What Data for This Category Looks Like
Daily documents for this category include tooling requirement documents, positioning datum schemes, fixture drawings, gauge operation manuals, acceptance reports, calibration records, and change histories. Data sources include product 3D models and GD&T, process routes, coordinate measuring machine data, on-site trial feedback, and metrology systems. Updates are carried out per individual project, triggered by each product change, with scheduled synchronization of calibration records. Document structures cover structured parameters, geometric drawing descriptions, metrology values, and process texts. Fields include GD&T tolerance values, coordinate measuring machine measurement deviations, calibration cycles, change version numbers, and more. Units include millimeters, micrometers, days, and others.
Constraints Imposed on Vector Models and Indexes
The multi-type structure of category documents requires vector chunking to balance the integrity of text descriptions and structured parameters, avoiding disconnected associated information. The update rhythm based on projects and changes requires indexes to support incremental updates, reducing resource consumption from repeated processing of historical data. Data contains high-precision metrology information such as GD&T tolerances and coordinate measuring machine measurement values, requiring vector models to capture numerical correlations and geometric positioning logic to ensure retrieval results match business needs. The precise recall requirement in tender scenarios requires indexes to support filtering relevant documents by business tags, avoiding interference from irrelevant historical calibration records in retrieval results.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the length of documents such as tooling requirement documents and fixture drawing descriptions, retains complete associations between GD&T parameters and positioning datums |
| Overlap Length | 100–150 characters | Retains associated information across chunks for change histories and calibration records, prevents loss of context after chunking |
| Similarity Threshold | 0.75–0.85 | Filters irrelevant historical calibration records, accurately matches positioning schemes and acceptance reports required for tenders |
| Single Search Limit | Top 15 results | Controls the number of documents returned by retrieval, balances recall coverage and response speed |
defaultToken | 512 | Adapts to the embedding length of structured parameters, ensures numerical information such as GD&T tolerances is fully encoded |
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 Mistakes
- Phenomenon: After Docker container deployment, all vector retrieval results have identical scores. Cause: Business-specific embedding request parameters were not correctly passed within the container, causing all texts to generate identical default embedding vectors.
- Phenomenon: Index construction tasks remain stuck on processing the final batch of documents, with no progress updates. Cause: Incremental index trigger rules do not adapt to the project-based update rhythm, only performing full index operations and not processing newly added change history documents.
- Phenomenon: Hybrid retrieval request response time exceeds 10 seconds, and no valid results are returned for a single search. Cause: The number of recalled documents per search was not limited, causing excessive loading of coordinate measuring machine measurement data and fixture drawing text, exceeding system resource limits.
How to Verify Proper Configuration
- Upload a typical tooling requirement document and fixture drawing, check if the chunked text blocks retain complete associated information between GD&T parameters and positioning datums.
- Initiate a retrieval request, verify that the scores of returned results have obvious differences, with no identical cases.
- Review index construction logs, confirm that newly added calibration records and change history documents are correctly included in the index queue.
- Test the response time of hybrid retrieval, confirm that it meets the threshold set for business needs.
The values provided are common starting points and should be measured against the reader's own samples.