Vector vs vectorless RAG is the wrong debate
Most RAG debate compares implementations, how a system works, not retrieval strategies, what it is trying to do and why. So the list of methods keeps growing (naive RAG, GraphRAG, PageIndex and the next X-RAG), and teams keep switching methods and retuning hundreds of knobs. I’ve seen teams waste months this way before they find what is actually wrong with their search. One core retrieval problem is simpler than the jargon: localization, narrowing a million documents down to the few that likely hold the answer, efficiently.
Takeaways
- Vector search localizes in one flat sweep (embed, score everything, keep the top K). A vectorless method like PageIndex uses a table of contents with summaries and narrows down from the top. Same goal, different route.
- Embeddings, summaries and LLM reasoning are implementation choices. The question to ask is which localization strategy fits your problem.
- Localization matters beyond search. Keeping an agent's context small and relevant is the same problem, and sometimes plain grep is enough.
Have a system that needs a second opinion?