Architectures
Use these patterns as starting points, then validate retrieval quality, security, cost, and user outcomes for your domain.
A straightforward pipeline that retrieves relevant chunks and passes them to an LLM with the user question.
A course assistant that answers only from lecture notes, syllabus documents, and reading lists.
Use for first prototypes, small knowledge bases, and educational demonstrations.
Link to this patternAdds a reranker after initial retrieval to improve the order and relevance of context passed to the model.
A support assistant that retrieves many candidate passages, then reranks them so the most specific troubleshooting step appears first.
Use when initial retrieval returns many partially relevant chunks.
Link to this patternCombines keyword and vector retrieval to handle exact names, legal references, technical terms, and semantic queries.
A legal monitoring system that must match exact regulation numbers and also understand semantic descriptions of obligations.
Use for technical, legal, academic, or enterprise corpora with exact terminology.
Link to this patternUses metadata such as date, source, role, department, jurisdiction, or document type to constrain retrieval.
A public-administration assistant that retrieves only current procedures for the user's department and language.
Use when sources differ by authority, time, department, language, or access rights.
Link to this patternAllows an agent to plan, retrieve, inspect sources, call tools, and iterate before answering.
A research assistant that searches papers, checks definitions, calls a calculator, and synthesizes a sourced answer.
Use for tasks that require multiple searches, tool calls, or procedural reasoning.
Link to this patternUses entities and relationships to retrieve connected evidence and support synthesis across a knowledge graph.
An institutional knowledge explorer that connects people, projects, policies, departments, and documents.
Use for domains with important relationships such as research, policy, legal, or organizational knowledge.
Link to this patternRetrieves from text, images, tables, scans, diagrams, or media and assembles context for a multimodal or text model.
An archive assistant that searches scanned PDFs, tables, handwritten forms, and image captions.
Use for scanned archives, slide decks, forms, diagrams, and mixed media collections.
Link to this patternRuns models, embeddings, and vector storage locally or in a controlled environment for privacy-sensitive workloads.
A classroom or lab prototype running local documents, local embeddings, and a local model on controlled hardware.
Use for privacy-sensitive experiments, classrooms, and offline prototypes.
Link to this patternAdds governance, access control, monitoring, evaluation, and feedback loops around the RAG pipeline.
A company-wide assistant with access controls, trace logs, evaluation dashboards, source owners, and release gates.
Use when RAG answers affect operations, customers, policy, or regulated decisions.
Link to this patternGraphRAG field guide
GraphRAG is most valuable when relationships, corpus-level patterns, or multi-hop evidence matter. It is not an automatic upgrade over well-tuned hybrid retrieval.
Text remains the primary evidence, while graph links organize chunks, entities, hierarchy, or neighborhoods for better traversal and retrieval.
The graph stores claims or relationships as structured knowledge, often alongside source passages for provenance.
Graph traversal identifies connected evidence while vector or keyword retrieval supplies the source text used for generation and citation.
| Retrieval mode | What it does |
|---|---|
| Similarity-based | Find semantically related nodes or linked passages. |
| Local or entity-centric | Combine a focal entity, its relationships, and supporting text. |
| Global or community-based | Synthesize across community summaries to answer corpus-level questions. |
| DRIFT or exploratory | Use community information to broaden a local search and generate follow-up questions. |
| Multi-hop | Traverse several relationships when no single passage contains the answer. |
Advanced patterns
Each pattern changes a different control point in the system. Compare the mechanism, useful tasks, risks, and minimum controls before adopting one.
| Pattern | Control mechanism | Useful for | Principal risk | Minimum control |
|---|---|---|---|---|
| Agentic RAG | An agent plans searches, selects tools, inspects evidence, and may iterate. | Multi-step research, tasks spanning several systems, and questions requiring adaptive retrieval. | Unbounded loops, hidden tool errors, latency, cost, and unpredictable decision paths. | Tool allowlists, iteration limits, trace logging, source requirements, and task-specific tests. |
| Corrective RAG (CRAG) | A retrieval evaluator judges whether evidence is sufficient and triggers another retrieval action when needed. | Pipelines where the first retrieval is often incomplete, ambiguous, or off-topic. | A weak evaluator may reject good evidence or accept bad evidence and add unnecessary searches. | Calibrated relevance checks, explicit fallback routes, and evaluation of correction decisions. |
| Self-RAG | The model uses reflection or critique signals to decide when to retrieve and assess its own generation. | Research settings exploring adaptive retrieval and generation-time self-critique. | Self-evaluation is not independent verification and may reproduce the generator's blind spots. | External evidence checks, held-out evaluation, and human review for high-impact use. |
| GraphRAG | Graph structure organizes entities, relationships, communities, or paths used during retrieval and synthesis. | Connected knowledge, corpus-level questions, entity exploration, and multi-hop evidence. | Incorrect graph extraction, expensive indexing, stale relationships, and reasoning unsupported by source text. | Source-linked edges, extraction audits, update workflows, and separate retrieval/reasoning evaluation. |