What this category’s data looks like
This category’s data originates from formulation systems, internal mixers, Mooney viscometers, vulcanizing machines, and laboratory testing equipment. Daily generated documents include:
- Formulation sheets, with rubber species ratios and additive dosages, using units phr and kg
- Weighing records, with feeding batch numbers and weights, using units kg
- Mixing curves, with torque and temperature readings, using units N·m and ℃
- Vulcanization process cards, with temperature and time settings, using units ℃ and min
- Finished product performance reports
Data updates occur per raw material truck, per rubber batch, per product mold run. Most fields include clear physical units and batch identifiers.
What constraints these characteristics impose on knowledge base retrieval and recall
High-frequency batch-level updates require retrieval and recall to support incremental synchronization and time-based filtering. This prevents use of expired process parameters or weighing records. Mixed document types, including time-series curves, text records, and test reports, require retrieval strategies that cover both structured parameters and unstructured text. Physical units attached to fields must be a required condition for retrieval matching. This prevents invalid recall caused by unit mismatches. Cross-upstream-downstream batch traceability requirements require retrieval to associate data across raw material, mixing, and vulcanization stages. This avoids information silos.
How to configure settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the vector retrieval logic of FastGPT v4.17.0, aligns with the correlation strength requirements for rubber category process parameters, and filters low-correlation results |
limit | Top 8 | Documents associated with a single rubber batch typically do not exceed 6. Retaining redundancy covers multi-dimensional traceability needs for defect analysis |
searchMode | mixedRecall | Matches both formulation text keywords and vector features of process parameters, covering multiple document types including mixing curves and process cards |
usingReRank | Enabled | Time-series features of mixing curves can interfere with vector similarity ranking. Reranking corrects result priority, aligning with defect analysis logic |
collectionIds | Bind collection IDs for mixing process library, vulcanization record library, and finished product inspection library | Limits retrieval scope, excludes irrelevant content such as administrative documents, and improves retrieval accuracy |
datasetSearchUsingExtensionQuery | Calibrate based on actual measurements | Enable upstream and downstream batch extension queries as needed, adapting to cross-stage quality traceability requirements |
Three common configuration mistakes
- Symptom: Retrieval returns results where the
similarityvalue exceeds 10000, with all results redundant and duplicated. Cause: No reasonable similarity threshold set for the rubber category’s parameter characteristics, using default similarity calculation logic from general scenarios. - Symptom: After uploading documents such as formulation sheets and vulcanization process cards, no index entries are generated in the backend, and the file status shows unindexed. Cause: No available channel for the m3e vector model added in platform configurations, preventing normal invocation of vector generation services.
- Symptom: The workflow throws a "Cannot convert undefined or null to object" error during runtime. Cause:
collectionIdsnot specified in retrieval configurations, no knowledge base collection for the target rubber category bound, leading to an empty retrieval dataset that triggers a type conversion error.
How to confirm the configuration is correct
- Upload a mixing curve CSV file and a vulcanization process card. Wait for indexing to finish, then retrieve "160°C, 10 minute vulcanization". Verify that the
similarityvalues of returned results fall within the 0.75–0.85 range. - View the knowledge base collection binding configuration. Confirm that
collectionIdspoints to the three dedicated collections: mixing process library, vulcanization record library, and finished product inspection library. - Trigger a defect analysis retrieval for "under-cure". Check that the number of returned results is under 8, and that the results include weighing records, mixing data, and vulcanization process cards for the corresponding batch.
- After enabling
usingReRank, retrieve "abnormal mixing temperature" again. Check that result ranking prioritizes temperature offset data recorded by the internal mixer.
The values provided are common starting points, and should be validated against the reader’s own samples.