What data in this category looks like
Flow, pressure, temperature, and analytical instrument manufacturers, as well as metrology laboratories, provide data for this category. Core documents include assembly and commissioning sheets, calibration procedures, original calibration records, error curves, metrology certificates, and software version logs. Data sources cover design documents, standard instrument monitoring data, calibration software outputs, laboratory information systems, and customer range requirements. New records are generated each time an instrument is calibrated, so update frequency aligns with individual calibration cycles. Individual calibration records include fields such as instrument number, calibration date, environmental parameters, error values, and traceability numbers. Field formats and units across different documents match the characteristics of the instrument category.
What constraints do these characteristics impose on knowledge base retrieval and recall
Documents in this category are mostly a mix of structured and semi-structured documents generated during a single calibration. The high-frequency update feature requires the retrieval system to support incremental synchronization, to avoid data lag that reduces review efficiency. Dispersed multi-source data sources require the retrieval system to balance structured field matching and unstructured semantic understanding, to ensure that key information such as traceability chains and error calculations is accurately retrieved. The number of calibration-related documents for a single instrument is limited, but reviews must cover all procedures, records, and certificates across the full process. The retrieval scope must be precisely bound to the core dataset, while filtering irrelevant documents to avoid interfering with review judgments. Additionally, calibration results directly relate to customer production processes. The accuracy of retrieval directly affects review pass rates, so retrieval thresholds and sorting logic must be strictly controlled.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Traceability information and error value matching for calibration records have high requirements. A threshold that is too low will introduce irrelevant documents, while a threshold that is too high will miss valid calibration records |
limit | Top 8 | The number of calibration-related documents for a single instrument is limited. Too many results will exceed the context window, while too few will not cover multi-dimensional calibration requirements |
usingReRank | Enabled | Error judgment and traceability chain verification for calibration records require precise sorting, to avoid low-relevance documents interfering with the review process |
searchMode | mixedRecall | There are both structured fields such as instrument number and calibration date, and unstructured error curve descriptions. Hybrid mode balances precise matching and semantic understanding |
datasetSearchUsingExtensionQuery | Enabled | User input review requests may not explicitly mention instrument models. Extended queries can supplement keywords such as range and calibration standards, to improve retrieval coverage |
collectionIds | Bind three datasets: calibration records, calibration procedures, and metrology certificates | These are the core document sources for inspection and process record review, and can filter unnecessary irrelevant data |
Three common misconfigurations
- Symptom: No knowledge base selection option appears when calling tools. Cause: The dataset set for the corresponding category is not bound in the system configuration, or dataset retrieval permissions are not enabled.
- Symptom: The re-ranking model passes testing after deployment, but
usingReRankreturns false for every retrieval result. Cause: The re-ranking model is not bound to the corresponding calibration category dataset, or the deployed re-ranking model does not correctly adapt to FastGPT's interface format. - Symptom: Retrieval results differ significantly between online chat and API calls. Cause: The API call does not carry the correct
collectionIdsparameter, or thedatasetSearchUsingExtensionQueryfunction is not enabled, leading to inconsistent retrieval scope between online chat and API calls.
How to confirm correct configuration
- Check dataset binding: Confirm that the
collectionIdsparameter points to the three core datasets of calibration records, calibration procedures, and metrology certificates, with no irrelevant datasets included. FastGPT v4.17.0 supports bulk binding of dataset sets. - Verify retrieval threshold: Input a review request that includes instrument model and calibration date, check that the
similarityscores of returned results fall within the preset interval, and confirm that no irrelevant documents with abnormally low scores are retrieved. - Test re-ranking function: Trigger retrieval and check the sorting logic of returned results. Confirm that highly relevant documents such as error curves and metrology certificates are ranked first, to verify that the
usingReRankfunction is working correctly. - Compare call results: Use the same query text and parameters to initiate retrieval via online chat and API respectively. Check that the number and relevance of retrieved results are consistent, to ensure that retrieval configurations for API and online chat are synchronized.
The values given are common starting points and should be measured against the reader's own samples.