Data Profile for This Category
The data for this category originates primarily from wafer fab MES systems, equipment FDC modules, defect detection and metrology tools, process recipe libraries, and experimental datasets. Daily generated documents include process flow cards, batch run records, equipment recipes, SPC timing charts, exception handling sheets, defect maps, and yield reports. Data updates align with wafer batch transfers, real-time equipment parameter reporting, and daily process ledger synchronization. Documents contain structured numerical fields, visual wafer position maps, and long-sequence run records. Fields include wafer batch identifiers, process parameters, equipment numbers, and defect position coordinates.
Constraints Imposed on Vector Models and Indexing
This category’s data characteristics create multiple constraints for the vector model and indexing workflow. Data sources with varied update cycles require support for incremental indexing and real-time parameter synchronization. This eliminates performance losses from full index rebuilding. Documents with visual wafer maps and SPC timing charts need support for non-pure text vectorization. This ensures the parameter association relationships within charts are accurately captured. Nanoscale-precision process parameters and cross-process document links require vector models to support fine-grained semantic recognition. Indexes must enable multi-dimensional field retrieval to meet cross-module review requirements.
Configuration Settings
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the valid single-section information length of process flow cards and equipment recipes in wafer manufacturing documents. Prevents cross-process semantic breaks |
| Overlap Length | 100–150 characters | Covers process parameter associations across chunks. Prevents chunking from breaking the timing logic of SPC charts |
| Similarity Threshold | 0.72–0.85 | Adapts to subtle difference recognition for nanoscale parameters. Filters irrelevant process document interference |
| Single Search Limit | Top 10 results | Matches the number of associated documents for single-batch wafer review. Avoids redundant retrieval slowing review efficiency |
maxToken | 4096 token | Adapts to long-sequence batch run records. Ensures full embedding into review context |
defaultToken | Calibrated via actual testing | Adapts to input length limits of different vector models. Compatibility with various public and commercial indexing model access channels |
Three Common Configuration Errors
- Symptom: No available channels are displayed after connecting a vector model, or models fail to load under the default group. Cause: The index model enable toggle is not selected on the model supplier configuration page, or the model is not bound to the intended usage group.
- Symptom: The BGE-M3 model deployed via ollama cannot generate vectors. Logs show timeouts or vector dimension mismatches. Cause: The vector model’s port and access path are not configured correctly, or input text length exceeds the model’s
maxTokenlimit. - Symptom: Cross-review retrieval results do not match expectations, including excessive irrelevant non-process documents. Cause: The similarity threshold is set too low, or chunk length does not align with document structure, leading to semantic fragments disrupting retrieval logic.
How to Confirm Successful Configuration
- Access the FastGPT v4.17.0 model management page. Confirm the index model is enabled and bound to the current active business group.
- Upload a representative wafer defect map to trigger the vector generation flow. Review generation logs to confirm no dimension errors or timeouts occur.
- Submit a cross-process process review query. Verify retrieval results include the corresponding process flow cards and exception handling sheets.
- Upload a new batch run record to test incremental index synchronization. Confirm the new data can be properly retrieved and searched.
The values provided above are common starting points. They should be evaluated against the reader’s own samples.