What the Data for This Category Looks Like
Core data for this category comes from wafer test maps, manufacturing execution systems (MES), bonding and molding equipment, ATE test machines, X-ray/SAM inspection equipment, and customer specification documents. Daily stored documents include packaging process cards, bonding and molding process parameter sheets, incoming wafer batch maps, test program version lists, yield analysis reports, and failure physical analysis reports. Data update frequency adjusts based on production batches, individual chip test results, and daily equipment operating status. Each single record contains fields such as wafer number, test bin value, equipment operating duration, and process parameter thresholds. Field specifications flexibly adjust based on document type, with no unified fixed format.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Data for this category comes from multiple production and inspection equipment sources, and update frequency dynamically adjusts with production batches and individual test results. This creates multiple constraints for knowledge base retrieval and recall. Multi-source and heterogeneous document types, including visual wafer maps, process parameter sheets, and test reports, require retrieval logic adapted to unstructured and semi-structured data parsing. Fields have no unified fixed specifications, so flexible keyword matching must be supported, without relying on fixed field retrieval to avoid missing key associated information. At the same time, the accuracy of retrieval results directly affects production compliance. Recall precision must be strictly controlled to prevent serious production issues such as mixed batches and misuse of process parameters caused by incorrect matching.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Semiconductor packaging and testing process parameters and wafer batch matching require high precision. This range filters low-relevance results and avoids mixed batch risks |
limit | Top 10–15 entries | The number of documents associated with a single production batch is limited. This range covers required process parameters, equipment logs, and test data, and avoids context overload |
searchMode | mixedRecall | Adapts to multi-source heterogeneous documents. It matches semantically related process parameters and accurately retrieves fixed identification fields such as wafer numbers |
usingReRank | Enabled | For high-value documents such as failure analysis reports and yield data, reranking improves the precision of the top 3 results and reduces manual troubleshooting costs |
datasetSearchUsingExtensionQuery | Enabled | Supplements derivative questions related to equipment inspection, matches actual user questioning habits, and covers more business scenarios |
collectionIds | Configured separately by production dataset, equipment dataset, and test dataset | Manages data by source, enabling precise recall of documents from corresponding links and avoiding interference from cross-category data |
Three Common Configuration Mistakes
- Phenomenon: Knowledge base recall content is not correctly filled into the <Reference> tag in the System prompt, and the returned result has no associated document reference. Cause: The context splicing configuration for knowledge base retrieval results is not enabled, or the format of the reference tag is not specified in the configuration.
- Phenomenon: The file name list in the knowledge base cannot be exported, and there is no corresponding export option in the interface. Cause: The file metadata storage switch for the dataset is not enabled, or the metadata export permission is not enabled in the system settings.
- Phenomenon: The number of retrieval results is far lower than expected, or equipment inspection records for the corresponding device cannot be matched. Cause: The
similaritythreshold is set too high, filtering some weakly relevant but actually matching equipment logs, orsearchModeis only configured for vector retrieval, and keyword matching for retrieving fixed number fields is not enabled.
How to Verify Proper Configuration
- Submit an equipment inspection query that includes a wafer number, and check whether the returned result includes packaging process cards and test program documents for the corresponding batch.
- View the <Reference> tag in the chat interface, and confirm that the recalled document content matches the business scenario of the query.
- Check the dataset metadata list, and confirm that the file names and upload times of all uploaded files can be viewed.
- Adjust the
similaritythreshold, and verify whether the number of recall results changes under different thresholds to meet business requirements.
The values provided here are common starting points and should be measured against the reader's own samples.