check whether the recall results only include documents from the corresponding production stages.
- View the matching scores returned by the retrieval, and confirm that the scores fall within the reasonable range that meets the threshold requirements set by the
similarityparameter. - Trigger the re-ranking logic, and check whether the order of recall results prioritizes core review documents such as partial discharge tests and insulation parameters.
- Check the parameter binding of workflow nodes, and confirm that
collectionIdshave correctly linked to the knowledge base collections of the corresponding production batches.
The values provided here are common starting points, and should be evaluated against the reader's own samples.