What Data Looks Like for This Category
This segment’s data comes primarily from flow, pressure, temperature, and analytical instrument manufacturers, plus metrology laboratories. Daily 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 and environmental monitoring data, calibration software outputs, laboratory information systems, and customer range requirements. Data updates occur after each calibration for every instrument. Document fields include calibration date, tested instrument model, standard instrument serial number, error curve values, metrology certificate serial number, and more. Units match the measured parameters, such as Pa for pressure and ℃ for temperature.
Constraints for Vector Models and Indexes
This category’s data has a high update frequency, and single data entries link to multiple documents. Indexes must support incremental updates to adapt to document changes after each calibration. Error curves, original calibration records, and similar content hold large amounts of structured and semi-structured numerical text. Vector models must support long text chunking and semantic association of numerical fields. Calibration records are core deliverables, so recall accuracy must be guaranteed. Indexes must strictly match document traceability fields. Data formats vary widely across sources. Unified preprocessing before indexing is required to avoid embedding vector deviations.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Ensures semantic integrity for long texts such as calibration procedures and error curves, avoids breaking critical numerical associations during chunking |
| Overlap Length | 100–200 characters | Preserves contextual links between adjacent chunks, prevents content like error curves and calibration steps from being split apart |
defaultToken | 8192 | Adapts to overall embedding needs for long documents, matches the typical total length of most industrial instrument calibration documents |
| Similarity Threshold | 0.75–0.85 | Filters low-match irrelevant calibration records, ensures targeted recall results |
| Single Search Limit | Top 5 entries | Controls context window size, avoids exceeding model processing limits |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
Three Common Configuration Mistakes
- The interface only displays one vector model, but multiple vector results correspond to a single data set. This happens when version markers are not added for multiple document versions of the same calibration record, leading to repeated indexing of the same data.
- Index tasks stay stuck in the "indexing" status with no progress. This happens when a single error curve document is too long and was not split using the configured chunk length, causing embedding process timeouts that do not complete.
- The locally used vector model is incompatible with the server-side model, leading to deviations in recall results. This happens when matching embedding dimensions are not confirmed, and consistency of vector formats output by the model is not verified.
How to Confirm Proper Configuration
- Upload a typical calibration procedure document, and check if chunked text fragments retain complete calibration steps and numerical information.
- Initiate a query related to equipment inspection, and verify that the number of recall results matches the single search limit setting.
- Compare the embedding dimensions of the local and server-side vector models, and confirm that their values are consistent.
- Check the indexing task logs in FastGPT v4.17.0, and confirm that there are no prompt messages for timeouts or format errors.
The values provided here are common starting points, and should be validated against relevant samples.