How we built the RAG pipeline that doesn't hallucinate, boring engineering decisions that mattered

RAG is everywhere. RAG that doesn't hallucinate is rarer.

The boring engineering decisions that actually mattered for us:

- Chunking strategy by document type, not one-size-fits-all

- Hybrid search (dense + sparse) with reranking, not pure vector

- Source attribution required in every response, not optional

- 'Don't know' as a first-class answer, not a fallback

- Per-source freshness tracking so the agent knows when data is stale

The dramatic stuff (model choice, embedding model) mattered less than the chunking and retrieval boring work.

What's the unglamorous engineering decision that improved your RAG most?