EnterpriseRAG-Bench

Why we included this project

Public RAG benchmarks are usually built from clean, well-formatted documents, which is not what enterprise search actually looks like. EnterpriseRAG-Bench instead gives you a corpus of about half a million documents spread across nine realistic sources, from Slack channels and Gmail threads to Linear tickets, meeting transcripts, and support tickets, plus 500 questions to test against. The questions are written to trip up naive systems: some require pulling together conflicting documents, some test whether the model knows when the answer is not in the corpus at all. The whole thing simulates a fictional company called Redwood Inference, so the data carries the internal messiness of a real organization. If your corpus looks different, the code can generate a similar benchmark for your own industry or company size, which makes it a useful starting point for teams that want a test set closer to their actual environment.