Devii · AI & ML · 2026-03-10 · 9 min read
RAG Pipelines: Chunking, Citations, And Offline Eval Before Production
Established ML-engineering practices for retrieval-augmented generation: no model hype, just quality controls.
Retrieval-augmented generation combines search over a corpus with a language model. Failure modes are usually **retrieval** failures (wrong chunk, stale doc, bad filters) rather than “the model forgot facts.”
In **2026** product teams increasingly ship RAG behind **eval harnesses**: golden questions, citation checks, and latency SLOs tied to specific embedding and reranker versions. Vendor models change; your evaluation set should be the regression gate.
Enforce **citations** to source spans, log retrieval scores, version the index like **schema migrations**, and treat prompt tweaks as controlled experiments-roll back when offline eval or safety filters regress.