Snapshot-tested RAG: reliable, not magic

TL;DR Useful RAG = clean docs + evaluated retriever + constrained prompt. Else, polite hallucination.

20 Mar 2026 · Nicolas Albert

1 min read Reading 0%

Key takeaways

RAG (retrieval-augmented generation) promises « answer from your docs ». Without discipline, it invents credible citations. Three lines of defense.

1. Clean documents

We clean, split into coherent chunks (not mid-sentence), and store in pgvector. A rotten chunk = a rotten source.

We measure precision@k: on a labelled question set, do the right chunks surface? If not, we fix chunking or embedding before touching the LLM.

The instruction forbids answering without retrieved context and requires citing the source. Not « per the docs », but « source #2 ».

RAG isn't magic: it's a pipeline you test like code.

Me contacter