A RAG pipeline ingests documents, cleans and chunks text, generates embeddings, stores vectors and metadata, retrieves candidates, reranks them, and supplies selected context to a language model. Chunk size, overlap, metadata filters, embedding choice, and reranking all affect retrieval quality, latency, and cost.

Vector similarity is not authorization. Apply tenant and document access filters before context reaches the model. Treat retrieved text as untrusted because it may contain prompt injection. Embeddings and indexes are derived data and need refresh/rebuild plans when source documents or embedding models change.

Evaluate retrieval recall separately from answer quality. Track source grounding, latency, token use, cost per successful task, and stale-document behavior. Provide citations or source snippets when users need to verify answers.