I'm quite sure that this is one of the things you can let an LLM build for you. Run it against a small test corpus, check results, improve, repeat.

Doesn't have to take a lot of your time, mostly wall time, gpu ticks, cpu ticks, network.