> RAG_ARCHITECTURES_v1.0
24 Canonical Architecture Families
72 maturity-graded configurations from linear dense retrieval to ColBERT, GraphRAG, and hierarchical multi-agent teams.
Naive Direct RAG
Linear single-step dense retrieval pipeline directly feeding context into generator. Configured for prototype maturity target.
Two-Stage Cross-Encoder Reranked RAG
Dense vector retrieval followed by deep cross-encoder token-level reranking. Configured for prototype maturity target.
Hybrid Sparse-Dense (BM25 + Dense) RAG
Dual parallel retrieval querying lexical inverted index and dense embeddings merged via RRF. Configured for prototype maturity target.
ColBERT Late-Interaction Token RAG
Fine-grained token-level late interaction computing MaxSim similarity across token vectors. Configured for prototype maturity target.
Hierarchical Parent-Child Chunking RAG
Indexes small child chunks for granular search precision while retrieving parent sections for context. Configured for prototype maturity target.
Graph-Augmented RAG (GraphRAG)
Extracts knowledge graph entities and Leiden community clusters for global corpus summarization. Configured for prototype maturity target.
Multi-Query Expansion RAG
Generates diverse alternative query formulations to bridge vocabulary mismatches. Configured for prototype maturity target.
Contextual Conversational Query Rewriting RAG
Resolves coreferences and dialogue history into standalone search queries. Configured for prototype maturity target.
Self-Querying Structured Filter Extraction RAG
Parses natural language into structured relational metadata filters combined with vector search. Configured for prototype maturity target.
Dynamic Propositional Chunking RAG
Decomposes complex text into atomic factual propositions prior to embedding and indexing. Configured for prototype maturity target.
Self-RAG (Self-Reflective Adaptive RAG)
Uses reflection tokens to dynamically retrieve, critique relevance, and verify support. Configured for prototype maturity target.
Corrective RAG (CRAG) with Fallback
Evaluates retrieval confidence and triggers web search or document refinement on low confidence. Configured for prototype maturity target.
Adaptive Query-Routed Multi-Index RAG
Classifies incoming query intent to dynamically route to specialized domain vector or SQL indices. Configured for prototype maturity target.
ReAct Autonomous Tool-Augmented RAG
Interleaves reasoning thoughts and tool actions in an iterative loop until answer convergence. Configured for prototype maturity target.
Plan-and-Execute Agentic RAG
Formulates an explicit multi-step plan before delegating sub-queries to retrieval agents. Configured for prototype maturity target.
Reflection & Self-Critique Agentic RAG
Generates initial responses, evaluates factual consistency, and recursively edits errors. Configured for prototype maturity target.
Hierarchical Supervisor-Worker Multi-Agent RAG
Supervisor agent coordinates specialized domain worker agents across partitioned document corpuses. Configured for prototype maturity target.
Debate & Adversarial Consensus Multi-Agent RAG
Independent adversarial agents challenge retrieved claims to eliminate hallucinated consensus. Configured for prototype maturity target.
Generator-Critic-Auditor Triad RAG
Three-agent architecture separating response drafting, factual criticism, and regulatory auditing. Configured for prototype maturity target.
Long-Context In-Memory RAG with KV Caching
Leverages large LLM context windows with prefix KV cache reuse for low-latency multi-document reasoning. Configured for prototype maturity target.
Edge & On-Device Embedded RAG
Executes lightweight quantized embeddings and vector search fully local on edge/mobile devices. Configured for prototype maturity target.
Multi-Modal Vision-Document RAG
Extracts and embeds visual charts, diagrams, and figures alongside textual paragraphs. Configured for prototype maturity target.
Privacy-Preserving Encrypted Vector RAG
Executes vector similarity search over homomorphically encrypted shards or differential privacy noise. Configured for prototype maturity target.
Regulated Enterprise Compliance-Locked RAG
Strict zero-retention air-gapped pipeline with cryptographic provenance, audit trails, and PII masking. Configured for prototype maturity target.