Source - Stop Saying RAG Is Dead

Obsidian Metadata

type	source
url	https://hamel.dev/notes/llm/rag/not_dead.html
authors	Hamel Husain, Ben Clavié
published	2025-07-12

Source: Stop Saying RAG Is Dead (7-part series)

URL: https://hamel.dev/notes/llm/rag/not_dead.html

The 2023 recipe (chunk → single vector → cosine) deserves criticism, but retrieval matters more than ever. LLMs are frozen at training time, and million-token windows don’t change the economics.

1. Ben Clavié: “I don’t use RAG, I just retrieve documents”. Naive single-vector search is what’s dead.
2. Nandan Thakur: modern IR evals for RAG. Classic metrics optimize for the #1 result; RAG needs coverage, diversity, relevance (FreshStack). BEIR winners can fail at RAG.
3. Orion Weller: reasoning retrievers that follow instructions (Rank1 generates relevance reasoning).
4. Antoine Chaffin: late interaction (ColBERT) keeps token-level detail; 150M models beat 7B on reasoning retrieval.
5. Bryan Bischof & Ayush Chaurasia: multiple representations and routing instead of one perfect embedding.
6. Kelly Hong: Source - Context Rot.
7. Jo Kristian Bergum: you probably don’t need a graph DB for GraphRAG; CSV or Postgres works.

→ RAG Evaluation, Is RAG Dead