Add docs for setting MapReduceChain token_max arg

[ShantanuNair]

• Description: Add docs on how to set and use the token_max arg for MapReduceChains
• Issue: Several issues have been repeatedly raised regarding this, since it is non-intuitive how to pass this arg. This makes usage with larger models such as gpt-3.5-turbo-16k much easier. I also explain how it's a sliding scale between how much the inputs will be summarized vs tokens left for generation.
• Tag maintainer: @baskaryan @hwchase17
• Twitter handle: More PRs incoming! :) https://twitter.com/ShantanuNair

Related Github discussion: https://github.com/hwchase17/langchain/discussions/2746
Related issues: #1349, #7043, #434, #6397, #6714, #6191

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)			Jul 11, 2023 11:30am

[ShantanuNair]

Updated to reflect token_max arg in chain initialization from #7204

[hwchase17]

i think it should be referencing #7204

[ShantanuNair]

@hwchase17 @baskaryan

[baskaryan]

don't think ReduceDocumentsChain is mentioned anywhere before this? should we update the custom map reduce section first, so that this is better contextualized?

[ShantanuNair]

@baskaryan I'll get to that and update the PR.

[leo-gan]

@ShantanuNair Hi , could you, please, resolve the merging issues? After that ping me and I push this PR for the review. Thanks!

[efriis]

Hey @ShantanuNair any chance you got around to resolving the merge conflicts?

[efriis]

Actually - I think most of these chains docs have been rewritten with LCEL. Closing this, and let me know if you think it's still important somewhere else!