This configuration guide is based on FastGPT v4.17.0.
What this category of data looks like
Data for this category primarily comes from silicon wafer incoming material records, workshop MES systems, operation logs of diffusion furnaces, coating equipment, and printing presses, test results from EL and IV testing equipment, and process data archived in laboratories. Update cadences cover each production batch, each individual cell’s testing nodes, and daily summary statistics. Daily documentation includes process cards, diffusion and coating parameter tables, printing and sintering operation records, EL defect detection images, IV grading data tables, and yield analysis reports. Fields include process parameter values, equipment operating duration, detection area coordinates, electrical performance indicators, and more. Units include degrees Celsius, seconds, square millimeters, and other standard units.
What constraints do these characteristics impose on multi-turn dialogue and prompt engineering
Data sources for this category are scattered, update cadences are frequent, and multi-format process, inspection, and summary documents are involved. These factors create multiple constraints for the multi-turn dialogue and prompt engineering link. Field formats vary across different production batches, so prompts must clearly specify field extraction rules and unit alignment requirements. Wafer-level and batch-level data are stored together, so the chatId and messages parameters must be used to distinguish dialogue contexts and avoid interference from cross-batch data. Unstructured descriptions of EL detection images and semi-structured IV data require gradual guidance of the model to align with document formats during multi-turn dialogue, to prevent deviation in extraction results.
How to set configurations
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
chatId | Bind to the unique identifier of the production batch | Distinguish dialogue contexts of different production batches to avoid confusion from cross-batch data |
messages | Automatically truncate old context every 15–20 dialogue turns | Adapt to frequently updated production data for this category, prevent context overflow |
variables | Preset production batch number, inspection date, and equipment model variables | Quickly locate process and inspection documents for the target batch |
| Maximum context length | 8000–12000 token | Cover the full set of process, inspection, and summary documents for a single batch, calibrated via actual testing |
| Context retention duration | 7200 seconds | Match the update interval of daily summary data, avoid interference from expired data in active dialogue |
| Prompt update cycle | Once per day | Align with the update rhythm of daily summary data, ensure prompts match the current day’s document formats |
Three common configuration errors
- Markdown tables output by the model fail to render correctly, with displayed content truncated. This occurs when compatible table formats are not explicitly specified in the prompt, and the maximum character length of a single table is not limited.
- The error
...[hide 38432 char ｜ insufficient_quota 当前分组上游负载已饱和，请稍后再试appears during dialogue. This happens when reasonable context truncation rules are not configured, causing the total token count of a single request to exceed the load limit. - After the client closes the SSE connection directly, the corresponding dialogue record is not saved to the database. This is caused by not enabling automatic persistence for the
messagesparameter, with dialogue data only saved after the full request is returned.
How to verify correct configuration
- Enter a query that includes a specific production batch number, confirm that the dialogue context only associates document data for that batch.
- Initiate multiple consecutive dialogue turns, confirm that old context is automatically truncated once the number of dialogue rounds reaches the preset threshold.
- Initiate a query that includes a Markdown table, confirm that the output format matches the preset compatible rules.
- Simulate a scenario where the client disconnects the SSE connection, confirm that the dialogue record has been saved to the database.
The values provided are common starting points and should be measured against the reader's own samples.