RAG in production: what breaks that tutorials never show
The most expensive failure mode is the one that doesn't look like a failure.
- Published
- 27 May 2026draft
- Category
- AI in production
- Reading
- 10 min
This piece is published as an outline: the statement and the structure are settled, the full text is coming.
The silent failure
When retrieval returns nothing useful the model answers anyway, with total confidence. How to catch it before the customer does.
Chunking decides what can be asked
Fragment size is not an infrastructure parameter: it defines the set of answerable questions.
The expired key
A real case: embeddings that started returning 401 and a bot that kept answering — vague and with no knowledge base — for hours.
Evaluate or believe
A set of questions with known answers is worth more than any temperature tweak.
Write to me
If you are thinking about building something with AI and want to talk it through with someone who has put it into production, write to me. I answer myself.
hola@giron.ai