What the Data for This Category Looks Like
Data for this category comes primarily from daily outputs of wafer fab process integration, equipment, yield, and related departments. It includes process flow cards, batch run records, equipment recipes, SPC statistical charts, exception handling sheets, defect maps, and yield reports. Data sources include MES systems, equipment FDC modules, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental data. Update cadences align with each wafer batch production run, real-time parameter reporting, and daily summary archiving. Fields typically include nanoscale process parameters, batch numbers, defect coordinates, and other professional content.
Constraints on Vector Models and Indexes
The ultra-long process chain, high-frequency update properties, and specialized field structure of this category impose multiple constraints on vector models and indexes. Massive process document sets and cross-module coupled process data require indexes to support hierarchical retrieval and context-aware matching. Real-time, batch-based update cadences require indexes to support incremental synchronization, avoiding the time cost of full reconstruction. Specialized fields such as nanoscale process parameters and defect coordinates require vector models to adapt to high-dimensional professional semantic encoding accuracy, while retaining association logic between fields.
Configuration Settings
This configuration is adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Wafer fabrication documents often include continuous process steps. This length preserves process context while avoiding overloading individual chunks |
| Overlap Length | 100–200 characters | Process parameters and steps are linked front to back. Overlap prevents context breaks |
maxToken | 8192 token | Must accommodate long text blocks containing multiple pages of process cards, avoiding truncation of critical parameters |
defaultToken | Set from representative sample documents | defaultToken is the default token count for direct text chunking; tune it on representative source documents. See the v4.17.2 model-field reference. It is not a vector dimension, embedding-model input cap, or retrieval/LLM context. |
| Single Recall Limit | Top 8 entries | Wafer fabrication processes have many associated nodes. Sufficient candidate results must be retained for subsequent filtering |
| Similarity Threshold | 0.72–0.80 | Professional terminology matching requires high precision, avoiding incorrect recall of unrelated process documents |
Three Common Configuration Mistakes
- Knowledge base upload status remains stuck in indexing for extended periods. Cause: Incremental index update rules are not configured. Full index reconstruction for massive wafer batch data takes too long.
- Embedding rate limit errors occur during vectorization. Cause: Thread count is not adjusted. Concurrent requests exceed the call limits of the vector service. Reduce thread count to mitigate this issue.
- Text chunk loss occurs when setting chunk length to 3000 characters. Cause: Long text chunk splitting does not preserve process logic boundaries. Some cross-process associated parameters are incorrectly truncated and cannot be correctly identified.
How to Verify Correct Configuration
- Upload a single typical process flow card. Check that chunking results preserve continuous process step context, with no obvious truncation or splitting errors.
- Submit a query containing specialized process parameters. Verify that the matching accuracy of recalled results meets preset professional scenario requirements.
- View the index update log. Confirm that real-time uploaded batch data completes incremental synchronization within the set time window.
- Test vectorization success rates across different chunk lengths. Confirm that configured chunking rules do not trigger text loss or abnormal errors.
The values provided are common starting points and should be measured against the reader's own samples.