What the data for this category looks like
Data sources for the photovoltaic module encapsulation process include battery sheet grading records, glass adhesive film and backsheet batch ledgers, workshop MES systems, string welding and lamination equipment logs, EL/IV tester outputs, insulation voltage withstand test records, and finished product packing lists. Data updates in real time per individual module and per material batch. Each document includes fields such as layout parameters, material batch number, lamination process parameters, EL crack image annotations, IV test power values, insulation voltage withstand values, and other relevant fields. Field units cover current gear, degrees Celsius, seconds, watts, and other professional measurement standards.
Constraints for Multi-turn Dialogue and Prompt Engineering
Data for this category updates in real time per individual module and material batch. Multi-turn dialogue must strictly track the full-chain ledger for a single batch to avoid cross-module data interference. Fields include specialized process parameters and test results. Prompts must explicitly specify field extraction rules exclusive to photovoltaic encapsulation to prevent confusion with parameters from other electronic manufacturing categories. Unstructured content such as EL crack images and IV test reports requires guiding the model to identify corresponding annotation information during dialogue. Additionally, ensure dialogue context only retains data associated with the current batch. Redundant information will disrupt summary and report generation logic.
Configuration Settings
This configuration is adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chatId | Bind the unique component batch identifier for the current work order | Ensure multi-turn dialogue context accurately links to the corresponding component batch |
variables | Bind material ledger and process record fields for the current component | Precisely call real-time production data for a single batch, exclude irrelevant content |
messages | Concatenate context in the order of the full-process ledger for a single component, no more than 7500 characters | Control context length to prevent model overload and focus on summary logic for the current batch |
| Large model call retry count | 3 times, 2-second interval | Address occasional empty responses from large models and improve call stability |
| Output format validation | Mandatory inclusion of three required fields: IV power value, lamination temperature, EL crack status | Meet core test requirements for photovoltaic encapsulation ledgers and ensure output completeness |
| Token consumption monitoring threshold | Calibrated via actual testing, no more than 12000 tokens per call | Adapt to the number of fields and document length of photovoltaic encapsulation data and control call costs |
Three Common Mistakes
- Phenomenon: Garbled end quotation marks appear in AI dialogue node output, which is automatically corrected to standard quotation marks afterward. Cause: The prompt did not explicitly specify the quotation mark type for output formatting, leading to conflicts between sub-responses that use different format rules.
- Phenomenon: Large models occasionally return empty responses with no output content. Cause: No retry mechanism was configured, context length exceeds the model's supported limit, or the
variablesparameter was not bound correctly resulting in missing data. - Phenomenon: Photovoltaic encapsulation-exclusive ledger fields cannot be called after configuring global variables in the workflow. Cause: The binding scope of global variables was not limited to the current work order, leading to calls to irrelevant data from other batches.
How to Verify Proper Configuration
- Manually enter the batch ID for a single component to trigger the dialogue node, and check if the returned content only includes ledger data for that batch.
- Simulate a scenario where the large model returns an empty response and verify if the preset retry mechanism is triggered, with a clear error prompt returned after 3 retries.
- Configure output format validation rules, intentionally delete required fields, and check if format validation errors are triggered.
- View dialogue logs to confirm that the
messagesparameter only contains context content for the current batch and no redundant cross-batch data.
The values provided are common starting points and should be evaluated against the reader's own samples.