RAG isn't one architecture, it's a technique with several common shapes, each suited to different problems. Here are the seven worth knowing (more exist, these cover most real systems):
- Naive / Standard RAG — embed the query, retrieve the top chunks, drop them into the prompt, generate. Lowest complexity. Fits FAQ bots and simple support chat over clean documentation.
- Hybrid RAG — combines vector search (good at meaning) with keyword search (good at exact matches — IDs, names, product codes). Common in enterprise and e-commerce search, where both matter.
- RAG with Memory — standard RAG has no memory of earlier turns, so "can you explain more about that?" breaks it. This adds a conversation-memory layer on top. Needed for any assistant that has to handle follow-ups.
- Graph RAG — standard RAG treats chunks as independent, which loses relationships between entities. Graph RAG structures data as a knowledge graph instead — entities as nodes, relationships as edges, same idea as Neo4j. More complex to build, but suited to fraud detection, legal work, research — anything where the connections between facts matter as much as the facts themselves.
- Agentic RAG — for questions that need more than one retrieval, like "explain the 2025 gold price hike and compare it to other metals." The system breaks the question into sub-queries, may call other tools (web search, finance APIs), and pulls the answer together across multiple steps.
- Multimodal RAG — standard RAG is text-only. This extends retrieval to images, video, and audio — relevant for medical imaging, security footage, anything where the source data isn't text.
- Self RAG (Self-Reflective RAG) — instead of generating once and returning it, the model drafts an answer, critiques its own draft, and decides whether it needs to retrieve more before finalizing. Slower, but suited to research or regulated industries where a wrong answer costs more than a slow one.
In production, teams rarely pick just one. Combining a couple of these — hybrid plus memory, or agentic plus graph — is normal, not an exception.
That closes out the series: what RAG is, why it beats plain generation, how the pipeline is actually built, and the shapes it takes in practice.