What data for this category looks like
Data sources for this category include wafer test maps, MES systems, bonding and molding production equipment, ATE test machines, X-ray/SAM testing equipment, and customer specification documents. Daily generated documents cover packaging process cards, bonding and molding parameter sheets, wafer incoming material maps, test program version files, yield reports, and failure analysis reports. Data is updated per batch, single-chip test results, and daily intervals. Documents include dedicated fields and units such as equipment number, test bin code, yield value (unit: %), and failure location.
Constraints on knowledge base retrieval and recall
Data for this category is scattered and updated frequently. Real-time indexing synchronization for multi-source heterogeneous data must be supported. Fine-grained data from single-chip tests and batch-level documents require precise matching. Retrieval and recall granularity must be adjustable flexibly. Documents such as packaging process cards and test programs include dedicated fields like equipment number and test bin code. Field-based precise retrieval must be supported. Daily and batch-level updates require the retrieval link to have low latency to avoid disrupting production rhythms. Structural differences across document types require adapting multi-modal retrieval recall logic. Reasonable recall weights must be ensured for different document types.
Recommended Configuration Settings
The following configurations apply to FastGPT v4.17.0:
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category of documents contains a large number of precise parameters. Low-match results must be filtered to avoid risks of mixed batches or incorrect programs |
limit | Top 8–12 entries | Single-batch retrieval needs to cover multiple document types such as process cards and test programs. 8-12 entries balances recall completeness and retrieval efficiency |
searchMode | mixedRecall | This category has both parameterized documents and image-text documents. Hybrid retrieval balances semantic matching and precise field matching |
usingReRank | Enabled | Recalled results must be reordered by business relevance. Core documents such as bonding parameters and test programs must be returned first |
datasetSearchUsingExtensionQuery | Enabled | Can expand relevant production terms based on user queries. This adapts to professional vocabulary expressions in the packaging and testing field |
collectionIds | Collection IDs divided by business domain | Split collections by domain such as packaging processes and test data. This accurately limits retrieval scope and avoids interference from irrelevant data |
Three Common Configuration Errors
- Phenomenon: After the knowledge base completes a full index, the number of search return results is far lower than expected or no matching results are found. Cause: Collections are not divided by business domain, or the
collectionIdsconfiguration is incorrect, leading to incorrect limitation of the retrieval scope. - Phenomenon: Clicking the knowledge base retrieval interface causes continuous loading and timeout. Cause: The
limitparameter is not adjusted, leading to too many low-quality results being recalled and link timeout, orusingReRankis not enabled, leading to delay in the reordering step. - Phenomenon: Retrieval results include outdated test programs unrelated to the current batch. Cause: No document update time-based filtering rule is configured, or the
similaritythreshold is set too low, leading to accidental recall of old parameter documents.
How to Confirm Correct Configuration
- Retrieve the wafer incoming material map for a specified batch, check whether the returned results include the exclusive parameters and equipment number of the corresponding batch.
- Trigger a batch data update, verify that the retrieval link can return the latest documents within the production-allowed latency range.
- Compare the sorting of retrieval results with and without
usingReRankenabled, confirm that core business documents are returned first. - Test retrieval of single-chip test data, confirm that corresponding yield reports and failure analysis records can be matched precisely.
The values provided here are common starting points and should be evaluated against the reader's own samples.