What the data for this category looks like
Photovoltaic cell manufacturing quality traceability data typically comes from silicon wafer incoming material records, workshop MES systems, operating parameters of diffusion furnaces/coating equipment/printing presses, EL defect detection images, IV performance test data, and laboratory yield reports. Data update cycles cover each production batch, per-product inspection, and daily aggregation. Documents include process cards, diffusion and coating parameter logs, printing and sintering records, EL detection images, IV grading data, and yield reports. Fields cover professional parameters such as equipment temperature, defect coordinates, and photoelectric conversion efficiency.
What constraints do these characteristics impose on the model access and configuration phase
The multi-source and heterogeneous data structure requires the model to support both structured parameter parsing and unstructured image recognition, so targeted multimodal-related parameters need to be configured. Frequently updated production data means the context length of a single input may fluctuate greatly. If the context range is not reasonably limited, token limit exceeded errors are likely to occur. Different documents have large differences in length; the content length of a single process card or EL detection image far exceeds that of general scenarios, so the maximum length of a single quote must be configured to avoid a single piece of data occupying too many context resources. Precise matching of professional fields requires prompts to strictly correspond to the proprietary parameter names of photovoltaic manufacturing, otherwise deviations may occur in defect analysis results. For example, the cause of hidden crack defects cannot be correctly associated with diffusion furnace temperature parameters.
How to configure the settings
This document is written based on FastGPT v4.17.0. The following configuration items are optimized for the photovoltaic cell manufacturing scenario:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000-12000 token | Adapts to the integrated context length of single-batch process parameters, EL detection images, and IV data |
vision | Enabled | Used to parse visual defects such as hidden cracks and color differences in EL detection images |
quoteMaxToken | 2000-3000 characters | Limits the reference length of a single process card or test report to avoid context overflow |
temperature | 0.1-0.3 | Ensures accuracy of defect analysis results and reduces unnecessary creative output |
AIPROXY_API_ENDPOINT | Fill in the address assigned by the FastGPT platform | Basic network configuration for completing model calls |
AIPROXY_API_TOKEN | Fill in the key assigned by the FastGPT platform | Identity verification for completing model calls |
Three common configuration mistakes
- Phenomenon: Returns a "context length exceeded" error, or conversation content is forcibly truncated. Cause: The
maxContextparameter is not configured, and multiple batches of process data and EL detection images are imported at one time, causing the context token count to exceed the limit. - Phenomenon: In the problem optimization stage, preset environment variables fail to load normally, and variables referenced in the prompt show as empty. Cause: Environment variables are not correctly bound on the model configuration page, or variable naming does not meet platform requirements.
- Phenomenon: Spaces in process parameters included in the prompt are ignored, causing the model to fail to correctly match equipment operating parameters. Cause: The prompt does not use a fixed format such as parentheses or colons to wrap parameters, and does not adapt to the space recognition logic.
How to verify the configuration is complete
- Upload a single EL detection image and the corresponding batch of process parameters, check whether the model can correctly identify the defect type and associate it with the corresponding process parameters.
- Input context containing multiple batches of data, check whether the
maxContextlimit is triggered, and confirm that the context length configuration meets expectations. - Reference the configured environment variables in the prompt, run a test conversation, and check whether the variables load and take effect normally.
- After switching the vector model, regenerate the knowledge base vector index, test the number and relevance of recall results, and confirm that the configuration takes effect.
The values given are common starting points and should be measured against the reader's own samples.