05 · Agentic RAG

What you’ll learn

Classic RAG retrieves once. Agentic RAG wraps retrieval in a small reasoning loop: the agent grades the retrieved context, and if it looks weak it rewrites the query and retrieves again. This self-correction dramatically improves answers to vague or poorly-worded questions.

Watch the trace: when the top similarity falls below a threshold, you’ll see a think step rewrite the query and a second retrieval pass.

The loop

flowchart TD
  Q[Question] --> R1[Retrieve]
  R1 --> J{Relevant<br/>enough?}
  J -- yes --> G[Generate answer]
  J -- no --> RW[Rewrite query]
  RW --> R2[Retrieve again]
  R2 --> G

import { agenticRagPipeline, loadCorpus } from '@lib/js';

const corpus = await loadCorpus('knowledge-base', import.meta.env.BASE_URL);
const { answer, trace } = agenticRagPipeline('how does it correct itself', corpus);

// Look for the `critique` step (relevance score vs threshold) and a possible
// `think` step that rewrites the query before a second retrieval.

from _shared.data import load_corpus
from agentic_rag import agentic_rag_pipeline

corpus = load_corpus("knowledge-base")
result = agentic_rag_pipeline("how does it correct itself", corpus)
for step in result["trace"]:
    print(step["kind"], "-", step["detail"])

Patterns you can add

• Relevance grading. Score retrieved chunks and discard weak ones.
• Query rewriting / expansion. Add synonyms or entities to recover recall.
• Decomposition. Split a multi-part question into sub-questions, retrieve for each.
• Loop limit. Always cap iterations so the agent can’t spin forever.

Here a single agent reasons over retrieval. Next we split the work across several specialised agents — a multi-agent system.