What This Category of Data Entails
The data covers the entire packaging process and testing links. Sources include wafer test maps, MES systems, bonding and molding equipment logs, ATE tester data, X-ray/SAM inspection results, and customer specification requirements. Daily documents include packaging process cards, bonding and molding parameter lists, wafer incoming material maps, test program version files, yield reports, and failure analysis reports. Data updates follow production batches, single-chip testing, and daily summaries. Different documents have dedicated fields such as batch numbers, parameter thresholds, test bin values, and failure locations. Field units vary based on document type.
Constraints on Knowledge Base Retrieval and Recall
This category of data comes from multiple sources and has a high update frequency. It updates in real time alongside production batches, single-chip testing, and daily summaries. This creates requirements for the timeliness of incremental knowledge base synchronization. Document types include parameter lists, test maps, reports, and other categories. They contain dedicated fields such as batch numbers, test bin values, and failure locations. This requires the retrieval link to support precise filtering by specific metadata. Core business links need to align wafer maps with physical batches, and require recall of full-link documents associated with the target batch. Due to the serious consequences of errors, the precision of recall results must be ensured to avoid irrelevant documents being included in the context.
Configuration Settings
This document applies to FastGPT v4.17.0. The following configurations are based on the knowledge base retrieval function of this version.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.78–0.82 | Meets the precision matching requirements of this category, filters low-relevance outdated test programs or non-target batch documents |
limit | Top 10 entries | Matches the typical number of associated documents per batch, controls total context length to fit model input limits |
usingReRank | Enabled | Optimizes recall ranking for long or multi-type documents, prevents failure analysis reports from being pushed down by low-relevance parameter lists |
datasetSearchUsingExtensionQuery | Enabled | Expands query semantics for dedicated fields such as batch numbers and test bin values, improves recall efficiency when aligning wafer maps with physical batches |
collectionIds | Specify the collection ID of the corresponding packaging and testing knowledge base | Isolates documents from other categories, avoids mixing irrelevant data from wafer manufacturing or complete machine testing |
Three Common Misconfigurations
- Phenomenon: Retrieval time exceeds 10 seconds, and the interface displays a timeout status code. Cause: The
limitparameter is not adjusted, and too many unnecessary documents are recalled, causing excessive computing resource usage in the embedding and ranking links. - Phenomenon: Non-target batch packaging process cards are included in retrieval results, with empty fields. Cause: The
collectionIdsfilter for the dedicated knowledge base is not configured, causing cross-collection documents to be recalled. - Phenomenon: Associated content from wafer incoming material maps cannot be retrieved. Cause:
datasetSearchUsingExtensionQueryis not enabled, and query semantics for metadata such as batch numbers are not expanded, making it impossible to match batch identifiers in the maps.
How to Verify Successful Configuration
- Upload packaging process cards and failure analysis reports for a single batch, run a query that includes the batch number, and confirm whether the target documents appear in the recall results.
- View retrieval latency, and adjust relevant parameters based on deployed hardware configurations to ensure latency falls within business-acceptable thresholds.
- Check the ranking of retrieval results, confirm that core documents such as failure analysis reports appear in top positions.
- Verify cross-document associated retrieval: input a test bin value, and confirm whether corresponding yield reports and ATE test data are recalled.
The values given here are common starting points and should be measured against the reader's own samples.