RAG isn't one thing. It's a family of pipelines, each with different trade-offs around latency, cost, accuracy, and complexity. The RAG Atlas maps ten of the most common retrieval patterns visually, with animated data-flow diagrams you can interact with.
Open the Interactive RAG Atlas →
Ten patterns covered
1. Vanilla RAG — the minimal baseline. Query → embed → vector search → LLM. Fastest and simplest, but brittle on ambiguous queries.
2. Hybrid RAG — BM25 + vector search, fused with Reciprocal Rank Fusion. Best for technical terms, error codes, and proper nouns that semantic search misses.
3. Rerank-first RAG — over-retrieve a wide candidate set, then cross-encode for precision. Significant accuracy improvement at the cost of extra latency.
4. Query Expansion — an LLM generates multiple query variants to cast a wider retrieval net. Rescues vague or vocabulary-mismatched queries.
5. HyDE — Hypothetical Document Embeddings. The LLM drafts a plausible answer first; that answer's embedding retrieves better than the raw question.
6. Conversational RAG — condenses chat history into a standalone question before retrieval. Keeps multi-turn context accurate without stuffing the full transcript.
7. Router RAG — a classifier dispatches queries to the right source (docs, code index, SQL, web). Avoids polluting context by searching the wrong corpus.
8. Iterative / Corrective RAG — retrieve, draft, critique, refine, repeat. A self-correcting loop for complex questions and sparse corpora.
9. Graph RAG — traverses a knowledge graph from query entities, then summarises communities. Captures relationships that chunk-based retrieval structurally misses.
10. Structured RAG — the LLM generates SQL or tool calls for tabular data. Exact answers from authoritative structured sources with no hallucination on numeric facts (when SQL is correct).
Each pattern in the interactive version includes an animated flow diagram, a node-by-node inspector (what it does, why it exists, how it fails, and how to tune it), and a live simulator where you can toggle chunk size, top-k, reranker on/off, and hybrid mode, and see the effects on latency, cost, and accuracy risk in real time.