What This Category of Data Looks Like
Semiconductor packaging and testing scenario business data sources include wafer test maps, MES systems, bonding and molding equipment, ATE test machines, X-ray/SAM inspection equipment, and customer specification documents. Daily documents include packaging process cards, bonding and molding parameter tables, wafer incoming material maps, test program version files, yield reports, and failure analysis reports. Data update frequency covers each production batch, each chip test link, and daily synchronization. Document fields include equipment parameter number, test bin value, yield proportion, failure type, and other items. Units involve dedicated electronic manufacturing parameters such as voltage, current, time, and batch number.
Constraints Imposed on Knowledge Base Retrieval and Recall
The multi-source and heterogeneous data sources of this category impose multiple constraints on knowledge base retrieval and recall. Retrieval tools must support multiple file formats including drawings, parameter tables, and test reports. The high-frequency update rhythm requires support for refreshing recall results by batch or in real time, to avoid using expired packaging parameters or yield data. Documents contain dedicated electronic manufacturing fields such as test bin value, bonding pull force parameter, and batch serial number, which require accurate matching of field association relationships to prevent parameter misalignment caused by fuzzy retrieval. Parallel production of multiple batches requires limiting the retrieval scope to the production batch data corresponding to the current bidding project, to avoid interference from cross-batch data.
Configuration Settings
These configurations are developed based on FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matching dedicated fields such as packaging parameters and test bin values requires high accuracy to avoid mixing low-correlation results |
limit | Top 8 entries | Bid documents need to cover multiple types of parameters, and a small number of accurate results can support content preparation while avoiding redundant information interfering with typesetting |
searchMode | fullTextRecall | The query examples emphasize identifiers or specialist terms; test this mode with representative data, not as a product default. |
usingReRank | Enabled | Sort recalled documents of multiple types by relevance, prioritizing yield reports or process cards related to the current bidding project |
datasetSearchUsingExtensionQuery | Enabled | Automatically supplement packaging and testing-related terms for vague expressions in bidding requirements to improve retrieval accuracy |
collectionIds | Bound knowledge base collection per project | Limit the retrieval scope to the batch data corresponding to the current bidding project, preventing interference from cross-batch data |
Three Common Misconfigurations
- Phenomenon: A 504 timeout status code is returned during retrieval, or logs show retrieval time exceeds 60 seconds. Cause: The
collectionIdsparameter is not limited to the current project's knowledge base collection, and the retrieval scope covers all historical data, leading to overly long recall time. - Phenomenon: Retrieval results do not include packaging process cards or yield data in the knowledge base, returning empty results or unrelated documents. Cause:
datasetSearchUsingExtensionQueryis not enabled, and dedicated terms are not supplemented for vague expressions in bidding requirements, leading to a mismatch between retrieval terms and knowledge base content. - Phenomenon: Knowledge base import tasks get stuck during training, or an error indicating failed download or server error is prompted when exporting the knowledge base. Cause: Packaging and testing documents are not split by batch when importing, and the single-batch import data volume exceeds the system processing limit, or cross-environment knowledge base permissions are not configured correctly during migration.
How to Verify Proper Configuration
- Perform a simulated retrieval, enter the keywords of the current bidding project, and check whether the recalled results include the packaging parameters and yield data of the corresponding batch.
- View the retrieval logs to confirm that the
similaritythreshold andlimitnumber of matching results meet expectations, with no abnormal timeout records. - Test the retrieval restriction for cross-batch data, enter keywords of other batches, and confirm that documents from unrelated batches are not recalled.
- Check the knowledge base import task status, confirm that no entries are stuck during training, and the export function can be triggered normally.
The values given here are common starting points and should be measured against the reader's own samples.