02 · Primer
In one paragraph
RAG gives context, agency adds control. Together they create retrieval systems that can plan, adapt, and self-check instead of following one rigid search pattern.
On this page4 sections
Agentic RAG adds planning and decision-making to retrieval-augmented generation, so systems can choose smarter retrieval paths and verify answers before responding.
How RAG works (visual)
Three stages run in sequence on every query. Base RAG is linear; agentic RAG can branch, retry, and verify before producing final output.
- Retrieve. Fetch relevant chunks from trusted sources based on query intent.
- Augment. Inject the retrieved context into prompt state before generation.
- Generate. Produce answers grounded in evidence instead of model memory alone.
What changes when you add agency
Four control layers sit on top of the base loop. Each one turns a fixed step into a decision.
- Planning layer. Decides what to fetch first, when to reformulate a query, and how deep retrieval should go.
- Adaptive retrieval. Chooses retrieval strategy per question: semantic, keyword, hybrid, or tool-assisted.
- Tool orchestration. Calls calculators, APIs, and domain systems when static documents are not enough.
- Verification loop. Validates claims, cites sources, and retries when confidence falls below threshold.
Reference architecture
Four layers, bottom-up. Each layer has a narrow contract with the layers on either side.
- Interface layer. User question intake. Session context. Response formatting.
- Agent control layer. Planner. Retriever selector. Critic / evaluator.
- Knowledge layer. Vector store. Structured DB. External systems.
- Ops layer. Auth and permissions. Tracing and metrics. Fallback policies.
Keep retrieval sources versioned and observable so answer quality is auditable over time.
Frequently asked
How is it different from standard RAG?
Standard RAG typically follows one fixed retrieval path. Agentic RAG can choose strategies, retry searches, and validate outputs before returning.
Does it eliminate hallucinations?
No system eliminates hallucinations entirely. Agentic RAG reduces risk by grounding outputs in sources and adding verification steps.
Where does it perform best?
Knowledge-heavy tasks with variable question types — support, internal search, policy workflows.
What metric matters most?
Grounded answer rate: the percentage of responses backed by relevant, correct citations and verified tool outputs.
Further reading