What's this about?
GRASP teaches an agentic retrieval system *how* to search, not just *what*: deciding when retrieval is needed at all, and whether a lexical (keyword) or semantic (meaning-based) search fits the query (HF 5↑).
What exactly is "agentic RAG"?
RAG (retrieval-augmented generation) lets a model look things up before answering instead of relying on memory. Classic RAG does one fixed retrieval; agentic RAG lets the model loop — reason, search, read, search again — like a researcher rather than a student with one allowed library visit. The open problem is search strategy: exact keywords find names and error codes; semantic search finds concepts; using the wrong one silently misses the answer.
Why does it matter?
Search-strategy selection is a daily pain in every production RAG system; policies that learn granularity decisions transfer directly to real assistants and coding agents.