What Data Looks Like for This Category
This category of data originates from product drawings and 3D models, CAPP/CAM systems, machine tool collection terminals, tool management systems, MES systems, and measurement and inspection tool data. Daily documents include process routes, operation cards, tool lists, CNC programs, first-piece inspection records, patrol inspection forms, and measuring tool calibration records. Data updates trigger with each work order, each workpiece, sampling batch, or tool change. Documents contain structured and unstructured fields such as CNC program code, tool parameters, dimensional deviation values, and inspection results.
Constraints for Vector Models and Indexing
This category has mixed data formats, including CNC program code, structured parameter tables, and natural language description documents. Data is updated frequently with work orders and tool changes. Differences in multi-source data formats require the indexing link to support multi-modal text parsing. Frequent updates demand adaptation to incremental indexing strategies, to avoid resource consumption from full reindexing. Documents include key exception data such as process association information and dimensional deviations. Vector models must support semantic encoding of code text and structured parameters. The maxToken configuration must be set sufficiently to cover long text segments, preventing truncation of critical G-code and process parameters.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the semantic integrity of individual segments of process documents and CNC programs, avoiding cross-process information fragmentation |
| Overlap Length | 100–150 characters | Retains contextual associations between processes, preventing loss of links between datums and clamping plans |
| Similarity Threshold | 0.72–0.85 | Calibrated via on-site testing, filters low-match exception reports and process documents, avoiding mismatched tool wear and dimensional drift data |
| Single Search Limit | Top 10 entries | Matches the single-batch data volume of shop floor patrol inspections and first-piece inspections, ensuring recall results cover typical exception scenarios |
maxToken | 8000 | Adapts to long text processing for CNC programs and process routes, avoiding truncation of critical G-code and process parameters |
defaultToken | 4096 | Provides sufficient margin for vector encoding of basic text segments |
| Reranking | Enabled | Optimizes semantic ranking of exception reports, prioritizing recall of historical processing documents matching the current exception type |
Three Common Misconfigurations
- After importing a text dataset into the knowledge base, the dataset status remains stuck in "indexing". The cause is failure to adapt to frequently updated metal cutting data, using a full indexing strategy without configuring incremental indexing logic, resulting in excessive indexing time for large volumes of CNC programs and process documents.
- Custom vector model requests always trigger large model calls. The cause is failure to correctly bind the dedicated call path of the vector model in channel configuration, confusing and misconfiguring the vector model and LLM channels.
- The semantic similarity value returned by the vector model falls outside the 0-1 range. The cause is failure to adjust the output normalization configuration of the vector model, directly using uncalibrated similarity scores that do not align with FastGPT's filtering logic.
How to Confirm Proper Configuration
- Upload a single CNC program document containing G-code, check whether the indexing completion time meets expectations, and confirm that the incremental indexing trigger logic operates correctly.
- Enter a query describing tool wear or dimensional drift, check whether the similarity scores of recall results fall within the 0-1 range.
- After configuring multi-replica deployment, initiate a batch indexing request, confirm that indexing progress is consistent across all replicas.
- Switch to the custom vector model channel, initiate a vector query request, confirm that returned results only include vector matching data, and no LLM generation logic is triggered.
The values provided are common starting points and should be measured against the reader's own samples.