What data for this category looks like
Data for this category comes from instrument design documents, standard device and environmental monitoring data, calibration software exports, laboratory information systems, and customer range requirements. Daily documents include assembly and commissioning sheets, calibration procedures, original calibration records, error curves, metrology certificates, and software version records. Data updates follow a per-instrument, per-calibration cycle. Each document includes fields such as instrument number, calibration date, error value, metrology grade, and standard device number. Units include specialized metrology units such as pascals, degrees Celsius, milliliters per minute, and other metrology-specific units.
What constraints this category's characteristics impose on knowledge base retrieval and recall
This category's data updates frequently, so retrieval and recall systems must support filtering the latest documents by calibration date to avoid referencing expired calibration records. The data includes specialized metrology fields and units, so retrieval must match field accuracy and unit consistency to prevent matching irrelevant range or error data. Multi-source data formats vary, including structured calibration records and unstructured error curves. Retrieval logic must adapt to different content structures to avoid missing key traceability information. Defect analysis requires associating multiple documents, so retrieval upper limits must support calling multiple associated records simultaneously.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.6–0.7 | This category's data includes precise metrology units and fields, so low-correlation results need to be filtered to avoid referencing calibration records with mismatched error values |
limit | Top 10 entries | The number of calibration records and associated documents for a single instrument is limited. Recalling too many entries increases context redundancy. 10 entries total around 4000 tokens, which aligns with the default context limit of FastGPT v4.17.0 |
usingReRank | Enabled | For unstructured error curves and calibration procedures, re-ranking improves semantic matching accuracy and filters incorrectly recalled generic documents |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval keywords to adapt to the exact matching requirements of structured fields such as instrument numbers and calibration dates |
searchMode | mixedRecall | Combines semantic matching and keyword matching to cover retrieval requirements for specialized metrology terms and structured fields |
Three Common Mistakes
- Symptom: Knowledge base search tests return no results, and calibration records for the specified instrument cannot be found. Cause: The
similaritythreshold is set incorrectly, ordatasetSearchUsingExtensionQueryis not enabled, resulting in failure to match structured fields such as instrument numbers. - Symptom: The points consumed per single retrieval exceed expectations, making batch defect analysis impossible. Cause: The
limitparameter's recall count is not restricted, resulting in too many redundant documents being recalled and consuming excessive point quota. - Symptom: Retrieved documents include expired calibration records that cannot be used for traceability of current defect issues. Cause: Retrieval logic sorted by update time is not configured, resulting in recall of calibration data older than the current batch.
How to Confirm Proper Configuration
- Upload a calibration record document, use the instrument number as the retrieval keyword, and verify whether the target document is included in the returned results.
- Adjust the
similarityparameter to 0.65, retrieve keywords related to error curves, and verify whether the matching accuracy of the returned results meets expectations. - Check the update time of the retrieved documents to confirm that the latest calibration records are displayed first.
- Count the token consumption per single retrieval to confirm that it does not exceed the preset context limit.
The values given are common starting points and should be measured against the reader's own samples.