gpt-oss:20b if you can run it, it's pretty good

You would use a RAG for augmenting the local model, I'll experiment with that soon