Data Overview for This Category
Data is created for individual transformers and high-voltage electrical products. 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 sources include design BOMs, silicon steel sheet and insulating material data, vacuum and drying equipment records, oil test results, and test station inspection reports. Updates occur once per product or per batch of raw materials. Document fields cover product structure parameters, test and detection values, and material specification information.
Constraints for Multi-Turn Dialogue and Prompt Engineering
This category’s data features per-unit or batch updates, and a high share of professional test documents. These traits create multiple constraints for multi-turn dialogue and prompt engineering workflows. Bind exclusive parameters and test data to individual products to avoid context confusion, so chatId must link to product numbers or batch numbers. Prompts must adapt to professional electrical terminology, and output formats must be limited to standardized tender document and quotation text, with unnecessary formats prohibited. Retain product parameter memory from each dialogue turn, so subsequent questions can directly call insulating oil data and test report content for the corresponding batch. Add parameter verification logic to prevent mismatches between insulating parameters and test results.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chatId | Bind to unique product number | Ensure dialogue context is linked to exclusive data for a single product, avoid cross-product parameter confusion |
messages | Retain the first 20 turns of dialogue content | Cover full workflow memory for a single product from parameter confirmation to quotation adjustment |
variables | Pass product BOM and test report ID | Allow dialogue nodes to directly call insulating oil and withstand voltage test data for the corresponding batch |
| Output format restriction | Use only plain text, disable Markdown syntax | Match formal typesetting requirements for tender documents, avoid format chaos |
| Single recall limit | Top 3 matching documents | Focus on most relevant test reports and BOM data, reduce irrelevant information interference |
| Maximum response length | 1500–2000 token | Cover core content of quotation and tender documents for a single product, avoid truncation of critical parameters |
Three Common Misconfigurations
- Output results still contain Markdown syntax. This occurs when prompts fail to explicitly prohibit non-plain text formats, or format verification rules do not take effect.
- Multi-turn dialogue context is confused, with cross-product parameter mixing. This occurs when
chatIdis not used to bind unique identifiers for individual products, leading to shared dialogue contexts across different products. - Workflow branches fail to exit correctly, repeatedly entering the question classification node. This occurs when the return value of the dialogue node is not configured to directly output results, and no termination condition for branch jumps is set.
How to Validate Proper Configuration
- Initiate a test dialogue, enter the product number of a transformer, and check whether the output content only includes test data and quotation information for that product, with no parameters from other products.
- Check the output result to confirm that no Markdown headings, lists, or other elements appear, and only plain text content is present.
- Trigger a workflow test, verify that after the dialogue node returns a result, the workflow directly enters the completion stage, and does not repeatedly enter the question classification node.
- After publishing the application, attempt to initiate a dialogue directly, confirm that no interactive buttons need to be clicked to enter the question-and-answer session.
The values provided here are common starting points, and should be evaluated against the reader's own samples.