What this category of data looks like
This category of data includes documents such as packaging process cards, bonding and molding parameters, wafer incoming material drawings, test program versions, yield reports, and failure analysis reports. Data sources cover wafer test maps, MES systems, bonding and molding equipment, ATE test machines, X-ray/SAM testing equipment, and customer specification requirements. Data is updated by production batch, single-chip test result, and daily frequency. Each individual document contains batch identifiers, parameter thresholds, test bin distributions, failure causes, and other relevant content. Units are adjusted based on equipment parameters and customer requirements.
What constraints do these characteristics impose on the citation and traceability process
The data characteristics of this category impose multiple constraints on the citation and traceability workflow. Multi-source and heterogeneous data sources include both structured parameters directly transmitted by equipment and non-standard documents provided by customers. It is necessary to clearly mark data collection equipment and source channels during traceability. The high-frequency update feature requires associating batch numbers and update times with traceability records to ensure cited content is the valid version of the corresponding production batch. Multiple types of documents covering processes, testing, and failure analysis require distinct traceability rules for structured parameters and unstructured text, to avoid mismatched citations across links. Fine-grained single-chip test data must be accurately bound to the corresponding production batch and test link, to prevent the risk of mixed batches or incorrect program usage.
How to configure the settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
quoteMaxToken | 800–1200 token | Adapt to complete citation of long texts such as packaging process cards and bonding parameters, and avoid truncating critical thresholds |
responseData | Enabled | Return original data fields for traceability, matching structured content from yield reports and failure analysis reports |
detail | Enabled | Display batch numbers and update times of cited documents to resolve confusion caused by cross-batch citations |
| Single Retrieval Limit | Top 6 entries | Cover core process documents within a single batch, avoiding interference from redundant data |
| Matching Similarity Threshold | 0.85–0.90 | Calibrated through actual testing, suitable for precise matching of wafer incoming material drawings and test program versions |
Three Common Configuration Errors
- Symptom: Q&A output does not include document identifiers for cited sources, or the cited document version does not match the current production batch. Cause: The
responseDataconfiguration item is not enabled, or the mapping relationship between documents and production batches is not correctly established. - Symptom: The number of context entries displayed on the interface does not match the number sent to the inference interface, for example, 30 entries are displayed on the front end but 310 entries are received by the interface. Cause: The single retrieval limit is not correctly configured, and retrieved data is not filtered by production batch.
- Symptom: Citation source annotation content cannot be removed from Q&A results. Cause: The traceability detail switch is not turned off, or the configuration item is not switched to non-traceability mode.
How to Verify Correct Configuration
- Submit a query containing process parameters or batch keywords, and check whether the output result includes source identifiers and batch information of the corresponding documents.
- View the request logs of the inference interface to verify whether the number of retrieved documents matches the configured single retrieval limit.
- Modify the values of configuration items to verify whether the cited content in Q&A results changes with configuration adjustments.
- Use the FastGPT v4.17.0 platform to test cross-batch queries, and confirm that the cited document version matches the target production batch.
The values provided here are common starting points and should be measured against the reader's own samples.