Saltar a contenido

28 RAG variants

A working catalogue with brief notes on when to pick each one.

Family Variant When to use
Foundational Naive Baseline; good for small corpora and prototyping
Foundational Advanced Naive + reranker, hybrid search
Foundational Modular Compose your own components
Quality Self-RAG Self-critique loop, retrieval as needed
Quality CRAG (Corrective RAG) Detects bad retrieval and corrects
Quality Adaptive RAG Adjusts strategy per query difficulty
Agents Agentic RAG Multi-step reasoning with tool calls
Agents ReAct Reasoning + acting interleaved
Knowledge graphs GraphRAG KG-backed retrieval
Knowledge graphs KAG Knowledge-augmented generation
Knowledge graphs CAG Cache-augmented generation
Query rewriting HyDE Generate hypothetical doc, retrieve on it
Query rewriting RAG-Fusion Multi-query expansion + RRF
Hierarchical RAPTOR Tree of chunk summaries
Hierarchical FLARE Forward-looking active retrieval
Long context Long RAG Larger chunks, fewer hops
Hybrid Hybrid RAG Dense + sparse + reranker
Conversational Conversational RAG Maintains turn history
Multi-modal Multi-Modal RAG Text + images + tables

The full list (28) lives in RagType enum. Each variant has its own factory builder.