What the data for this category looks like
Data mainly comes from silicon wafer incoming material ledgers, workshop MES systems, operation logs of diffusion furnaces, coating equipment, and printing presses, EL and IV test reports, and laboratory quality analysis data. Update cycles cover each production batch, each individual product inspection, and daily summaries. Daily documents include process cards, diffusion and coating parameter tables, printing and sintering operation records, EL defect detection documents, IV performance grading data, and yield analysis reports. Document fields include batch numbers, equipment operating parameters, detection point coordinates, operation timestamps, and other relevant fields.
What constraints these characteristics impose on the vector models and indexes link
The high-frequency, fine-grained data updates of this category require indexes to support incremental synchronization, and avoid resource consumption and delays caused by full reconstruction. The multi-dimensional document structure at the wafer and batch level requires chunking strategies that match the smallest business unit, to ensure accurate association of process information for a single batch or single product during retrieval. Multi-source and heterogeneous document types and fields require unified field mapping and format standardization before index construction, to prevent field mismatch issues during retrieval. The strong association between process parameters and test results requires vector models to capture semantic associations between numerical parameters and business scenarios, improving matching accuracy for cross-review.
How to set the configurations
These configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the average length of single-batch process documents, avoids mixing content across batches |
| Overlap Length | 100–200 characters | Covers cross-block associations of process parameters, prevents key parameters from being truncated by chunking |
maxToken | 8192 | Adapts to the input upper limit of mainstream vector models, avoids truncation of core process information |
| Similarity Threshold | 0.75–0.85, calibrated via actual testing | Balances precise matching of process parameters and identification of abnormal deviations |
| Single Retrieval Upper Limit | Top 10 results | Covers multiple types of documents associated with a single batch, avoids excessive retrieval results interfering with review |
defaultToken | 512 | Serves as the base length for default chunking, adapts to semantic extraction of general process parameter text |
Three common misconfigurations
- Phenomenon: An "undefined model must match" error occurs when a non-preconfigured vector model is selected. Cause: The API key and permission for the corresponding model have not been added in the system configuration in advance, or the model input format does not match the chunking rules preset by the system.
- Phenomenon: Semantic retrieval scores are high, but irrelevant process parameters appear in cross-review results. Cause: Exclusive field mapping has not been performed for the process fields of photovoltaic cells, and the vector model has captured irrelevant semantic associations from non-core parameters.
- Phenomenon: A single batch of data generates multiple unrelated vector results. Cause: Unified chunking length and overlap rules have not been set, or multi-source data has not been bound with a unified batch number field.
How to confirm the configuration is correctly applied
- Import a single batch of process documents and test data, perform a retrieval test, and check whether the recalled results cover the core business fields of this batch.
- View the vector model call logs, confirm that all configured models can normally return embedding results without abnormal errors.
- Adjust chunk-related parameters, compare retrieval matching results under different configurations, and confirm that the parameter values match the length characteristics of the business data.
- Test retrieval results for multiple sets of associated data, confirm that associated fields such as batch numbers can be matched correctly.
The values provided are common starting points and should be measured against the reader's own samples.