LLMOps Checklist for RAG Systems: Fix Pipelines Before Blaming the Model

A practical LLMOps checklist for production RAG systems highlights that hallucinations and runaway inference costs are typically caused by weak operational practices, not the underlying model. Key recommendations include hybrid retrieval combining dense embeddings with sparse BM25, which can raise recall by 15–30%, and a cross-encoder reranker that teams report reduces obvious hallucinations by 50–60%. Prompt version control is advised to ensure deterministic rollbacks, with templates stored in a registry and tied to CI evaluation checks. Automated evaluation using a golden Q&A set should gate deployments, blocking releases on regressions such as a drop greater than 3% in recall. Cost controls like model routing based on query complexity and semantic caching of repeated answers round out the sprint-based implementation roadmap.