Definition

Agentic retrieval is the pattern where AI agents use embedding-based search (and often multi-hop / tool loops) to fetch evidence for reasoning, coding, or long-horizon tasks — beyond single-shot RAG.

Key Points

  • Embedding quality directly affects which passages an agent ever sees
  • Stronger embeddings can reduce downstream token spend in search-agent stacks (NVIDIA claims with nemotron-3-ultra search agent)
  • Typical workloads: enterprise RAG, code retrieval, agent memory over long conversations

Sources

Jul 2026: modus context-warehouse seed frames selective context composition as alternative to over-retrieval (2026-08-02-tnw-modus-10m-seed).