What this category's data looks like
Core data for this category comes from the MES system, bonding and molding equipment, ATE testers, X-ray/SAM inspection equipment, and customer specification documents. Daily generated documents include packaging process cards, bonding and molding parameter sheets, wafer incoming material maps, test program version files, yield reports, and failure analysis reports. Data updates occur per production batch, per individual chip test, and daily summary. Fields include batch numbers, process parameter values, test bin codes, yield values, and more. Units include degrees Celsius, volts, batch numbers, and others. Some documents contain spatial coordinate information in wafer map format.
What constraints these characteristics impose on knowledge base retrieval and recall
The data for this category is multi-source and heterogeneous, covering device logs, test data, drawings, reports, and other types. This requires the retrieval system to support parsing rules adapted to different formats. High-frequency production data needs incremental recall to avoid repeated retrieval of expired historical ledgers. Unique identifier fields such as spatial coordinate fields in wafer incoming maps and test program version numbers require flexible switching between exact and fuzzy matching. Cross-document association needs, such as aligning wafer incoming maps with test data for corresponding batches, require retrieval to associate content across multiple knowledge base collections. It also requires filtering irrelevant cross-category documents to avoid disrupting production ledger summary logic.
How to configure settings
Choose a retrieval mode through representative-query testing; configure category, batch, or entity filters separately because the mode itself does not enforce those boundaries.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Threshold for matching precise business fields such as packaging parameters and test bin codes, to avoid low-relevance results being included in production ledger summaries |
limit | Top 10 entries | Limits the number of documents retrieved per call, adapts to the summary requirements of multi-source short texts in production ledgers, and controls context length |
usingReRank | Enabled | Performs secondary ranking on long-text results such as yield reports and failure analysis reports, to improve content relevance for summary reports |
datasetSearchUsingExtensionQuery | Enabled | Expands the retrieval scope to cover original ledger data exported from the MES system and associated content from device logs |
collectionIds | Collection IDs grouped by business domain | Distinguishes knowledge base collections of different types such as packaging process cards and test data, to avoid cross-domain documents interfering with retrieval results |
Three common mistakes
- Symptom: Calling the knowledge base retrieval interface returns an unsupported format error, but the corresponding document link can be downloaded normally as a docx file via a browser. Cause: Document parsing rules are not configured to adapt to business documents in docx format such as packaging process cards and yield reports, or binary stream format was not correctly recognized during upload.
- Symptom: Retrieval results include both wafer manufacturing documents and packaging and testing content, and irrelevant results cannot be filtered. Cause: The
collectionIdsparameter is not configured by business domain, leading to cross-knowledge-base collection retrieval. - Symptom: The number of retrieved results is far lower than expected, and cannot cover that day's test data. Cause: The
limitparameter is set too low, ordatasetSearchUsingExtensionQueryis not enabled, so associated data from device logs is not covered.
How to confirm proper configuration
- Upload a packaging process card in docx format, initiate a retrieval, and confirm that returned results under FastGPT v4.17.0 include this document and have correct format parsing.
- Check the retrieval log to confirm that the values of configuration parameters such as
similarityandlimitmatch the preset values. - Test cross-document associated retrieval by entering "wafer incoming map number + test program version", and confirm that ledger data for the corresponding batch is returned.
- Verify that when
usingReRankis enabled, long-text failure analysis reports have higher sorting priority than low-relevance results.
The values provided here are common starting points, and should be evaluated against your own samples.