What are the different types of RAG?

There are three main variations on plain RAG — RAG Fusion, Hybrid Search, and HyDE — each patching a specific weakness in the basic search-then-generate approach. They don't replace RAG; they change how the retrieval step finds documents before the model ever writes its answer.

What are the three main types of RAG?

Each one fixes a different blind spot in ordinary retrieval:

• RAG Fusion: instead of searching with just your original question, the system asks the LLM to rephrase it several different ways, runs a search for each version, then merges the results and re-ranks them using a method called reciprocal rank fusion. One phrasing of your question might miss a document that a slightly different phrasing would catch, so this covers more angles.
• Hybrid Search: this combines old-school keyword search with modern semantic (vector) search. Semantic search is great at understanding meaning, but it can blur past exact terms like a product code, a model number, or a person's name. Keyword search catches those, so pairing the two covers more ground.
• HyDE (Hypothetical Document Embeddings): the system asks the LLM to write a hypothetical answer to your question first — even one that might be factually wrong — and then searches using that made-up answer's text instead of your original question. A fake answer often sits closer, in meaning-space, to a real matching document than the bare question does, so it retrieves better hits.

What do these RAG types have in common?

None of them replace basic RAG. They're add-ons that change how the retrieval step works before the LLM ever writes the real answer. Each attacks the same failure — the search step pulling back the wrong documents — just from a different angle.

You can even stack them, running hybrid search and RAG Fusion together in the same pipeline.

What about GraphRAG?

There's also GraphRAG, which swaps the plain document index for a knowledge graph — a structured web of entities and the relationships between them. Instead of retrieving loose text chunks, it can follow connections across your data, which helps with questions that need facts joined from several places. It's a genuinely different technique, covered on its own elsewhere.