CRAG: leaderboard

4,409 factual QA pairs across five domains and eight question types, with mock web-search and knowledge-graph APIs, from Meta (2024); LLMs alone scored at most 34% at release, plain RAG 44%.

Source: arxiv.org.

Interactive version: theaggregate.ai/benchmark?slug=crag · How It Works · Data refreshed daily, snapshot 2026-09-05.