What Data for This Category Looks Like
Data for this category comes from enterprise metrology rooms, metrology management systems of CNAS accredited laboratories, LIMS platforms, external calibration institution certificates, standard equipment, and environmental monitoring data. Daily documents include gauge ledgers, calibration plans, calibration certificates, interim check records, original test records, reference material ledgers, and deviation disposition forms. Data updates trigger on calibration cycles, completion of single tests, or abnormal events. Documents include both structured fields such as instrument number, calibration validity period, accuracy grade, and measurement unit, and unstructured content such as measurement process descriptions in calibration certificates and uncertainty statements. Core fields include associated identifiers like instrument number and traceability number.
How These Characteristics Impact Knowledge Base Retrieval and Recall
Data sources are scattered and formats vary. Retrieval must support both structured field matching and unstructured semantic understanding. Update cycles are irregular. On-demand incremental updates and synchronization based on actual trigger conditions must be supported. Traceability chain associated fields require retrieval to cross-reference multiple records of the same instrument across documents. Isolated recall of single pieces of data must be avoided. Document content links directly to quality judgment. Recall result confidence must meet metrological compliance requirements. Low-correlation results may affect audit accuracy.
Configuration Settings
This configuration is based on FastGPT v4.17.0
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | High-confidence matching is required for calibration certificates and test records, to avoid mixing low-correlation non-metrological documents into retrieval results |
limit | Top 6 entries | Individual metrology documents have long content. Too many recall results will exceed the context window limit and affect large model understanding |
usingReRank | Enabled | For retrieval scenarios mixing structured ledgers and unstructured certificates, reranking optimizes sorting accuracy and improves the recall priority of key records |
datasetSearchUsingExtensionQuery | Enabled | Retrieval scope needs to be expanded via instrument numbers and traceability identifiers to associate and recall full-cycle records for the same metrology instrument |
searchMode | mixedRecall | Balances exact matching of structured fields and semantic understanding of unstructured content, adapting to multiple types of metrology documents |
Three Common Configuration Mistakes
- Retrieval returns only 2 results. This occurs because the
limitparameter is set to top 2 entries, failing to cover all associated metrology records. - Uploaded scanned calibration certificate PDFs display garbled text. This happens because OCR preprocessing is not enabled, so valid text content cannot be extracted from scanned documents.
- The large model cannot reference original text fragments after calling MySQL metrology data. This is because
datasetSearchUsingExtensionQueryis not enabled, and database query results are not synchronized as retrievable knowledge base fragments.
How to Verify Correct Configuration
- Upload a scanned calibration certificate PDF, and check whether retrieval results can extract valid text content such as measured values and uncertainty statements.
- Enter the number of a specific instrument, and check whether ledgers, previous calibration records, and interim check reports for that instrument can be recalled.
- View the retrieval log to confirm that the configurations for
usingReRankanddatasetSearchUsingExtensionQueryare in effect, with no error prompts. - Test entering the number of an expired calibration instrument, and check whether corresponding abnormal records and associated subsequent disposition information can be recalled.
The values provided are common starting points and should be measured against the reader's own samples.