What the Data for This Category Looks Like
Core data for this category comes from business workflows of non-standard equipment manufacturers including cabinet chassis and ventilation equipment. Daily documents include expansion drawings, CNC cutting programs, bending process cards, nesting diagrams, first-piece records, and process flow cards. Data sources cover CAD/CAM tools, order BOMs, sheet metal inventory systems, laser cutters, bending machines, and MES systems. Data updates trigger with orders, daily production, or changeover operations. Core fields include bending compensation, sheet thickness, cutting dimensions, bending angles, nesting utilization rate, and more. Some documents are structured parameter text, while others are raw program text exported from equipment.
Constraints These Characteristics Impose on Vector Models and Indexing
The multi-variety, small-batch production model leads to high data update frequency. Indexes need to support incremental updates to avoid full repeated calculations. Diverse document types include structured process parameters, CNC program text, and CAD-transcribed content. Chunking rules must adapt to different formats to prevent core fields from being split apart. Core fields such as bending compensation and cutting dimensions directly affect production quality. Indexes need to prioritize embedding these high-value fields, while controlling chunk length to ensure embedding accuracy. Frequent changeovers and order changes also require index recall speed to meet real-time audit needs, avoiding excessive delays that disrupt production cycles.
Configuration Settings
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the core information length of documents such as bending process cards and nesting diagrams, preventing key parameters in a single chunk from being split |
| Overlap Length | 100–150 characters | Covers core fields that span chunks such as bending coefficient and sheet thickness, ensuring information completeness |
| Similarity Threshold | 0.75–0.85 | Distinguishes similar process parameters from incorrect compensation values, reducing false recall rates during supplier audits |
| Single Search Limit | Top 8 results | Matches the distribution range of core matching results in multi-variety, small-batch scenarios |
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. |
defaultToken | 512 | Adapts to input restrictions of local lightweight vector models, preventing embedding tasks from timing out |
Three Common Mistakes
- Phenomenon: Newly added knowledge base data remains in indexing status for an extended period. Cause: Did not adapt to the high update frequency of this category, did not enable incremental index configuration, leading to repeated full index execution.
- Phenomenon: Custom index configuration produces no valid recall results. Cause: Did not include core production fields such as bending compensation and cutting dimensions in the embedding scope, only processing general text.
- Phenomenon: Vector calculation times out after connecting a local m3e model. Cause: Did not adjust
defaultTokento match the input length limit of the local model, causing embedding tasks to exceed the model's processing upper limit.
How to Confirm Proper Configuration
- Review index task logs, verify that the chunk count for each document matches the expected number for 800–1200 character chunks.
- Input a query that includes bending coefficient and sheet thickness, confirm that the recall results include the corresponding core fields.
- Upload a new first-piece record, confirm that the index task completes within the preset time range.
- Review vector storage usage, confirm that the preset capacity threshold is not exceeded.
The values provided here are common starting points, and should be evaluated against the reader's own samples.