Request for a new approach to combineDocument Chain for efficient GPT processing

[zhn1010]

Hi. I hope you are doing well.

I am writing to request a new feature in the LangChain library that combines the benefits of both the Stuffing and Map Reduce methods for working with GPT language models. The proposed approach involves sending parallel requests to GPT, but instead of treating each request as a single document (as in the MapReduceDocumentsChain), the requests will concatenate multiple documents until they reach the maximum token size of the GPT prompt. GPT will then attempt to find an answer for each concatenated document.

It reduces the number of requests sent to GPT, similar to the StuffDocumentsChain method, but also allows for parallel requests like the MapReduceDocumentsChain method. It can be beneficial when documents are chunked and the length of them is fairly short.

We believe that this new approach would be a valuable addition to the LangChain library, allowing users to work more efficiently with GPT language models and increasing the library's usefulness for various natural language processing tasks.

Thank you very much for considering our request.