Beating GPT-5.6 Sol on retrieval with 100x cheaper open models

“Most teams' best training data is just sitting in their databases. The problem is that turning raw data into something usable is hard, and letting agents read, search, and mutate data cheaply at scal...