Building a Summarization System with LangChain - Part 3 Using ChatGPT Turbo

Key Takeaways

okay this video will be just a quick one to follow up how could you do the summarization with the chat GPT turbo API so I'm going to show you we've gone through two videos now of doing it with the standard one and just recapping there you could do it either through the summarization chains or you could actually just set up a large language model and pass it in so we've got the exact same article about coinbase that we had in the last video so you can see the facts there and doing the fact extraction we get that with the standard API so with the chechi BT API there are a couple of different things we need to think about one that this is built for chat the big difference with this is you now have different sort of prompts and different inputs based on whether it's the part of the system whether it's a user whether it's the ai's response that raises some interesting issues of that we need to think about okay how are we going to package this up so if you look at just doing a standard chat with this we just pass in a human message and that probably has a system message on top of that but just doing this this is an example from their docs we're passing a human message and you can see it does it quite nicely so just doing a normal summary the way I've done it is that you use the system message to set up the context of what the whole thing is about so in this case the system message is you're an expert at making strong factual summarizations take the article submitted by the user and produce a factual useful summary here we're not using the user to give any instructions the user is just going to be putting in the actual article So the instructions are coming in the system message here you can think of these as like placeholder tokens that tell the model different kinds of information so a system message token tells it setting up the context setting up the task often that kind of thing and then the human message is what you would expect to get from the user so we take this we're passing our article as the human message so we've got these messages this list of messages and if we're actually doing it in chat we would have system human AI response Etc and this messages list can get reasonably long doing this but here because we really just want to translate one thing and then get it back we're not actually setting this up as conversational chat so we can just pass in the article there and sure enough it gives us a really nice summary back just from doing that it's treated it like a I guess in some ways you could think of it maybe treating it like a conversation but it's eliminated the bits about the conversation because we've set the system to basically just take the article submitted by the user and produce a factual summary does it work for the bullet lists ones yeah you can use it for this ones again here you would change the system message your human message is going to be the article just like before but the system message here is you're an expert at making strong factual summarization so the same as before but now extract the key facts out of this text don't include opinions give each fact a number and keep them short sentences and we can see that it gives back pretty nice information not that different not hugely different than what we were getting before the thing to think about here though is that each of these models tensed a little bit of style or personality to itself so you want to play around with prompt sometimes you're going to have to change the prompt for the kind of article or the kind of documentation Etc that you're summarizing so that works what about making the triples so normally what I was doing to make the triples was I was passing in these facts but one of the cool things with this is that we can actually just pass in so we can do it like that but we can also just pass in the article Itself by passing in the article itself we can get triples and this is like doing it the old way with this is doing one way so here we're basically just passing these in but we can also do it like this where we've got so actually sorry that was just extracting the the facts out there and what I was playing around with was that you could pass this in but actually with the chat gbt it seems to get better response if you actually just pass in the full article so rather than pass the list of facts which would you can play with this you're just passing the responses dot content in here but what I've passed in is the whole article and you can see that it's giving us number for each triple it's given us the subject to predicate object for each of these so it would be pretty easy to even apps chat should be T to write a piece and python code that would extract these out into whatever format you wanted to use in your knowledge graph once we've got them in this kind of format it makes it very easy to do and looking at it it seems to have done quite a nice job of doing it like this which seems different than the previous one was doing it's cut off a little bit at the end this because of the number of Max tokens that I've set so that's something that you could play around with so as always if you've got any questions please put them in the comments if this was useful to you please click and subscribe and I'll see you in the next video