What data looks like for this category
Daily documents for sheet metal cutting and bending include unfolding drawings, CNC cutting programs, bending process cards, nesting diagrams, first-piece records, and process flow cards. Data sources cover CAD/CAM software, order BOM systems, sheet metal inventory ledgers, laser cutting machines, bending equipment, and MES systems. Data updates trigger with order creation, daily production, and equipment model changes. Each model change generates new process parameters and drawing files. Document fields include sheet thickness, bending angle, hole position coordinates, leftover material size, and more. Units mostly follow physical measurement standards such as millimeters and degrees.
Constraints imposed on vector models and indexing
Diverse document types include parameter text from vector drawings, G-code from CNC programs, and tabular process card content. This requires vector modeling to adapt to content splitting logic for different formats. High-frequency data updates require indexes to support incremental updates, avoiding performance loss from full reconstruction. Fields include specific physical units. Vector generation must retain features linked to units, preventing misclassification of parameters with different units as similar content. The multi-variety, small-batch production mode causes large fluctuations in data volume. Index sharding and expansion strategies must adapt to dynamically changing data scale.
Configuration settings
This guide is written for FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Sheet metal cutting and bending documents contain continuous process logic and parameters. This length fully retains core content such as bending compensation and nesting rules |
maxToken | 8192 | Adapts to the maximum supported length of most mainstream embedding models, preventing long text from bending process cards from being truncated |
| Similarity Threshold | 0.75–0.85 | Differentiates similar but distinct bending schemes, adapting to process matching needs for multi-variety, small-batch production |
| Single Search Limit | Top 8 entries | Matches the conventional number of similar process documents in sheet metal production, avoiding interference from redundant results |
defaultToken | 1024 | Balances vector generation efficiency and text coverage, adapting to process parameter retrieval for most short texts |
| Reranking | Calibrated based on actual testing | Similarity of sheet metal documents relies on precise matching of process parameters. Reranking rules must be adjusted for the specific business scenario |
Three common configuration mistakes
- Phenomenon: The interface displays "No available channel" or "This token does not have permission to use the model". Cause: The bce-embedding channel is not configured as the default embedding model, or the channel's access key has not been granted usage permission for the corresponding model.
- Phenomenon: The number of retrieval results remains zero. Cause: Chunk length is set too large, causing G-code text from CNC cutting programs to fail proper splitting, preventing valid vectors from being generated.
- Phenomenon: Vector generation tasks time out. Cause: The
maxTokensetting exceeds the maximum supported length of the embedding model, causing text processing to fail.
How to confirm configurations are correctly applied
- View FastGPT embedding model call logs to confirm every vector generation request returns successfully with no error messages.
- Manually upload a bending process card, run a similar document retrieval, and verify the number of returned results matches the single search limit setting.
- Adjust the similarity threshold, observe changes in retrieval result matching accuracy, and confirm the threshold aligns with current business matching requirements.
- Upload a new sheet metal cutting document, check index update records, and confirm the new document has been added to the retrieval scope.
The values given are common starting points and should be measured against the reader's own samples.