What This Category’s Data Looks Like
Data for this category comes from product 3D models and GD&T, process routes, coordinate measuring machine (CMM) data, on-site trial feedback, and metrology systems. Daily documents include fixture requirement specifications, positioning datum schemes, fixture drawings, gauge operation manuals, acceptance reports, calibration records, and change histories. Data updates are triggered by project milestones and product changes. New data is also generated during regular calibration cycles. Documents contain fields such as GD&T tolerance parameters, CMM measurement values, and details of positioning and clamping schemes. Some documents include binary attachments linked to engineering drawings.
Constraints Imposed on Vector Models and Indexing
Data characteristics for this category impose multiple constraints on the vector model and indexing workflow. Documents with engineering drawings require multi-modal vector encoding, to align with multi-modal vector model call requirements. Update rhythms, triggered by projects and product changes, require indexes to support incremental updates and versioned management. This avoids re-indexing historical data. High-precision fields such as GD&T tolerances and CMM measurement values require vector models to retain detailed accuracy. This prevents loss of critical positioning information after encoding. Structured metrology data and unstructured operation documents are mixed. This requires unified chunking rules to adapt to content lengths across different formats.
Configuration Settings
Configuration items for FastGPT v4.17.0 are listed below:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the average length of fixture drawing descriptions and acceptance reports in this category, prevents single chunks from containing cross-page drawing annotations |
| Overlap Length | 100–150 characters | Covers cross-chunk critical information for positioning datum schemes and change histories, prevents chunking from breaking continuous tolerance descriptions |
defaultToken | 4096 token | Matches the default input length of most open-source vector models, aligns with the average total token count of single documents in this category |
maxToken | 8192 token | Supports batch indexing scenarios containing multiple calibration records, meets the encoding needs for long texts |
| Similarity Threshold | 0.72–0.80 | Filters low-match irrelevant fixture documents, retains positioning schemes that highly match current process requirements |
| Reranking | Enabled, top 3 results | Re-sorts recall results, corrects tolerance parameter matching deviations that may occur during vector encoding |
Three Common Configuration Errors
- Phenomenon: Embedding model connection to OneAPI returns authentication failure or connection timeout errors. Cause: OneAPI access endpoints and keys are not configured correctly, or the vector model call format does not match the parameters required by OneAPI.
- Phenomenon: After enabling the Rerank model configuration, online recall tests do not return reranked results. Cause: The corresponding Rerank model is not bound during the indexing link, or the number of pre-recall results is less than the minimum required for reranking.
- Phenomenon: Attempts to access Tongyi multi-modal vector models fail. Cause: Custom multi-modal vector model adaptation logic is not added to the system, and the model’s input and output formats are not mapped correctly.
How to Confirm Configuration Is Complete
- Navigate to the vector model management page, verify that connectivity tests for configured Embedding models and Rerank models return successful results.
- Run an indexing test for a single fixture drawing document, confirm that the chunked content matches the key information of the original document as expected.
- Initiate an online recall test, enter a query related to positioning datums, confirm that the returned results show changes in reranked order.
- Check index logs, confirm that incremental updates only process newly added calibration records and change histories, and do not re-index historical data.
The values provided here are common starting points, and should be evaluated against the user’s own samples.