What the data for this category looks like
Data for this category comes from formula systems, raw material batch ledgers, internal mixers, Mooney viscometers, vulcanizing machines, and laboratory testing equipment. Daily documents include formula sheets, weighing records, mixing curves, rubber compound quick test reports, vulcanization process cards, mold usage records, and finished product performance test reports. Data updates occur per raw material batch, per rubber compound batch, and per product mold run. Document fields include batch number, feeding weight, mixing temperature, vulcanization duration, finished product tensile strength, and other items. Some data uses time-series curve format.
What constraints these characteristics impose on the vector models and indexes link
The high-frequency update rhythm of this category requires vector indexes to support incremental synchronization, avoiding performance overhead from full index rebuilding. Diverse document formats require retaining associated fields such as batch numbers, mold numbers, and raw material batch numbers during chunking, ensuring recall results can directly link to specific production links. Multiple data types (text records, mixing time-series parameters, finished product test values) require vector models to adapt to mixed-format text conversion. The index dimension must also match the vector length output by the model. Additionally, the rigor of quality traceability requires index results to retain original data source fields, preventing recall results from disconnecting from production batches and reducing defect analysis accuracy.
How to set configurations
The configurations on this page are based on the interface and parameter logic of FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale for this setting |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the semantic integrity of long texts such as mixing curves and process cards, avoiding loss of key production parameters across chunks |
| Overlap Length | 50–80 characters | Retains cross-chunk associated fields such as batch numbers and mold numbers, ensuring semantic continuity between chunks |
| Similarity Threshold | 0.75–0.85 | Matches the traceability accuracy requirements of this category, filtering out low-correlation non-production link data |
defaultToken | 4096 | Adapts to the total token length of single-batch documents, avoiding truncation of core content such as formulas and process parameters |
maxToken | 2048 | Limits the token upper limit for a single recall result, avoiding context overflow overload |
| Single Search Limit | Top 10 results | Controls the number of recall results, balancing traceability efficiency and information completeness |
Three common mistakes
- Phenomenon: After uploading documents such as mixing curves and weighing records, no corresponding index entries appear in the vector database, and the interface shows an "index failed" status. Cause: Source fields such as batch numbers were not retained in the chunking configuration, causing chunked data to fail to match index rules.
- Phenomenon: When calling the m3e model, a "no available channels" prompt appears, and vectors cannot be generated. Cause: The calling key and service address of m3e were not configured according to official documentation, or the usage permission for the corresponding model was not activated.
- Phenomenon: Repeating the same defect analysis query triggers vector database retrieval every time, and no cache is hit. Cause: FastGPT's cache optimization configuration was not enabled, or the cache validity period was set too short to cover high-frequency query scenarios.
How to confirm correct configuration
- Upload a standard formula sheet and mixing curve document, check whether the document entries for the corresponding batch are included in the vector database index list.
- Execute a defect analysis query, verify that the returned recall results include associated fields such as batch numbers and vulcanization parameters.
- View the model call log, confirm that the m3e model call request has successfully generated vectors, with no channel error messages.
- Repeat the same query, confirm that the return speed of the retrieval result meets expectations, and no repeated vector database call records appear.
The values provided here are common starting points and should be measured against the reader's own samples.