What data in this category looks like
Semiconductor wafer fabrication production ledgers cover multi-source heterogeneous data. Sources include MES systems, equipment FDC collection modules, defect detection and metrology equipment, and process recipe libraries. Daily documents include process flow cards, batch run records, equipment recipes, SPC analysis charts, exception handling sheets, defect maps, and yield reports.
Data update cadences cover per-wafer batch generation, real-time equipment parameter synchronization, and daily summary statistics. Fields include wafer batch numbers, nanoscale process parameters, equipment operating temperatures, defect coordinates, and yield values. Units include nanometers, degrees Celsius, batch numbers, and others.
What constraints do these characteristics impose on the vector model and indexing stage
Multi-source heterogeneous data formats require vector models to adapt to both structured parameter text and unstructured chart descriptions, with unified vectorization processing logic. High-frequency update cadences require indexes to support incremental loading, avoiding full index reprocessing of archived historical data. Tiny differences in nanoscale process parameters require models to retain numerical semantic associations, ensuring accurate matching of parameter drifts for similar processes during retrieval. Massive process ledgers and batch records require index recall efficiency to adapt to high-concurrency queries, avoiding retrieval delays that impact production response speeds.
How to configure the settings
FastGPT v4.17.0 supports adjustments to the following configuration items. The specific configurations are as follows:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
defaultToken | 8192 token | Adapts to long text passages of wafer fabrication process documents, covering complete process parameters and batch records |
maxToken | 16384 token | Processes long ledger reports including defect map descriptions and SPC analyses, avoiding truncation of critical process information |
| Chunk Length | 800–1200 characters | Splits individual process flow cards and batch run records, balancing semantic completeness and index recall efficiency |
| Overlap Length | 100–150 characters | Retains process parameter associations across chunks, avoiding loss of process connection information after splitting |
| Similarity Threshold | Calibrated via actual testing | Must adapt to tiny differences in nanoscale parameters, avoiding false recall of irrelevant batch data |
| Single Search Limit | Top 10 results | Covers associated retrieval needs for multiple processes, while controlling context length |
Three common errors
- Phenomenon: Semantic retrieval returns results with abnormally high similarity scores. Cause: The matching weight of the vector model was not adjusted for the numerical semantics of nanoscale process parameters, resulting in over-amplification of numerical associations.
- Phenomenon: Index tasks remain in a running state with no data written. Cause: Incremental index trigger rules were not configured, and full indexes cannot complete loading under high-frequency updated wafer batch data.
- Phenomenon: Recall results cannot be triggered within the preset retrieval template, while normal question answering works normally. Cause: The retrieval template does not include semantic matching logic for core fields such as wafer batch numbers and process parameters, only matching general keywords.
How to confirm the configuration is correct
- Import a single wafer batch run record and process flow card, check if the split text completely retains process parameters and batch identifiers.
- Trigger an incremental index task, verify if high-frequency updated ledger data can complete index loading within the expected time.
- Submit a retrieval request including process parameters and defect descriptions, check if the semantic matching logic of returned results meets business expectations.
- View the vector model call logs, confirm that the configurations for
defaultTokenandmaxTokendid not trigger truncation errors.
The values provided are common starting points and should be measured against the reader's own samples.