RAG patterns — 7 approaches from vector to agentic

Key Takeaways

Reference catalog of 7 RAG patterns, each with when-to-use, trade-offs, and components.

1. Classic Vector RAG — docs → chunks → embeddings → similarity search → LLM. Fast, cheap, offline-capable. Our stack: FalkorDB + MLX, ~1ms/query.
2. Graph-Boosted RAG — vector search + 1-hop knowledge graph expansion. Finds related docs vectors miss. Our stack: FalkorDB vectors + explicit/tag/semantic edges, ~75% precision.
3. Tree-Based Reasoning (PageIndex) — no vectors. LLM builds hierarchical tree, reasons over it. For 100+ page professional docs.
4. Late Interaction (ColBERT) — token-level similarity, 10x better on exact terms. Heavy but precise.
5. Hypothetical Document Embeddings (HyDE) — LLM generates ideal answer first, embeds that. Great for abstract questions.
6. Hybrid — BM25 (keywords) + vectors + graph, with reciprocal rank fusion. Best of all worlds.
7. Agentic RAG — LLM decides what to search, iterates, self-corrects. Highest quality, highest cost.

Decision matrix: small docs + budget → Classic Vector. Linked docs → Graph-Boosted. Long docs → Tree-Based. Need keywords + semantics → Hybrid. Complex multi-step → Agentic.

Our implementation (solograph): Hybrid approach — FalkorDB vectors + knowledge graph + BM25, with graph-boosted expansion. The LLM Wiki (index.md) sits above all of this as a compiled navigation layer.

Connections

• context-graphs-summary — knowledge graph in RAG = context graphs applied to document retrieval
• privacy-as-architecture — offline RAG with MLX = privacy-first retrieval, no cloud dependency
• solo-methodology — solograph implements patterns 1+2+6 (vector + graph + hybrid)