What data for this category includes
Data sources for photovoltaic cell manufacturing processes and testing include silicon wafer incoming inspection records, production MES system logs, real-time operating parameters of diffusion furnaces, coating equipment, and printing presses, EL defect detection images, IV grading performance data, and laboratory yield reports. Data updates follow this rhythm: once per batch of incoming silicon wafers, real-time updates for each product’s testing data, and daily aggregation of production yield data. Daily documents include process cards, individual device operation logs, and single-batch testing reports. Fields include furnace temperature, deposition rate, defect coordinates, open-circuit voltage, and similar items. General percentage-based statistical fields are not included.
What constraints these characteristics impose on multi-turn dialogue and prompt engineering
Photovoltaic cell manufacturing process data has strong batch correlation, high real-time requirements, and diverse document types. Multi-turn dialogue must strictly limit the context scope to avoid cross-batch parameter confusion. Each product’s testing data is independent. Dialogue must retain batch numbers or wafer numbers as context variables to ensure matching between queries and retrieved data. Process parameters have high precision requirements. Prompts must explicitly limit use of only official document data from the current batch, and prohibit generation of unfiled parameter values. Data updates occur frequently. The retrieval scope of dialogue context must be refreshed regularly to avoid using expired process parameters.
Configuration Settings
The following configurations are based on FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Context window size | 8000–12000 token | Adapts to the long text length of photovoltaic process documents, retains sufficient multi-turn dialogue context |
variables binding rules | Bind production batch number and wafer number variables | Associates current dialogue queries with process data and testing reports for the corresponding batch |
chatId configuration | Fixed session ID, persistent storage | Ensures coherent context for multi-turn dialogue, avoids session interruptions or permission errors |
messages truncation threshold | Retain the most recent 15–20 dialogue messages | Filters redundant history, prevents context overflow that impacts response accuracy |
| Prompt template configuration | Limit retrieval to process documents for the current batch | Ensures responses only use official process data linked to the current dialogue, avoids cross-batch parameter confusion |
Three Common Mistakes
- Phenomenon: An error stating "No permission to operate this conversation record" is returned when sending consecutive dialogue requests in a short time. Cause:
chatIdis not correctly bound, leading to session identifier conflicts or unauthorized access to the current dialogue context. - Phenomenon: Dialogue records are lost after refreshing the dialogue page. Cause:
chatIdis not persisted to session storage, or the dialogue history retention function is not enabled in the configuration. - Phenomenon: Generated process parameters do not match the current batch. Cause: Production batch number variables are not bound, and the prompt does not limit retrieval to process documents for the current batch.
How to Confirm Proper Configuration
- Send a query that includes a specific production batch number, verify that the returned results only link to process cards, device operation logs, and testing data for that batch.
- Send 3 consecutive related process queries, verify that the dialogue context remains coherent, with no context loss or parameter confusion.
- Send 2 dialogue requests for different batches in a short time, verify that independent session results are generated for each, with no permission errors.
- Refresh the dialogue page, verify that the original dialogue records are retained normally, and no new dialogue is displayed.
The values provided are common starting points, and should be evaluated against the reader’s own samples.