RAG pipelines that survive production

A retrieval-augmented system has four moving parts: the corpus, the chunker, the index, and the prompt. Each one quietly degrades in different ways once real traffic hits it.

This is the checklist we run before any RAG pipeline goes live — and again every quarter after.

The corpus is a living thing

Pick a refresh cadence early. Stale corpora are the most common cause of confidently wrong answers, and the easiest one to engineer around with a scheduled re-index.

Chunking is a product decision

There is no universally correct chunk size. There is only the size that gives your retriever enough context to answer the questions your users actually ask. Build a small eval set, then tune.