Building a Summarization System with LangChain and GPT-3 - Part 1

In this video, you will learn how to build a summarization chain using LangChain and GPT-3. We will explore the three types of document chains for summarization, the pros and cons of each and cover it all in code for you to easily use.

My Links:

Github:

00:00 Intro
03:44 Map Reduce CombineDocument Chain
07:22 Stuffing CombineDocument Chain
11:39 RefineCombineDocument Chain

#LangChain #BuildingAppswithLLMs

[mbrochh]

This was an excellent summary of how to do summaries with LangChain :) Sadly, their docs are a bit all over the place, so Youtube content like this is a real life saver at the moment.

[SouravGhosh-yboc]

This is an absolute Gem of an Playlist.

[seththunder]

You literally deserve million subs. I literally understood everything from A to Z.

[WillMcCartneyAI]

Dude you're a bit of a legend for all of this - thanks!

[derhoc]

Thanks for this great video! You deserve way more subscribers! My problem is that these summaries are often way too short. Is there a token limit for the output somewhere in the source code that I didn't see?

[shuvojyotirakshit]

Can you pls use gpt 3.5 for semantic search on local text files

[creativeuser]

For me, it sounds logical to say: full summary = sum of summary of parts. Not sure why we summarized the summaries.

Great content btw! Keep it up

[yazanrisheh]

Hey sam what about map rerank? isnt it also part of it? How can we use it?

[ArugolluSubrahmanyam]

Could you please clarify if we employ the same approach for utilizing Azure's OpenAI for summarization as we do for OpenAI?

[samarthsarin]

I tried using the steps used in this video
But one challenge i see is that my document is very big so when I split it around 15-20 docs are being created and then when I pass it to a open source llm model like vicuna or llama2 it goes out of memory or if document is small then it gives blank output.
Can you please create a video with open source systems and not limit the document to just 4 pieces. As in basically I would like to see behavior of open source llms with very large document size. It would be really helpful. Plesse let me know your thoughts

[jaisingh]

Can you create one video where you can tell how can we do with Hugging Face model inference api
and which models we can use

[danielsommer]

Great video! Thanks! Is there a difference between CharacterTextSplitter and And would using that reduce content loss?

[janakiraman]

Where is the text file. Cant find it in github or colab

[jpiabrantes]

Would love low-level step-by-step tutorial on building these things. Lots of info is loss by using only this high-level library calls.

[shuchishah]

i tried map_reduce in my application. when i upload document of 30 pages and i my chunk size is 1000, it takes around 19 seconds, which is quite a lot time. what can i do to reduce this time?

[LearningWorldChatGPT]

Great video !
Hi Sam!
One question: I took a pre-trained Hugging Face model to do a summary task (didn't want to use OpenAI API). But the summary was very bad (I did it for the Portuguese language) ... I would have fine-tuned it better, right? Would you have to perform this fine-tuning with a lot of data? (since it's an LLM)

[thenbaplayer]

Thank you for your video but could you make a video with more suitable for beginners from zero to that.

[roopyekollu]

Hi Sam, I see you took only top 3 documents to run the chain. Were you able to pass whole docs and succeeded? I'm running into token length limitation when I do that irrespective of the chain type

[VaunaKiller]

Hey, its not the first time I see this interface that allows to run python snippets steb-by-step, midex with some markdown inbetween.
Is it some web-based IDE, or like an extension for VSCode / other desktop IDE?
Btw, thanks for sharing!

[vviftbg]

Hi Sam, Thanks for the video. How can i use this for more than 4097 tokens? It threw up an error saying that I requested for 40, 914 tokens and it asked me to reduce my prompt.