RAG is “retrieval augmented generation”
Augmented because we attach the external context into the context window with the LLM
ppl say its dead but its actually not because LLMs always need external context which is obviously not trained within the model weights
RAG is just the natural way to provide external information when it is needed.
just using vector search will not work
- vector search is a tool
- rag is a pattern using that tool (+ others)
- rag can use smarter retrieval too: keyword, metadata filters, hybrid, rerankers, etc
- vector-only rag ≠ good rag
want to build RAG that doesn’t suck?
- hybrid retrieval (vector + keyword + filters)
- good chunking strategy
- reranking or scoring
- prompt engineering or agent-based orchestration
- evals on hallucination & relevance
Transclude of RAG-2025-09-23-10.00.07.excalidraw
references:
https://maven.com/p/569540/i-don-t-use-rag-i-just-retrieve-documents