Traditional RAG just retrieves and summarizes. Agentic RAG “reasons” about how to search.
When a user asks “Find me a creative artist workspace in Hackney under £200”—the agent parses intent, builds filters, calls the right search, and responds conversationally.
Building this used to mean stitching together embedding pipelines, vector databases, and agent frameworks.
With Vector Search 2.0 + ADK, it’s:
- One Collection with auto-embeddings
- One Python function for hybrid search
- One Agent definition
My latest post walks through building a travel agent that searches 2,000 London Airbnb listings—with a runnable notebook to try it yourself: