All configuration recommendations on this page are compatible with FastGPT v4.17.0.
What the 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 cover product 3D models and GD&T, process routes, coordinate measuring machine data, on-site trial feedback, and metrology systems. Data updates follow project progress, and are executed with each product change and regular calibration. Documents include structured geometric dimensioning and tolerancing parameters, coordinate measurement values, semi-structured operation instructions and change records, with fields covering millimeter-level tolerance values, calibration dates, detection counts and other professional content.
What constraints these characteristics impose on the vector model and index link
This category has diverse data types, including structured tolerance parameters, semi-structured operation instructions and transcribed drawing content. Update frequency fluctuates with project and product changes. This requires vector indexes to support incremental updates and adapt to data chunking for different formats.
Professional terms such as GD&T and positioning datums have high semantic accuracy requirements. Retrieval must filter low-relevance results to avoid process scheme deviations caused by mismatches. In addition, the need to splice long documents such as complete acceptance reports requires the vector model’s context processing capability to adapt to longer input lengths.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Covers the reasonable length of single-block text such as tooling requirement documents and fixture drawing descriptions, avoiding overly fragmented or overly long chunks |
| Overlap Length | 100–150 characters | Retains contextual association for continuous content such as positioning datums and process routes, preventing key information from being split |
maxToken | 8192 token | Adapts to the context splicing requirements of long documents such as complete acceptance reports, and aligns with the input upper limit of general vector models |
| Similarity Threshold | 0.85–0.90 | Meets the high-precision requirements for tooling data, filtering out low-relevance retrieval results |
| Single Search Limit | Top 3 results | Process schemes are unique, and excessive results will increase the screening cost for engineers |
| Query Optimization | Enabled | Corrects colloquial expressions of tooling terms and improves retrieval matching accuracy |
Three common errors
- A large number of irrelevant gauge operation manual contents appear in retrieval results. The cause is that the chunk overlap length is set too small, leading to the context of positioning datums being fragmented and preventing accurate matching results.
- Insufficient retrieval accuracy when using the m3e embedding model. The cause is failure to adjust the chunking strategy for professional terms such as GD&T and coordinate measuring machine data, resulting in incorrect encoding of professional semantics.
- After submitting tooling change data and calling the create training order function, no updated index is generated. The cause is failure to correctly associate data such as change histories and calibration records with the training task, resulting in the index not being updated synchronously.
How to confirm the configuration is set correctly
- Upload a mixed document containing a tooling requirement document and fixture drawing descriptions, and check that the length of the chunked text blocks falls within the preset 800–1200 character range.
- Initiate a test query containing "positioning datum scheme", and check whether the similarity scores of the retrieval results fall within the preset 0.85–0.90 threshold range.
- Initiate two identical process queries, confirm that the response time of the second query is significantly shorter than the first, to verify that the caching mechanism is active.
- Submit updated data for a tooling change history, wait for the training order to complete, then initiate the corresponding query, and confirm that the retrieval results include the updated change content.
The values provided are common starting points and should be measured against the reader's own samples.