What Data Looks Like for This Category
Daily documents for this category include process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, yield reports, and more. Data sources cover MES systems, equipment FDC modules, defect detection and metrology equipment, wafer map files, process recipe libraries, and experimental datasets. Data updates run per wafer batch, in real time, or daily. Fields include nanoscale process parameters, batch numbers, defect spatial coordinates, yield values, and more. Some documents are structured tables, some are mixed text and image files, with no unified fixed format.
Constraints Imposed on Vector Models and Indexing Workflows
Multi-source heterogeneous document formats require indexes to support mixed-modal vectorization and retrieval. This enables unified encoding of structured parameters and mixed text-image content. High-frequency update cycles require indexes to support incremental updates. This adapts to small-batch, high-frequency wafer batch data writes. Strong correlation between nanoscale process parameters and defect spatial distribution requires vector models to meet encoding accuracy standards for fine-grained numerical and spatial features. Long process chain data across multiple steps requires indexes to support long-context associated retrieval. This avoids truncation of critical process step association information. Classified production data requires indexes to support on-premises deployment. This ensures no data leakage during retrieval.
Configuration Settings
These guidelines are based on FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the information density of single blocks of process parameters and defect descriptions in wafer manufacturing documents, avoiding truncation of critical nanoscale parameters |
| Overlap Length | 150–200 characters | Covers cross-block process step association information, preventing vector retrieval breaks between adjacent process steps |
| Similarity Threshold | 0.72–0.85 | Matches the feature matching accuracy of defect maps and yield reports, filtering low-correlation batch records |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
| Single Search Limit | Top 10 entries | Covers the number of associated process steps for a single wafer batch, ensuring sufficient process data is retrieved for quality attribution |
defaultToken | Calibrated via actual testing | Adapts to input length limits of different Embedding models, matching the total field count of wafer manufacturing data |
Three Common Misconfigurations
- Phenomenon: Image documents such as defect maps and wafer maps cannot be returned in retrieval results. Cause: Multimodal vectorization configuration is not enabled, or associated retrieval rules for image indexes are not configured.
- Phenomenon: The specified Embedding model cannot be found in the model selection list. Cause: The corresponding model's API key and access address are not correctly configured in FastGPT's model management module, or the model version is not on the supported list.
- Phenomenon: When importing structured batch run records, vector retrieval results for some fields are empty. Cause: Structured field metadata is not retained during chunking, or the index is not configured with vector encoding adaptation rules for structured data.
How to Verify Proper Configuration
- Upload a typical defect map document, execute a retrieval request containing defect feature descriptions, and check if the corresponding image files are included in the returned results.
- View the index update log to confirm that documents for each wafer batch complete vectorization and index construction within the set update frequency.
- Adjust the similarity threshold configuration, observe changes in retrieval result relevance, and confirm the threshold matches the business's retrieval requirements.
- Test long-context process chain retrieval, confirm that returned results do not truncate critical process step association information.
The values provided here are common starting points, and should be evaluated against the reader's own samples.