What data looks like for this category
Data for this category comes from multiple sources. These include product drawings and 3D models, CAPP/CAM systems, machine tool collection terminals, tool management systems, MES platforms, and measurement and inspection tool data. Daily documents cover process routes, operation cards, tool lists, CNC programs, first-piece inspection records, patrol inspection forms, and measuring tool calibration records. Fields include operation number, tool model, cutting parameters, dimensional tolerances, inspection results, and more. Common units are millimeters, revolutions per minute, and meters per minute. Data updates have no fixed cycle. Updates trigger with each work order, single-piece machining, sampling inspection, and tool replacement.
What constraints these characteristics impose on the vector models and indexing link
Data sources for this category are scattered and have diverse formats. The data includes both structured process parameters and inspection results. It also includes unstructured CNC program text and 3D model annotations. Multi-modal vector processing logic must be supported. Data updates have no fixed cycle. Updates trigger with work orders, single-piece machining, sampling inspections, and tool replacements. Incremental indexing must be supported. This avoids time-consuming full reindexing operations. Documents contain precise fields such as tool model, dimensional tolerance, and cutting parameters. Vector chunking must retain key process-related information. This avoids breaking process logic during splitting. Real-time data collected by machine tools and static ledger data must have differentiated indexing priorities. This ensures that process-related data is recalled first during retrieval.
How to set configurations
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Documents in this category include long texts such as CNC programs and process routes. Excessively long chunking will lose local associations, while excessively short chunking will disrupt process logic. This aligns with the default chunking rules of FastGPT v4.17.0 |
| Overlap Length | 100–150 characters | Process connection information between operations must be retained to avoid incomplete process logic during retrieval due to chunk breaks |
defaultToken | 1024 token | Aligns with the vector encoding length of common process parameters and program texts in this category, to avoid truncation of key information |
| Similarity Threshold | 0.75–0.85 | This category requires precise matching of fields such as tool model and tolerance requirements. A threshold that is too low will introduce irrelevant results, while a threshold that is too high will miss valid associated data |
| Single Search Limit | Top 8 results | Multi-dimensional association of process ledgers must be balanced. Too many single recalls will exceed context limits, while too few will fail to cover all associated operations |
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
- The first common issue: After importing production ledger data, custom tool model and operation number fields are not indexed. No corresponding content appears in search results. The cause is that the range of custom fields to be extracted is not specified in the import configuration. Only basic text content loads by default.
- The second common issue: The specified embedding model cannot be found in the model selection interface. A configuration error prompt pops up. The cause is that the API access parameters for the corresponding model are not configured in system settings. Or the current FastGPT version does not support the interface format of this model.
- The third common issue: After vectorizing ledger documents containing CNC program screenshots, only text fragments return during retrieval. No image-related results appear. The cause is that multi-modal vector processing configuration is not enabled. Vectorization rules for image data are not configured separately.
How to confirm the configuration is correct
- Upload a single operation card document. Check the index preview interface. Confirm that custom fields such as tool model and tolerance value have been correctly extracted and vectorized.
- Initiate a search containing tool model keywords. Verify the field matching status of returned results. Adjust configuration items to ranges that meet business requirements.
- Upload a document containing CNC program screenshots. Verify whether search results include text information associated with images. Confirm that the multi-modal configuration is effective.
- Check system operation logs. Confirm that there are no authentication errors or timeout prompts for vector model calls. Verify that configured values for
defaultTokenandmaxTokenmatch the set parameters.
The values provided here are common starting points and should be measured against the reader's own samples.