What the Data for This Category Looks Like
Quality audit data for photovoltaic cell manufacturing typically comes from silicon wafer incoming material records, production MES systems, operation logs of diffusion furnaces, coating equipment and printing presses, EL and IV test reports, and laboratory test data. Data update frequencies cover per-batch production, per-wafer test results, and daily summaries. Daily documents include process cards, diffusion and coating parameter tables, printing and sintering records, EL detection images, IV grading data, and yield reports. Data fields include process temperature and time parameters, EL crack markers, and IV electrical performance values. Field units use standard metrics for process parameters and physical units for test results.
Constraints Imposed on Multi-turn Dialogue and Prompt Words by These Characteristics
The multi-source and high-frequency update data feature requires multi-turn dialogue to bind chatId to associate production batches, to avoid mixing data across batches. Unstructured documents such as EL detection images and IV grading data require prompt words to clearly specify the field scope and format for data extraction, to adapt to different types of audit needs. Fine-grained data per wafer or per batch requires variables to preset variables such as batch number, equipment ID, and test date, to enable quick association with corresponding data sources. High-frequency updated production data requires the messages context to retain continuous questioning information within a single batch, while limiting the total length to avoid request timeouts.
Configuration Settings
This configuration is based on FastGPT v4.17.0. Specific settings are as follows:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chatId | Bind to production batch number | Associate full-process quality data for the same batch, avoid confusion across batch audits |
messages | Retain the most recent 20 rounds of dialogue context | Cover full-process questioning for single-batch audits, avoid loss of key clues |
variables | Preset batch number, equipment ID, and test date variables | Quickly associate MES, EL/IV test data for the corresponding batch |
| Maximum context length | 8000–12000 token | Accommodate text content such as process parameters and test reports for a single batch |
| Question Optimization | Enabled | Convert vague quality questions into precise queries that can match the knowledge base |
| Single Recall Upper Limit | Top 10 entries | Cover core documents required for single-batch audits, avoid interference from redundant data |
Three Common Misconfigurations
- Symptom: A connection error prompt appears when calling the model. Cause: Network parameters for model access are not correctly configured, and batch association parameters in
variablesare not bound, leading to abnormal request routing. - Symptom: After the knowledge base returns an EL detection image, the large model does not display the image. Cause: The prompt words do not explicitly require converting image links into displayable formats, and the calling rules for image placeholders are not specified.
- Symptom: Keyword variables such as batch number and test equipment ID cannot be extracted from the dialogue. Cause: The prompt words do not specify the field scope and format requirements for variables, and the automatic extraction configuration for
variablesis not enabled.
How to Verify Proper Configuration
- Initiate multi-turn questioning about process parameters for a single batch, check whether the
messagescontext retains previous batch number and parameter information - Upload an EL detection image and trigger a question, check whether the large model can correctly identify and describe abnormal positions in the image
- Enter a question that includes a batch number and test date, check whether
variablesautomatically extract and associate corresponding data documents - Test the model calling process, check that no connection error prompts appear, confirm that network configuration and model access are normal
The values provided are common starting points and should be measured against the reader's own samples.