What the Data for This Category Looks Like
Core data for sheet metal cutting and bending scenarios comes from CAD/CAM software, order BOM systems, sheet material inventory ledgers, laser cutting machines, bending machines, and MES systems. Daily documents include unfolded drawings, CNC cutting programs, bending process cards, nesting diagrams, first-piece records, and process transfer cards. Data updates occur per order, daily, or during each production changeover. Document fields include bending compensation parameters, hole position coordinates, sheet material specifications, process numbers, and other details. Common units are millimeters, square meters, and hours.
What Constraints Do These Characteristics Impose on the Vector Models and Indexes Link?
The multi-variety, small-batch production model of this category creates demand for high-frequency data updates. Indexes must support incremental updates and rapid reconstruction. Most documents are parameterized process and drawing content. Chunking must retain key parameter blocks and avoid splitting bending coefficients or hole position association information. Multi-source heterogeneous data formats require indexes to work with different parsing rules. Processing errors have severe consequences, so vector matching must ensure high recall rates and avoid missing key process parameters.
How to Set the Configuration
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the parameter block length of unfolded drawings and process cards, avoiding splitting key process parameters such as bending coefficients |
| Overlap Length | 50–80 characters | Retains bending sequence and hole position association information for adjacent blocks, preventing semantic fragmentation that affects retrieval accuracy |
| Similarity Threshold | 0.75–0.85 | Matches the tolerance range of sheet metal processing, balancing false positive and missed detection risks |
defaultToken | 1024 | Adapts to the single-block text token limit of bending process cards, ensuring complete parsing of bending steps |
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. |
| Single Recalled Entries | Top 6 | Covers the high-frequency retrieval needs of multi-variety, small-batch production, avoiding omission of historical records from the same process |
Three Common Misconfigurations
- Index progress gets stuck at "Indexing" with no further movement. The cause is a single oversized nesting diagram or complete unfolded drawing file, with no chunking parameters configured, leading to vector generation timeout.
- Locally generated vectors cannot be retrieved normally on the server. The cause is use of different vector models locally and on the server, with mismatched embedding dimensions leading to deviation in similarity calculation.
- Query results do not follow the preset knowledge base priority order. The cause is lack of configured knowledge base priority rules, so the system sorts retrieval results by index time by default.
How to Confirm the Configuration Is Correct
- Upload a single bending process card, check if the chunked text retains complete bending coefficients and process sequence information.
- Call the vector generation interface, verify that the returned embedding dimension matches the configured model dimension, and that
defaultTokenandmaxTokencomply with the limits of the selected model. - Upload multiple documents of the same type, check if the index progress completes within a reasonable time without stalling.
- After configuring knowledge base priority, initiate a test query, confirm that results from the specified knowledge base are returned first.
The values provided are common starting points and should be measured against the reader's own samples.