What production RAG actually needs
Retrieval quality, eval gates, and operational harnesses matter more than swapping the model.
Most RAG demos fail the same way in production: the model looks fine in a notebook, then quietly invents citations, drifts on domain jargon, or returns yesterday’s pricing as if it were truth.
The fix is rarely “use a bigger model.” It is usually:
- Retrieval that matches the job — hybrid search, chunking tuned to document shape, and metadata filters that actually encode business rules.
- Eval gates in CI — Context Precision and Faithfulness (or your domain equivalent) blocking merges when quality regresses.
- An operational harness — logging, replay, and human review paths so failures are diagnosable instead of mysterious.
At Traversia we treat RAG as infrastructure: latency budgets, regression suites, and the same care we give to a high-concurrency API gateway. The model is a component. The system is the product.
If you are shipping agents over private corpora, invest in the boundary — retrieval, tools, and evals — before you invest in another prompt rewrite.