What the data for this product category looks like
Data sources include design BOMs, purchase ledgers for silicon steel sheets and insulating materials, vacuum drying equipment logs, oil test reports, and high-voltage test station records. The data update schedule is that each product corresponds to a complete set of records, and corresponding ledgers are updated synchronously for each batch of raw materials. 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. Fields include insulating structure parameters, test voltage values, drying duration, partial discharge values, and other relevant metrics, with units covering professional measurement standards such as kilovolts, picocoulombs, and hours. Full-process data for a single product is independent and complete, and no cross-batch combined analysis is required.
What constraints these characteristics impose on multi-turn dialogue and prompt engineering
The independent traceability requirement for a single product means dialogue must be bound to a unique identifier to avoid cross-product data confusion. Multi-source documents have significant format differences, so prompts must guide dialogue to extract information using category-specific fields to avoid generalized analysis. Individual document lengths vary widely, so context length must be limited to a reasonable range to avoid overload that impacts response speed. In scenarios with many professional terms, prompts must limit analysis dimensions, focus on core links such as insulation and high-voltage testing, and reduce invalid output. Data updates occur at a high frequency, so dialogue contexts must prioritize calling the latest test and assembly records to ensure the timeliness of analysis results.
Configuration guidelines
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chatId | Bind to the unique production number of a single transformer | Establish full-process data association for a single product and avoid cross-product confusion |
messages | Limit context to the first 3 rounds of core dialogue + 1 core document summary | Adapt to differences in document lengths for this product category and avoid context overload |
variables | Preset two variables: insulation parameters and test batch | Match core analysis dimensions for this product category and simplify calls to professional parameters |
| Maximum context length | 800–1200 characters | Adapt to the average length of individual category documents and ensure response efficiency |
| Timeout period | 600 seconds | Adapt to the time required for professional analysis and avoid mid-run interruptions |
| Maximum number of recalled documents per call | Top 5 | Focus on core associated data and reduce interference from redundant information |
| Maximum single uploaded document size | 1000 MB | Adapt to upload requirements for large-sized single-category test reports |
Three common configuration errors
- Issue: Dialogue export records are not associated with a product number, and fields are empty. Cause: Traceability data for a single product is not bound via
chatId, only general dialogue content is stored. - Issue: After uploading attachments such as insulating oil reports and withstand voltage test reports, subsequent dialogue association cannot be completed. Cause: The allowed document type whitelist is not configured, or the function for binding attachments to dialogue is not enabled.
- Issue: After calling the api/v1/chat/completions interface, the returned result does not match the specified knowledge base tag. Cause: The knowledge base association identifier is not correctly passed in the request parameters, or the exclusive dataset for the corresponding product category is not bound.
How to verify the configuration is correct
- Enter the nameplate parameters of a single transformer when initiating a dialogue, and check whether the dialogue automatically associates the corresponding batch of assembly and test records.
- After uploading an insulating oil report, initiate a partial discharge anomaly analysis, and check whether the returned result includes specific test values from the report.
- Call the api/v1/chat/completions interface, pass the specified knowledge base tag, and check whether the returned result only associates the dataset for the corresponding product category.
- View the dialogue history, and check whether
chatIdmatches the unique production number of the single product.
The values provided here are common starting points and should be evaluated against the reader's own samples.