One small RAG change improved answer quality more than prompt tweaking

Hey everyone 👋

Something interesting we noticed while working on a recent RAG-based AI system:

We spent a lot less time improving prompts than most people would expect.

The bigger improvement came from fixing retrieval quality.

A few things that made the biggest difference:

→ better chunking strategy
→ filtering noisy context
→ reranking retrieved results before generation

Once retrieval improved, hallucinations dropped noticeably and responses became much more consistent.

One thing we’ve learned building AI systems:
better context usually beats more prompting.

Curious—what’s had the biggest impact on reliability in your AI projects so far?