What's this about?
A retrieval paper (29 upvotes on HF) rethinking how search agents use "relevance." Today's agents use relevance to pick the top-k most relevant documents — but the authors argue that relevance alone can't tell the agent *where* the evidence is inside a document, how to combine pieces, or how to verify it.
What exactly is "agentic search"?
Traditional search returns a ranked list and you read it. Agentic search puts an AI agent in the loop: it issues queries, reads results, decides what to look at next, and assembles an answer — iterating like a researcher. This paper says relevance should guide that whole interaction (localize, compose, verify evidence), not just be a one-shot filter for which documents to grab.
Why is this worth noting?
Retrieval quality is the ceiling on RAG and research agents. Reframing relevance from a ranking score into an interaction signal is a small conceptual shift that could meaningfully improve how agents gather and check evidence.