What data for this category looks like
Business data for metal cutting machining comes from product drawings and 3D models, CAPP/CAM systems, machine tool collection terminals, tool management systems, MES systems, and measuring 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. Fields include operation numbers, tool life, dimensional deviation, inspection results, and other items. Data updates trigger per work order, per workpiece, per sampling batch, or when tools are replaced.
What constraints these characteristics impose on vector models and indexing
Data sources for this category are scattered, with diverse field dimensions. Unified field semantics are required to ensure consistency during vector generation. High-frequency updated work order and tool replacement records require indexes to support incremental triggering, avoiding resource consumption from full index reconstruction. Document types have significant length differences: short-text inspection records and long-text CNC programs need adapted chunking rules. The sequential logic of operation execution and process associations require indexes to retain contextual connections. Chunking that cuts continuous machining links will otherwise reduce audit accuracy. Data precision requirements directly determine vector matching threshold settings, and false matching rates must be strictly controlled.
How to configure settings
These configurations are developed for FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the volume of single-process content in operation cards and CNC programs, matching the scope of single-process information for metal cutting documents |
| Overlap Length | 100–150 characters | Retains process associations between operations, avoiding chunking that cuts continuous machining logic |
maxToken | 4096 token | Adapts to the input upper limit of mainstream vector models, meeting the chunked vector generation requirements for long-text CNC programs |
defaultToken | 256 token | Adapts to the vector generation requirements for short-text inspection records and measuring tool calibration records, balancing accuracy and resource usage |
| Similarity Threshold | 0.75–0.85 | Filters low-matching non-associated records, avoiding misjudgments in inspection audits |
| Single Search Limit | Top 10 entries | Matches the conventional number of operations for machining work orders, controlling the total length of retrieved context |
Three common configuration errors
- Phenomenon: The sequence of operations in retrieval results does not match the actual machining process, and some associated process records are missing. Cause: Chunk overlap length is not configured, or index associations for operation sequence are not retained when automatic deduplication is enabled.
- Phenomenon: When calling the vector model, results generated by the large language model are returned, and no vector retrieval results are obtained. Cause: The call channel for the vector model is not correctly bound, and retrieval requests are incorrectly routed to large language model nodes.
- Phenomenon: The number of returned retrieval results is far lower than expected, or errors related to excessive resource usage occur. Cause: Incremental indexes are not configured based on work order update frequency, and frequent full index tasks consume excessive system resources.
How to verify correct configuration
- Upload a single operation card document. Check if chunked text fragments retain the association between operation numbers and process descriptions. Verify that chunking logic complies with
maxTokenconfiguration requirements. - Submit work order data that includes tool replacement records. Verify that the index triggers after data updates, with no delayed or missing records.
- Adjust the similarity threshold. Check if the matching degree of retrieval results meets business expectations, with no over-matching or under-matching.
- View system logs. Confirm that vector model call requests are not incorrectly routed to other nodes, and comply with configuration requirements.
The values provided are common starting points and should be measured against the reader's own samples.