What this category of data looks like
Photovoltaic cell manufacturing process and abnormal data typically originates from process, equipment, quality control, and production departments at crystalline silicon cell manufacturers. Daily documentation includes process cards, diffusion and coating parameters, printing and sintering records, EL detection images, IV grading data, and yield reports. Data sources include incoming silicon wafers, MES systems, diffusion furnaces, coating equipment, printing presses, EL and IV testing devices, and laboratory datasets. Updates occur per production batch, per individual piece testing, and daily aggregation. Document fields cover process parameter values, detection image identifiers, electrical performance metrics, defect types, and yield percentages. No unified fixed format exists, but all data must align with the production batch dimension.
What constraints do these characteristics impose on multi-turn dialogue and prompt engineering
This category of data has high update frequency and strong batch correlation. Multi-turn dialogue must strictly limit the number of context history rounds to avoid introducing redundant process data from non-current batches. Documents contain both structured parameters and unstructured images and detection reports. Prompts must clearly define field ranges and batch dimensions to prevent confusion between parameters of different production batches. Production processes are continuous, and surface microstructures are sensitive. Multi-turn dialogue must retain key process context to ensure continuity during abnormal analysis. Data sources are scattered, so prompts must guide calls to fields from corresponding sources to avoid generating content that deviates from actual production data.
How to set configurations
These configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Context history rounds | Previous 3–5 dialogue turns | Photovoltaic data updates frequently; overly long context will introduce redundant process parameters from non-current batches |
| Prompt variable binding | Bind batch-specific fields in variables | Match the fixed production batch dimension of process cards and IV data |
| Maximum tokens per reply | 8000–12000 tokens | Accommodate complete content of EL detection descriptions and yield reports |
| Context recall limit | Top 2 historical dialogues | Avoid mixing data from unrelated production batches in multi-turn dialogue |
| Error retry count | 2–3 retries | Address initial request failures caused by delayed equipment data uploads |
Three common mistakes
- Phenomenon: When using the same prompt to ask questions, FastGPT’s reply effect is weaker than that of native large model platforms. Cause: Batch-specific process parameters in
variablesare not bound, and the prompt does not match the actual data of the current production batch. - Phenomenon: Workflow dialogue logs are empty. Cause: Dialogue log storage configuration is not enabled, or
chatIdis not correctly passed in the request. - Phenomenon: In a Docker-deployed FastGPT instance, model testing works normally but workflow dialogue shows a failure. Cause: Workflow nodes are not correctly associated with the configured model, or environment variables do not sync the model access key.
How to confirm the configuration is correct
- Import a single batch’s process card and IV data, test whether the prompt can accurately extract process parameters and electrical performance indicators for that batch.
- Initiate a multi-turn dialogue, verify that the context history round limit is enforced, and that no more than the set number of historical dialogue contents are loaded.
- Trigger the abnormal report drafting process, check whether the returned content includes EL defect identification, IV grading results, and yield analysis.
- View the dialogue request log, confirm that
chatIdandmessagesparameters are correctly passed, and ensure logs are generated normally.
The values provided here are common starting points, and should be evaluated against the reader’s own samples.