footnote1

Agentic RAG: an AI research agent for your documents. It decides what to search for, searches again when the results are thin, and every claim in its answer links to the exact sentence it came from.

What it does · How it works · Evals · Run it (no API key needed) · Slides (PDF)

---

Why

"Chat with your documents" is the most common thing companies ask an LLM to do. The usual build is a fixed pipeline: search once with the user's words, paste the top results into a prompt, hope. It fails in two ways. A two-part question gets one search and half an answer. And the answer reads well even when one sentence in it came from nowhere.

footnote replaces the fixed pipeline with an **agent**. Claude is given search tools and works like a researcher: it picks the wording, runs a search per part of the question, reads what came back, rephrases and tries again if the results are thin, opens the full page when a passage is cut off, and stops when it has the evidence or is satisfied the documents do not cover it.

It is built around one rule: **an answer you cannot check is not an answer.**