What This Category of Data Looks Like
This category of data mainly comes from CAD/CAM design files, order BOMs, sheet metal inventory ledgers, laser cutter and bending machine operation logs, and MES production systems. Daily output documents include part unfolding drawings, CNC cutting programs, bending process cards, nesting diagrams, first-piece inspection records, and process transfer cards. Data updates are initiated when orders are delivered, daily production schedules run, or each mold is changed. Most fields cover sheet thickness, bending coefficient, hole diameter size, bending angle, and similar parameters. The primary units are millimeters and degrees.
What Constraints These Characteristics Impose on Vector Models and Indexes
This category of data has high update frequency and scattered trigger nodes. Indexes must support incremental updates to avoid resource consumption from full reconstruction. Document types are diverse. They include long process card texts with professional parameters, structured short texts such as CNC code, and Excel-format BOM and inventory data. The appropriate chunk length varies widely across different documents. Some documents contain professional fields like bending coefficient and hole diameter tolerance. Chunking must retain field associations to avoid breaking process logic. The multi-variety, small-batch production mode also causes obvious fluctuations in single-batch data volume. The index recall limit must adapt to query requests of different sizes.
Configuration Settings
The following configurations apply to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Ensures integrity of professional parameters for long texts such as bending process cards and nesting diagrams, and avoids damaging process logic |
| Overlap Length | 50–80 characters | Retains key information such as bending coefficient and process sequence across adjacent chunks, and prevents process logic breaks caused by chunk truncation |
maxToken | 8192 | Adapts to the input limit of Qwen/Qwen3-Embedding-8B models, and supports mixed text lengths of CNC code and process cards |
| Similarity Threshold | 0.72–0.85 | Filters low-match non-process results, and ensures recalled content is strongly relevant to sheet metal cutting and bending process questions |
| Single Search Limit | Top 8 entries | Adapts to single-order process query scenarios for multi-variety, small-batch production, and avoids recalling too many irrelevant documents |
defaultToken | 1024 | Matches vector generation requirements for basic queries, and balances recall accuracy and computing resource consumption |
Three Common Mistakes
- Symptom: Adding a new embedding model configuration with the same name overwrites the existing configuration. Cause: No unique model identifier is configured for different business scenarios, causing conflicts with identically named configuration items.
- Symptom: Query recall results have too coarse granularity, and cannot accurately match process questions for specific processes. Cause: Chunk length is set too large, and no chunking is done to the minimum granularity suitable for process Q&A.
- Symptom: Truncation errors occur during vector generation, with empty results or incomplete fields returned. Cause:
maxTokenis not set to adapt to the model input limit, and long texts are not truncated correctly.
How to Confirm Configurations Are Correct
- View the embedded model configuration list on the platform, confirm all model identifiers are unique and have no duplicate names.
- Upload a single bending process card document, manually check the chunking results, and verify that key process parameters are fully retained in the chunked content.
- Initiate a process-related query, and check whether the number of recalled results and the matching logic of the similarity threshold meet expectations.
- Test vector generation for long texts, confirm there are no truncation errors, and that generated vectors meet model specifications.
The values listed here are common starting points, and should be evaluated against samples relevant to the specific deployment.