ModelRefs / Naive RAG — AI Glossary
Naive RAG — AI Glossary
The basic retrieval-augmented generation pipeline: chunk documents, embed, store in vector DB, retrieve top-k by similarity, pass to LLM.
Overview
Naive RAG indexes documents as fixed-size chunks, encodes them with an embedding model, stores in a vector database, and at query time retrieves top-k chunks by cosine similarity to the query embedding. Simple to implement but fails on complex queries requiring multi-hop reasoning, synthesis, or precision beyond semantic similarity.
Reference details
| Topic | rag |
|---|---|
| Also known as | basic RAG, simple RAG |
| Last reviewed | 2026-06-24 |
Related terms
How it works
- Documents are split into fixed-size chunks at index time, with no attempt to align the cut to meaning.
- Each chunk is embedded once and stored with its vector, so the index is built before any question is known.
- At query time the question is embedded with the same model — the same embedding space is what makes the comparison meaningful.
- The k nearest chunks by vector similarity are returned, always k, whether or not any of them is relevant.
- Those chunks are concatenated into the prompt and the model answers from them, with no step that could reject a bad retrieval.
Example: Why always-k is the defining weakness
With k = 5 over a corpus of 10,000 chunks, every question returns exactly 5 — the top 0.05% by similarity. For a question the corpus answers well, that is 5 useful chunks. For a question it does not cover at all, it is still 5 chunks, now the 5 least-dissimilar irrelevant ones, handed to the model as context. The pipeline has no mechanism to return nothing, which is why a naive system answers out-of-scope questions confidently rather than abstaining.
Commonly confused with
The baseline pipeline — chunk, embed, retrieve top-k, generate — and the thing every other RAG variant is defined against. Naive is descriptive, not pejorative: it is the correct starting point, and the reason to name it is so that adopting a more elaborate variant is a measured decision against this baseline rather than an assumption that more machinery is better.
When to use it
Reach for it when:
- A first implementation, where it establishes a baseline the variants must beat.
- A homogeneous corpus and questions answerable from a single passage.
- Diagnosing which later addition actually helped, by removing it and comparing.
Reach for something else when:
- Questions need facts from several documents composed together.
- Out-of-scope questions must be declined rather than answered from the nearest chunks.
- The corpus mixes formats where one chunk size suits none of them.
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Naive RAG — AI Glossary.
Frequently asked questions
What is Naive RAG?
The basic retrieval-augmented generation pipeline: chunk documents, embed, store in vector DB, retrieve top-k by similarity, pass to LLM.
Is Naive RAG the same as basic RAG?
Yes — basic RAG, simple RAG are common aliases for Naive RAG.
What concepts are related to Naive RAG?
Closely related concepts include advanced rag, modular rag, retrieval pipeline.