How to build a Map-Reduce Summarization Chain / Workflow?

[djpecot]

Howdy! Looking to start using this tool in addition to Langchain for quick prototyping :)

What would be the best practice to implement a map-reduce style summarization of an entire document (like in Langchain)? Is there a best practice or way to perform any kind of custom summarization task on a document that exceeds the token limit for a single prompt? I browsed the docs, checked the UI, and checked issues and discussions here without finding anything.

[HenryHengZJ]

In ConveraationalRetrievalQAChain, click the Additional Parameters, from there you can select stuff, map reduce type

[djpecot]

Yep, I got this one working, however seems like it's just pulling random rows from my vectorstore and not giving an appropriate summary of the entire pdf I uploaded (just one):

I understand that the k value also can change how many documents are returned, however in this case I want to perform a map-reduce summarization chain across my entire document (I mean pdf in this case, not how a document is chunked).

[djpecot]

Here's another example:

[djpecot]

Random Thought:

I guess this process of iterating over each part of a document would be similar to the creation hypothetical document embeddings as well in that the LLM must interact with all available data, but not within the same completion.

[HenryHengZJ]

oh got it, you are trying to use summarization chain with map-reduce, instead of a conversational retrieval qa chain?

for now we still haven't integrated into flowise yet, but might be a good addition - https://js.langchain.com/docs/modules/chains/popular/summarize

[djpecot]

@HenryHengZJ bingo. I haven't dug into Flowwise too much yet, but is there a code runtime I can use to write custom langchain code? If I could do that, wouldn't necessarily need as a feature since I can write it myself :)

[kianwoon]

I'm also interested in a document summarisation chain for LangChain. It would be a valuable feature to add. While LangFlow doesn't currently offer this capability, they do support custom components through their beta feature, which is worth considering.

[luizcarvalho]

@djpecot Do you solved this question?