What the Data for This Category Looks Like
Photovoltaic module encapsulation quality traceability data sources include battery cell grading records, glass adhesive film backplane batch ledgers, workshop MES systems, string welding and lamination production equipment, EL/IV testers, and others. The data update frequency is that each module and each batch of materials generates a dedicated record, covering documents such as layout drawings, material batch records, lamination process parameters, EL defect images, IV test reports, insulation voltage withstand records, packing lists, and more. Fields include material batch number, string welding temperature and pressure, lamination time, EL crack coordinates, IV test power value, and other business identifiers.
What Constraints These Characteristics Impose on the "Multi-turn Dialogue and Prompt Words" Link
The dedicated data per single module, multi-type document formats, and high-frequency update rhythm of this category impose multiple constraints on multi-turn dialogue and prompt word configuration. First, bind dedicated fields such as batch numbers via variables to accurately filter redundant data across modules and prevent the model from confusing process records of different batches. Second, clearly distinguish the interpretation priority of text-based process records and image-based EL/IV test reports in the prompt words to adapt to different formats of document input. Finally, because defect analysis requires associating data from multiple links, guide users to supplement detailed parameters such as equipment number and test time in multi-turn dialogue to avoid generalized responses.
How to Set the Configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Single Search Limit | Top 6 entries | The associated documents for a single module are usually 5 to 7, avoiding excessive redundant content that interferes with model judgment |
| Maximum Context | 8000–12000 characters | Covers all process and test data for a single module, preventing truncation of critical business information |
| Variable Binding | Bind the batch number field in variables | Accurately filter documents for the corresponding batch and exclude irrelevant data across modules |
| Reference Content Limit | 2000 characters | Matches the average length of a single photovoltaic test report, controlling the amount of output information per dialogue turn |
| Question Optimization | Calibrated based on actual testing | Optimize the decomposition logic of initial questions for specific scenarios such as EL crack detection and power matching |
| Reordering | Enabled | Prioritize recalling process data from core defect links such as string welding and lamination |
Three Common Misconfigurations
- Phenomenon: After increasing the reference content limit to 2000 characters, the model outputs without citing any retrieved knowledge base paragraphs, and directly calls general knowledge for responses. Cause: The prompt words do not explicitly require the model to generate responses based on retrieved content, causing the model to ignore the configured retrieval scope.
- Phenomenon: The component number ID attached to the knowledge base paragraphs is sent to the model along with the content, interfering with analysis logic. Cause: The prompt words do not specify requirements to only extract business content from paragraphs and retain the format requirements of original fields, causing redundant fields to be mixed into the context.
- Phenomenon: Image content such as EL test images cannot be displayed on the dialogue interface, and only text descriptions are returned. Cause: The knowledge base is not configured to support multi-modal document uploads, or the interpretation priority of images is not specified in the prompt words, causing the model to fail to recognize image inputs.
How to Confirm the Configuration is Correct
- Manually enter a known component batch number with crack defects, check if the retrieved documents only include string welding and lamination process records and corresponding EL/IV test reports for that batch, with no irrelevant data from other batches.
- Adjust the reference content limit to 1500 characters, verify that the length of the cited content in the model output does not exceed this threshold, and clearly mark the source document type.
- Upload an EL crack image to the knowledge base, initiate a question including the image description, check if the dialogue interface normally displays the image and returns corresponding interpretation results.
- Bind the same custom model in both the dialogue interface and the simple workspace, verify that the question response logic and retrieval results are consistent between the two.
The values provided here are common starting points and should be evaluated against the reader's own samples.