What the data for this category looks like
Data for this category originates from design BOM, silicon steel sheet and insulating material procurement records, vacuum drying equipment logs, insulating oil test reports, and high-voltage test station records. Data is updated per individual product or each batch of materials. Corresponding documents include winding and core assembly records, vacuum drying logs, insulating oil test reports, partial discharge and withstand voltage test reports, and product nameplate parameter tables. Core fields include insulation distance, moisture content, partial discharge value, and withstand voltage level. Common units include millimeters, picocoulombs, kilovolts, and similar units.
Constraints imposed by these characteristics on citation sources and traceability
Single-product traceability requires citation traceability to be bound to specific product batches and serial numbers. This prevents document confusion across products. Field differences exist across multiple test document types. Retrieval must match the field types of the target document. For example, partial discharge data only associates with high-voltage test records. Updates follow a per-product or per-batch rhythm. Cited data sources must be the latest batch of test and assembly records. This avoids using expired data. Parameters related to insulation structure and cleanliness require strict rigor. Traceability must clearly state the citation source’s generation time and test environment. This ensures the accuracy of Q&A content.
The following configurations are based on FastGPT v4.17.0.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Similarity threshold | 0.75–0.85 | Matches text features of transformer test reports and assembly records, filters low-relevance non-process documents |
| Single search limit | Top 8 entries | Core process document count for a single product is concentrated. Too many citations will dilute the focus of responses |
quoteMaxToken | 1200–1500 token | Adapts to the text length of insulating oil reports and withstand voltage test reports, ensures complete citation of key parameters and conclusions |
| Citation display rule | Display source document type and generation time | Meets single-product traceability requirements, clarifies traceability time and document category |
detail | Enable full field display | High-voltage test data requires clear details such as partial discharge values and test environment, ensuring the rigor of Q&A content |
responseData | Only return data sources associated with product batches | Avoids incorrect citation of cross-batch documents, matches the single-product traceability requirements of this category |
Three common mistakes
- Phenomenon: Citation documents not added to the local knowledge base are returned in conversations. Cause: No permission binding rules for data sources are configured, causing the retrieval scope to exceed the preset local knowledge base range.
- Phenomenon: The number of citations exceeds the business expected upper limit. Cause: No single search limit parameter is set, or the similarity threshold is set too low, resulting in a large number of low-relevance redundant documents being retrieved.
- Phenomenon: The citation display area in conversations cannot be closed. Cause: The default state of the citation display switch is not configured correctly, or backend configurations are not synchronized to the frontend rendering logic.
How to confirm the configuration is correct
- Initiate a process question targeting a specific product batch. Verify that all returned citation sources belong to the test and assembly documents of that batch.
- Adjust the similarity threshold and single search limit. Initiate a test question. Verify that the number of returned citations matches the preset range.
- Check that the citation display area shows key information such as source document type and generation time. Confirm that the
detailconfiguration takes effect. - Try to modify the citation display rule configuration. Refresh the conversation. Verify that the display status is updated synchronously.
The values provided are common starting points and should be measured against the reader's own samples.