betterstuff is a stuff chain that loads as many documents as possible without exceeding context size
betterrefine is a refine chain with updated prompts, but for now it's functionally equivalent to the normal refine chain. I tried aditing the prompt to add a score to each answer and use that in refining, but I couldn't find a prompt that would make vic7b reliably write relevant scores.
Add two custom chains:
Other:
MODEL_MAX_TOKENSin env