We benchmark grounded retrieval the way teams actually use it: real workplace questions, blind-judged answers, corpora of 220K documents and up.
This report publishes the methodology, the judge rubric, and the results across connector mixes — so you can reproduce the numbers on your own corpus.