Retrieval-augmented generation: architecture that holds up in production
A working retrieval pipeline is straightforward. Making it reliable enough to trust with business questions is where the engineering is.
ZooBieDoo Infotech Engineering TeamRead
Key takeaways
- Consistent lighting and camera position beat model sophistication.
- Report precision and recall; accuracy hides imbalance.
- Human-in-the-loop delivers value sooner and generates labels.
- Decide on-device versus cloud inference before designing.
Control the capture conditions
The single highest-return investment in an industrial vision project is usually consistent lighting and fixed camera positioning. Variable conditions force the model to solve a much harder problem than the one you care about.
Where conditions cannot be controlled — outdoor inspection, handheld capture — the training data must represent that variability honestly, which means considerably more of it.
Class imbalance is the norm
Defect detection datasets are dominated by good examples. A model that predicts pass every time can appear highly accurate and be useless.
- Report precision and recall separately, never accuracy alone.
- Decide which error is more costly and tune the threshold accordingly.
- Consider anomaly detection where defects are rare and varied.
- Collect defect examples deliberately rather than waiting for them.
Decide the human role early
Full automation is rarely the right first step. Systems that flag for human confirmation deliver value sooner and build the labelled data needed to justify more automation later.
Design the operator interface properly — it determines throughput as much as the model does.
Plan for the edge
Factory and field deployments often cannot rely on connectivity or tolerate cloud latency. That constrains model size and requires a deliberate update mechanism for devices that are intermittently connected.
Decide early whether inference is on-device, on-premise or cloud, because it shapes the architecture and the model choices available.
Deployment and maintenance realities
A vision system that works in a trial frequently degrades in production for physical reasons: a camera drifts out of position, a lens gets dirty, lighting changes seasonally, or a supplier changes packaging.
Build monitoring for the input as well as the output — image brightness, sharpness and framing statistics reveal capture drift before accuracy visibly falls.
Plan for periodic recalibration and retraining, and make sure someone on site owns the physical setup. Vision projects fail in maintenance more often than in modelling.
Frequently asked questions
Fewer than expected when conditions are controlled and transfer learning is used; considerably more when conditions vary. Defect variety matters more than total count.
General object detection handles common categories well. Domain-specific defects almost always need training on your own examples.
Sometimes usable. Resolution, frame rate and mounting stability are the constraints to check before assuming they will do.
