What the data for this category looks like
Photovoltaic cell manufacturing business data typically comes from silicon wafer incoming materials, MES systems, diffusion furnaces, coating equipment, printing presses, EL and IV testing equipment, and laboratories. Data update cycles cover each batch of incoming materials, per-wafer testing, 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. Fields include batch number, etching depth, conversion efficiency, hidden crack level, sintering temperature, and others. Corresponding units include percentage, micrometers, degrees Celsius, and others.
What constraints do these characteristics impose on knowledge base retrieval and recall
High-frequency data updates require the knowledge base to support real-time incremental synchronization, preventing retrieval results from lagging behind latest production data. Multi-source, heterogeneous document types include structured parameter tables and unstructured detection images and reports. This requires the retrieval system to support both structured field matching and unstructured content understanding. High data volume from per-wafer testing requires reasonable limits on retrieval recall scope and quantity, to avoid context overload. Core business focuses on process window control and defect identification. Retrieval results must prioritize precise content related to conversion efficiency, hidden cracks, and color differences, and must not include irrelevant general manufacturing documents.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the specificity of photovoltaic cell manufacturing process parameters and quality inspection data, avoids interference from low-relevance general manufacturing documents, adjust based on actual testing |
limit | Top 8–12 results | Adapts to the high volume of per-batch and per-wafer testing data, avoids context overload that impacts generation efficiency |
searchMode | embedding | Prioritizes matching structured process parameters and inspection records, aligns with the standardized data characteristics of this category |
usingReRank | Enabled | Re-ranks retrieval results for EL detection images and IV grading data, improves the ranking priority of highly relevant results |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonymous terms related to processes, for example, expanding "texturing parameters" to "texturing etching depth, texture uniformity" |
collectionIds | Collection IDs corresponding to photovoltaic cell manufacturing process, quality inspection, and supplier documents | Limits retrieval scope, prevents mixing in knowledge base data from other categories |
Three common configuration errors
- No retrieval response or timeout: Failure to limit
collectionIdscauses the retrieval scope to cover redundant non-category data, exceeding system processing limits. - Knowledge base ID and chunk ID cannot be copied: Public access permission for the dataset is not configured in FastGPT v4.17.0 or later, and the front-end copy control is not enabled.
- Low retrieval accuracy:
usingReRankis not enabled, and thesimilaritythreshold is set too low, resulting in mixing of irrelevant component packaging documents.
How to verify successful configuration
- View the
collectionIdsconfiguration of the dataset, confirm it only includes document collections related to photovoltaic cell manufacturing processes, quality inspections, and supplier management. - Initiate a test retrieval, input "texturing process window parameters", and check if the similarity scores of returned results fall within the preset range.
- Check if the re-ranking function is active, confirm retrieval results prioritize highly relevant IV grading data or EL detection reports.
- Upload a new batch yield report, test incremental synchronization, and confirm retrieval results include the latest updated data.
The values listed here are common starting points, and should be adjusted based on the specific deployment samples.