introducing Company Knowledge Bench the first retrieval benchmark built on messy, real-world company knowledge soon agents will do a huge share of the work inside every company. all of it starts with finding the right information, but almost nobody can measure how well their agents actually find it public benchmarks don't help. they're either one narrow domain like law or medicine, or synthetic documents and questions. real company knowledge is messy, contradicts itself and goes out of date retrieval is the core of @kapa_ai, so we spent months building our own benchmark from 1,000 real-world queries by people and agents then we ran 7 retrievers on it: 1/ grep works surprisingly well, but it's 5x slower and floods the context with 40K tokens per query 2/ optimized agentic retrievers do far better. Kapa Deep scored highest (0.65) in ~5s with an eighth of the tokens 3/ we're still very early. lots of headroom left, especially as more agents run on company knowledge full write-up below