· 1 min read
Every week a business team shows us a RAG prototype that answered five questions correctly in a demo and now hallucinates in front of customers. The gap between "works on my documents" and "works in production" is where most AI budgets go to die.
The demo trap
A vector search over a few PDFs will look magical. The problems arrive with scale and messiness: duplicate documents, stale versions, tables that embed badly, and questions whose answers span three sources.
Our pre-ship checklist
- Chunking matched to the corpus. Legal contracts, support tickets and product specs need different strategies. One-size chunking is the most common silent failure.
- Retrieval evaluation before generation evaluation. If the right passage isn't in the context window, no prompt will save you. We measure recall on a golden set first.
- Grounded citations. Every answer links to its source passages. If the model can't cite it, the UI says "I don't know" — which users trust more, not less.
- Freshness pipeline. Documents change. An ingestion pipeline with deletion handling matters more than the embedding model choice.
- Cost ceilings. Per-query budgets enforced in code, not in hope.
What this means for your build
Ship the boring parts first: ingestion, evaluation, observability. The impressive demo takes a weekend; the system your team can rely on takes discipline. That's the part worth paying for.