What data for this category looks like
This category’s data primarily comes from business outputs of tooling departments at automotive, aerospace, and mechanical assembly enterprises, as well as specialized tooling suppliers. It includes documents such as tooling requirement statements, positioning datum schemes, fixture drawings, gauge operation manuals, acceptance reports, calibration records, and change histories. Data sources cover product digital models and GD&T, process routes, coordinate measuring machine data, on-site trial feedback, and metrology systems. Updates are driven by projects, triggered with each product change, and also paired with regular calibration actions to complete updates. Individual documents typically include clear fields such as geometric parameters, calibration cycles, and change nodes.
Constraints imposed on vector models and indexing
This category’s data mixes unstructured text, structured numerical reports, and engineering drawing descriptions. Significant field differences exist across different documents, which creates requirements for vector model adaptability. The update rhythm driven by projects and changes requires indexes to support incremental updates to avoid repeated processing of full datasets. Content containing high-precision geometric parameters such as coordinate measuring machine data and calibration records requires vector models to accurately capture numerical correlation relationships. Additionally, indexes must support filtering recall ranges by change nodes to ensure that the latest valid data is called during equipment inspection and maintenance Q&A.
Configuration settings
This configuration scheme applies to FastGPT v4.17.0. Specific configuration items are listed below:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | This category’s documents include engineering drawing descriptions, numerical calibration records, and text change histories. This range balances the integrity of information per chunk and the computational load of the vector model |
| Overlap Length | 100–150 characters | Documents such as positioning datum schemes and change histories contain continuously related content. Setting overlap avoids truncation of key related information |
| Similarity Threshold | 0.75–0.85 | Low-match outdated calibration records and non-target fixture schemes must be filtered out to ensure the match between recalled content and current equipment inspection and maintenance requirements |
| Rerank Switch | Enabled | Equipment inspection and maintenance Q&A needs to prioritize matching calibration and acceptance records related to faults. Reranking improves the sorting priority of highly relevant content |
maxToken | 4000–6000 token | Single-round Q&A needs to call associated content from multiple fixture drawings and calibration records. This range covers information requirements for typical scenarios |
defaultToken | Calibrated based on actual measurements | Adapt to the vector model input requirements for mixed structured and unstructured data in this category. Adjust based on actual recall effects |
Three common configuration errors
- Phenomenon: After enabling the Rerank model, the result sorting returned by the online recall test does not change. Cause: The recall switch associated with reranking was not enabled synchronously in the knowledge base index configuration, or the rerank model’s API key was not correctly bound to the corresponding service.
- Phenomenon: When calling the Embedding model via OneAPI, a connection timeout or 403 permission error is returned. Cause: The input format of the Embedding model was not adjusted for the structured numerical data of this category, or the OneAPI interface configuration did not match the calling protocol of the selected Embedding model.
- Phenomenon: Expired calibration records or pre-change fixture schemes are included in the index recall results. Cause: The incremental update rule filtered by change nodes was not configured, and the full index did not filter invalid historical data.
How to confirm the configuration is complete
- Enter the FastGPT knowledge base management page, view the running logs of completed index tasks, and confirm that the chunk length, overlap length, and other configuration items match the set values.
- Launch an online recall test for a typical equipment inspection question, check the sorting and matching degree of the recall results, and confirm that the rerank switch is active.
- Check the vector model call logs, confirm that the actual value of the
defaultTokenparameter matches the configured value, and there are no abnormal error messages. - View the index update records, confirm that incremental updates are only triggered at product change or regular calibration nodes, and no repeated processing of full historical data has occurred.
The values provided here are common starting points and should be measured against the reader's own samples.