What Data Looks Like for This Category
Business data for semiconductor wafer fabrication primarily comes from MES systems, equipment FDC collection modules, defect detection and metrology equipment, wafer map repositories, process recipe libraries, and experimental datasets. Daily documents include process flow cards, batch run records, equipment recipes, SPC control charts, exception handling sheets, defect atlases, and yield reports. Data updates follow wafer batch production, real-time equipment parameter reporting, and daily report generation. Most fields include nanoscale process parameters, batch unique identifiers, yield values, and defect spatial distribution coordinates.
What Constraints Do These Characteristics Impose on Vector Models and Indexes
High-frequency updates of real-time per-batch wafer data and daily reports require indexes to support incremental synchronization and low-latency refresh. This avoids excessive compute resource usage from batch index rebuilding. Long-process documents across hundreds of process steps and multi-source coupled parameters require chunking strategies adapted to cross-process associated text structures. This prevents splitting that breaks complete process logic. High-precision fields such as nanoscale parameters and defect spatial distributions require vector model embedding dimensions adapted to fine-grained feature extraction. This prevents loss of critical process details. Fast retrieval needs under limited rework windows require index recall efficiency to match the time constraints of bid document and quotation preparation.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-paragraph text length of process flow cards and yield reports, avoids splitting that breaks process logic |
| Overlap Length | 100–150 characters | Covers cross-process associated text boundaries, prevents loss of critical associated information from chunking |
| Similarity Threshold | 0.75–0.85 | Filters low-match device parameters and defect data, adapts to the high-precision matching needs of nanoscale parameters |
| Single Search Limit | Top 10 results | Controls the scale of retrieval results, matches the fast call needs of bid document and quotation preparation |
maxToken | 8000 | Adapts to embedding of long-text process documents, meets the overall indexing requirements of multi-process data |
defaultToken | 2000 | Limits the base length of a single embedded text, balances retrieval accuracy and compute resource consumption |
Three Common Mistakes
- The interface displays vector model access failure, returning status code 500. This occurs when the model channel's API key and access address are not correctly configured, and the access path of the deployed vector model is not adapted.
- The number of retrieval results does not match the configured single search limit, and empty results are returned. This occurs when the index has not synchronized the latest batch run records and defect atlas data, and the incremental update task is not triggered normally.
- The embedded vector dimension does not match the index requirements, and index construction cannot be completed. This occurs when the vector model's embedding dimension is not adjusted according to the fine-grained features of nanoscale parameters, and general-purpose vector model parameters are misused.
How to Confirm the Configuration Is Correct
- Enter the vector model management page of FastGPT v4.17.0, check that the configured model status is Connected, with no error prompts.
- Upload a wafer yield report as a test document, trigger index construction, and check that the index progress log shows no abnormal errors.
- Enter a search term containing nanoscale process parameters, verify that the retrieval results return matching process flow card and equipment recipe documents.
- Test the incremental synchronization function, upload a new batch run record, and confirm that the corresponding document entry is added to the index list.
The values provided are common starting points and should be measured against the reader's own samples.