Types of RAG and which one fits your documents

The main types of RAG are naive (standard) RAG, hybrid RAG, reranked RAG, contextual retrieval, graph RAG, agentic RAG and self correcting RAG. Most are the same search then answer loop with one part upgraded, so the real question is which part your documents need. For policies, manuals and contracts, hybrid search plus a reranker covers most needs; graph RAG suits questions about a whole archive, and agentic RAG suits questions that span several sources.

Every type here is defined in a paper or in official framework docs; we quote no benchmark of our own.

Where do the names come from?

A 2023 survey by Gao and colleagues, "Retrieval-Augmented Generation for Large Language Models: A Survey", sorts systems into three stages. Naive RAG is the "Retrieve-Read" loop: index, retrieve, generate. Advanced RAG adds steps before retrieval, such as rewriting or expanding the question, and after it, such as reranking and compressing what was found. Modular RAG treats each step as a swappable module, with routing between sources and loops instead of one straight pass.

Hybrid, reranked and contextual RAG are advanced RAG with one part named. Graph and agentic RAG are modular. Our glossary entry on RAG defines each briefly; this post is about fit.

Naive RAG: the baseline

Split documents into chunks, embed each one, embed the question, put the closest few chunks into the prompt and generate. The survey lists the weak points: retrieval misses relevant passages or returns irrelevant ones, the model can still add content the passages do not support, and near duplicate passages crowd the prompt.

Naive RAG works when people ask in the same words the documents use: an FAQ or a short staff handbook. It is the cheapest to build and run, with one search and one model call per question. Its quality depends mostly on how you split the files, covered in RAG chunking strategies.

Hybrid, reranked and contextual RAG: better search, same loop

Hybrid RAG runs keyword search (BM25) next to vector search and merges the two result lists, so exact strings such as part numbers, GST section numbers and error codes are found even when embeddings blur them. How to merge the lists is in our post on hybrid search with BM25 and vectors.

Reranked RAG adds a second pass. The Sentence Transformers docs describe a retriever that returns, for example, 100 candidates, then a cross-encoder that reads the question and each candidate together and scores how relevant it is. That is more accurate than comparing separate vectors, and too slow to run over a whole collection, so it only reorders the shortlist. The index does not change; each question takes a little longer.

Contextual retrieval, described by Anthropic in September 2024, has a model write a short note placing each chunk within its document, prepended before embedding and keyword indexing. On their datasets, the share of questions whose right chunk was missing from the top 20 fell from 5.7% to 3.7% with contextual embeddings, 2.9% adding contextual BM25, and 1.9% adding a reranker. The cost lands at indexing time: one model call per chunk, repeated when documents change.

Graph RAG: questions about the whole collection

Edge and colleagues at Microsoft Research showed in 2024 that standard RAG answers questions whose answer sits in specific passages, but fails at global questions such as "What are the main themes in the dataset?". GraphRAG has a language model extract a graph of entities and relationships, writes summaries for groups of closely related entities in advance, and answers a global question by combining partial answers from those summaries.

The GraphRAG docs offer global search for whole collection questions, local search for questions about a named entity, and a basic vector search for questions plain RAG handles well. The repository warns that "GraphRAG indexing can be an expensive operation" and advises starting small. It fits audit reports, research archives or years of customer complaints, where people ask what keeps coming up. It is the wrong tool for "what is our notice period?".

Agentic and self correcting RAG

Agentic RAG puts an AI agent in charge of retrieval. A 2025 survey by Singh and colleagues names four patterns: reflection, planning, tool use and several agents working together. In LangGraph's tutorial, the agent decides whether to search or answer directly, grades what it retrieved, and rewrites the question and searches again if the results miss. It fits questions that need a policy document, a CRM record and a price table together. Expect several model calls per question, varying response times, and harder testing, because the path changes per question.

Self correcting RAG grades the retrieval before trusting it. Self-RAG (Asai and colleagues, 2023) trains a model to retrieve only when needed and critique its own output. Corrective RAG (Yan and colleagues, 2024) uses a lightweight evaluator to score retrieved documents and falls back to web search when confidence is low. For company documents, the useful version is simpler: if no passage passes the relevance check, say the documents do not cover the question.

Which type fits your documents?

In what order should you add them?

1. Start with naive RAG and a test set of real questions.
2. If codes, names and numbers are missed, add hybrid search.
3. If the right passage is ranked too low, add a reranker.
4. If chunks lose their meaning out of context, prepend titles and headings, or add contextual notes.
5. Add a relevance check and an honest "not in the documents" reply.
6. Only then consider graph or agentic RAG, and only for the questions that need them.

Change one thing at a time and rerun the test set after each change.

Where to start with your own documents

Our RAG chatbot builds run vector and keyword search together, rerank the combined results, and score every change against a question set written with your team.

If you are unsure which type your documents need, send us a sample of files and the questions people actually ask. Book a consultation.