What Data for This Category Looks Like
Data for this category comes from design BOMs, silicon steel sheet and insulating material ledgers, vacuum drying equipment logs, oil test reports, and high-voltage test records from 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 equipment nameplate parameter tables. Data is updated per single product or batch of raw materials. Fields include parameters with clear units, such as insulation moisture content, partial discharge peak value, and withstand voltage value.
How These Characteristics Create Constraints for Vector Models and Indexing
Data for this category is updated per single product or batch of raw materials. This sets clear requirements for incremental index synchronization frequency. Documents contain two types of content: structured test parameters and non-standardized operation records. Vector models must balance semantic association of numeric fields and contextual completeness of text descriptions. Full-process traceability requirements for single products require indexes to support precise filtering by equipment number and batch number. This prevents cross-batch data from interfering with audit results. Insulation-related parameters have high precision requirements. Vector embedding must retain parameter units and threshold association information to prevent loss of key business semantics. The setup must also adapt to the small-batch data scale with high-frequency updates.
Configuration Settings
This configuration is written based on FastGPT v4.17.0.
| Config Item | Suggested Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Match the text length of single test reports for this category, avoid cutting key parameters |
| Overlap Length | 100–150 characters | Retain association information for insulation parameters between adjacent chunks, prevent semantic loss from chunk breakage |
maxToken | 8192 | Match the context window limit of mainstream vector models, adapt to long-text test records |
| Similarity Threshold | 0.75–0.85, calibrated based on actual measurements | Distinguish the correlation degree of normal and abnormal partial discharge and withstand voltage test data |
| Single Search Limit | Top 10 entries | Cover the number of associated test records for a single product, avoid redundant data interfering with audits |
| Maximum Context | 40000 characters | Carry all audit-related document content for a single product |
Three Common Misconfigurations
- Vectorization requests only accept single text slices, and batch processing is not enabled. The phenomenon is that only single text parameters appear in API call logs. Vector service monitoring shows low utilization of batch processing queues. The cause is that FastGPT batch vectorization trigger rules are not configured. Requests are submitted by default as single text entries.
- Index results do not match the current audit equipment. The phenomenon is that search returns test reports belonging to other batches or other equipment. The cause is that index filtering rules based on equipment number or batch number are not configured, leading to recall of cross-batch data.
- Key parameters are lost after chunking. The phenomenon is that fields such as insulation moisture content and partial discharge peak value have no corresponding semantic segments after vector embedding in audit documents. The cause is that the chunk length is set too small, separating parameters with units from their context.
How to Confirm Proper Configuration
- Check batch processing monitoring of the vector service, confirm that requests contain multiple text slices, to verify that batch vectorization configuration takes effect.
- Submit an audit query for a single device, check that returned results only include associated documents for that device, to verify that index filtering rules are configured correctly.
- Export chunked text fragments, confirm that insulation parameters, test data and context are not separated by chunking, to verify that chunking configuration is reasonable.
- Call the FastGPT API to obtain chunked index content, confirm that returned chunked data contains complete document structure and field information, to verify that index extraction configuration is correct.
The values provided here are common starting points, and should be evaluated against the reader's own samples.