This page applies to FastGPT v4.17.0.
What this category of data looks like
Daily data for SMT and electronic assembly includes BOM and coordinate files, line change lists, stencil and reflow oven temperature records, first article inspection reports, SPI/AOI inspection reports, maintenance records and traceability labels. Data sources include customer-provided BOMs, Gerber files, production MES systems, placement machines, SPI/AOI equipment, and reflow oven temperature collection devices. Update frequency follows triggers per work order and each line change; real-time collected oven temperature and defect data is updated in real time. Documents include structured fields such as part number, reference designator, component batch, coordinate parameters, plus unstructured defect descriptions and maintenance records, with fields attached to corresponding units such as millimeters and degrees Celsius.
What constraints these characteristics impose on the vector model and index link
This category has high-frequency, real-time data updates, so indexes must support incremental updates to avoid performance loss from full reconstruction. Document types cover both structured and unstructured content. When splitting documents, retain the integrity of key fields such as part numbers and coordinates to avoid semantic breaks. Multiple fields include professional terminology and units from the electronic manufacturing sector, so vector models must adapt to this professional semantic recognition to ensure the accuracy of retrieval results. Additionally, the data volume per work order fluctuates greatly, so the resource usage of the index must match the production rhythm to avoid disrupting normal operation of the production system.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the length of single-section process descriptions in SMT documents, avoiding damage to key information such as part numbers and coordinates during splitting |
| Overlap Length | 100–150 characters | Retains key related information across chunks, such as process connection content before and after line changes |
| Similarity Threshold | 0.75–0.85 | Distinguishes professional semantics between process documents and exception reports, avoiding false matching of unrelated defect records |
defaultToken | 8192 | Adapts to embedding processing for long documents such as complete BOM tables, aligning with the conventional length of industry documents |
maxToken | 16384 | Supports batch embedding of process data from multiple work orders, avoiding truncation of key content |
| Vector Model Channel | Specialized embedding model for the electronic manufacturing sector | Adapts to professional terminology and data formats in the SMT industry, improving the accuracy of embedding and retrieval |
Three common mistakes
- Phenomenon: Retrieval results include process documents not belonging to the target work order. Cause: Index shards are not associated by work order ID, leading to accidental recall of cross-work order data.
- Phenomenon: Embedding tasks trigger timeout errors. Cause: The
maxTokenparameter is not adjusted, leading to failed embedding after long documents such as complete stencil configuration tables are truncated, triggering a timeout. - Phenomenon: The matching accuracy of defect types in exception reports is low. Cause: A reasonable similarity threshold is not set, confusing defect records with similar semantics but belonging to different processes.
How to confirm correct configuration
- Upload the BOM and SPI report for a single work order, retrieve the corresponding process parameters, and verify that the work order number in the returned results matches the target work order.
- Check the disk usage of the knowledge base, confirming that the storage proportions of original files, split chunks, and embedded vectors meet expectations.
- Test an incremental update of a line change list, verifying whether the index automatically refreshes and can retrieve the latest content.
- Adjust the similarity threshold, compare retrieval results under different thresholds, and select a value that meets business requirements.
The values provided here are common starting points and should be evaluated against samples from the target deployment.