What Data Looks Like for This Category
Data sources for this category include design BOM, silicon steel sheet and insulating material procurement data, vacuum drying equipment logs, oil test reports, and partial discharge and withstand voltage test records from the test station. Data is updated per product and per batch of raw materials. Daily documents include winding and core assembly records, vacuum drying records, insulating oil reports, partial discharge test reports, and nameplate parameter sheets. Inspection data for a single product includes professional fields with units such as insulation moisture content (ppm), partial discharge value (pC), and withstand voltage value (kV).
Constraints for Multi-turn Dialogue and Prompt Engineering
The data is updated per product and per batch of raw materials. This structure requires multi-turn dialogue to bind chatId and the corresponding product ID to avoid cross-product data confusion. Professional inspection documents include fields with units such as insulation moisture content and partial discharge value. Prompts must limit retrieval scope to documents associated with the current product to prevent irrelevant data from interfering with review results. Single-product inspection data volumes are large. Multi-turn dialogue context must be restricted to the document set for a single product to avoid exceeding model processing limits. Each product’s inspection process has fixed steps. Prompts must pre-set review logic for corresponding steps to reduce manual intervention steps.
Configuration Settings
The following configurations follow the interface and parameter rules for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chatId | Bind to the unique product code of a single transformer | Manage dialogue context independently per product to avoid cross-product data interference |
messages | Retain the most recent 15–20 rounds of dialogue content | Dialogue rounds for a single product’s inspection process typically fall within this range; overly long contexts dilute critical review information |
variables | Pre-set product number, test date, and insulation grade variables | Quickly associate exclusive documents such as design BOM and test reports for the corresponding product |
| Maximum context length | 8000–12000 token | Total character count of inspection documents for a single product matches this range, adapting to the processing limits of general large language models |
| Single retrieval limit | Top 5 results | Number of relevant inspection documents for a single product is concentrated; excessive retrieval results interfere with model judgment |
| Image recognition switch | Enabled, limited to recognizing only test report-type images | Images for this category’s inspection data are mostly screenshots of partial discharge and withstand voltage test records, filtering irrelevant image interference |
Three Common Configuration Mistakes
- Enabling image recognition but failing to return image content. After uploading a JPEG image of a partial discharge test report, the model replies that it cannot recognize the image content. The cause is that the associated configuration for image recognition was not enabled in the AI dialogue node, or the uploaded image format does not meet supported requirements.
- Dialogue automatically cites associated files at the end. Dialogue output includes document links and fragments of cited insulating oil reports and test records. The cause is that the default "citation source files" setting in the dialogue node was not disabled, or the retrieval scope was not limited to exclusive documents for the current product.
- Data interference after multi-turn dialogue. Inspection data for cross-batch insulating materials appears in the dialogue context of the current transformer product. The cause is that
chatIdwas not used to bind the unique code for a single product, leading to confusion of dialogue contexts between different products.
How to Verify Proper Configuration
- Manually enter the product number of a specified transformer, check whether the dialogue automatically associates exclusive documents such as winding records and insulating oil reports for that product.
- Upload a properly formatted partial discharge test report image, verify whether the model can correctly recognize professional data such as discharge amount and test date in the image.
- Initiate more than 20 rounds of dialogue for the inspection process, check whether the context retains the most recent 15–20 rounds of content, and no cross-product data interference occurs.
- Review the settings of the dialogue node, confirm that the "citation source files" option has been disabled, or the retrieval scope has been limited to the document set for the current product.
The values provided here are common starting points, and should be evaluated against the reader’s own samples.