Data Characteristics for This Category
Semiconductor packaging and testing production data sources include wafer test maps, MES systems, bonding and molding equipment logs, ATE tester data, X-ray/SAM inspection reports, and customer specification documents. Data update frequency is divided into three levels based on production batches, single-chip test results, and daily equipment operating status. Daily documents include packaging process cards, bonding and molding parameter tables, wafer incoming material maps, test program version files, yield summary reports, and failure analysis reports. Fields cover process ID, batch identifier, test parameters, failure point coordinates, and more, resulting in multi-source heterogeneous data with non-uniform formats.
Constraints Imposed on Vector Models and Indexing
Multi-source and heterogeneous data types require vector models to support vectorization of both text and chart data. Frequently updated production data needs incremental index construction logic to avoid performance losses from full index reconstruction. Batch-related fields such as batch numbers and test program versions must be bound to the vector index as metadata, to ensure accurate association of multiple document types from the same batch during retrieval. Fine-grained single-chip test data requires reasonable control of chunk length to prevent splitting of key associated information. Daily summarized ledger data needs sufficient context length to support complete report generation logic.
How to Set Configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
defaultToken | 4096 | Adapts to the conventional token length of yield reports and failure analysis reports, ensuring key parameters are not truncated |
| Chunk Length | 800–1200 characters | Adapts to the single-segment text length of packaging process cards and test program version files, avoiding splitting core process information |
| Overlap Length | 100–150 characters | Retains context for cross-chunk associated fields such as batch numbers and test parameters, improving retrieval accuracy |
| Similarity Threshold | 0.75–0.85 | Filters low-match equipment logs and test reports, reducing irrelevant recall results |
| Single Search Limit | Top 10 entries | Controls the number of documents recalled in a single round, adapting to the batch summary requirements of production ledgers |
| Maximum Context | 16000 characters | Accommodates the summary content of multiple associated ledgers, meeting the context requirements for complete report generation |
Four Common Misconfiguration Issues
- The interface displays "Retrieval results are empty". This happens when the wafer batch number is not bound to the vector index as metadata, preventing association between wafer incoming material maps and production ledger data.
- A "Model not selectable" prompt appears during calls. This occurs when API access permissions for the corresponding Embedding model are not configured, or the model service port is not correctly mapped during local deployment.
- An error "Field format mismatch" occurs during index import. This is caused by inconsistent field naming rules for packaging process cards and test program version files, making associated fields unrecognizable during vectorization.
- Retrieval results do not include wafer test maps. This occurs when multi-modal vectorization configuration is not enabled, and image data is not separately stored in the vector index or associated with batch metadata.
How to Verify Correct Configuration
- Upload a single packaging process card, and confirm that the chunked text retains complete process parameters and batch identifiers.
- Submit a retrieval request that includes a specified wafer batch number, and verify that the recall results include all associated documents for the target batch.
- View the vector model operation logs in FastGPT v4.17.0, and confirm that no model call timeout or parameter incompatibility errors are present.
- Import a batch of test production data, and check that the index creation progress matches the preset batch or daily update rhythm.
Note: The configuration values provided here are common starting points. They should be evaluated against the reader's own sample data.