What data for this category looks like
Source data comes from design documents, standard instrument and environmental monitoring data, calibration software, laboratory information systems, and customer range requirements. Updates follow the cadence of a single device’s individual calibration. Daily documents include assembly and commissioning sheets, calibration procedures, original calibration records, error curves, metrology certificates, and software version logs. Each document contains fields such as instrument model, calibration date, standard instrument number, error value, measurement range, and qualification determination. Units include metrology-specific units such as pascals, degrees Celsius, and milliliters per minute.
What constraints these characteristics impose on knowledge base retrieval and recall
Data is updated in single-calibration units, and the size of individual data entries varies widely. Documents such as error curves may contain large amounts of time-series data. Retrieval must accurately match key identifiers such as instrument model and calibration date to avoid incorrect results from fuzzy matching. Metrology documents have strict requirements for field accuracy. Retrieval must prioritize matching core business fields to ensure precision. Data from different sources has inconsistent formats, so unified field mapping must be completed before inclusion in the retrieval scope. Audit-critical information such as traceability chains and environmental conditions must also be accurately retrievable.
How to set configurations
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Requires accurate matching of key fields such as instrument model and calibration date to avoid retrieving irrelevant documents with low similarity |
limit | Top 8 entries | The number of documents related to a single calibration is limited. Retrieving too many will increase context redundancy, while retrieving too few will fail to cover all calibration records required for a complete audit |
searchMode | embedding | Requires accurate matching of metrology documents’ semantics and fields; vector retrieval is prioritized to ensure retrieval accuracy |
usingReRank | Enabled | Long documents such as error curves and calibration procedures require re-ranking to filter the most relevant content fragments and improve retrieval precision |
datasetSearchUsingExtensionQuery | Enabled | Can automatically supplement keywords such as instrument model and measurement range based on user questions, adapting to the precise retrieval needs of audit scenarios |
collectionIds | Specify the collection ID of calibration documents for this category | Isolate non-calibration-related documents to avoid cross-category retrieval interfering with audit results |
Three common misconfigurations
- Retrieval results have very low matching with the query, returning a large number of irrelevant general calibration documents. This occurs when a reasonable
similaritythreshold is not set, or whendatasetSearchUsingExtensionQueryis not enabled to supplement key retrieval terms. - Retrieval times out or returns no results. This occurs when the
limitparameter is not restricted, resulting in the retrieval of too many long time-series error curve documents that exceed the system’s processing limit. - Retrieved documents do not include calibration records for the target instrument. This occurs when
collectionIdsare not specified, causing the retrieval scope to cover instrument documents from other categories and leading to mixed results.
How to confirm configurations are correctly set
- Enter a query that includes a specific instrument model and calibration date, then check whether returned results only include documents related to that instrument.
- Review retrieval logs to confirm that
usingReRankanddatasetSearchUsingExtensionQueryhave been properly enabled. - Adjust the
similaritythreshold to verify whether changes in the number of retrieved results across different thresholds align with expectations. - Verify that
collectionIdsonly bind calibration document collections for this category, avoiding cross-collection retrieval.
The values provided here are common starting points and should be evaluated against the reader’s own samples.