Building RAG pipelines in production: lessons from 20+ deployments

Chunking decides the ceiling

Retrieval can only return what you split well. Chunk by meaning, such as headings and sections, and keep enough context for each piece to stand alone.

Pick embeddings against your own data

Benchmarks are a starting point. Test two or three embedding models on real questions from your users before you commit.

Don't skip reranking

A reranker over the top results is often the cheapest accuracy gain in the pipeline, and it makes failures easier to debug.