Gemini 1.5 Pro for Code - Part 01

Key Takeaways

okay so in this video I'm going to look at using Gemini 1.5 Pro for doing a number of tasks around Cod and the first thing that I thought I'd take a look at is something that people were asking in the comments is could you get it to look at a repo and to basically ask it to build some examples based on that repo that kind of thing so the repo that I've chosen is crew AI some people mentioned this in the comments and also this is one of the projects that I've been meaning to make a tutorial for or much longer form tutorial for over time so basically crew AI is a framework for multi-agent apps or Bots and it uses Lang chain underneath in here you can see here there's some docks about it in the repo which I'll talk about which we'll bring in um telling you what it is telling how to install it telling it some various things that you can do with it that kind of thing as we go through so I've gone through and run this in Gemini I'm going to walk you through exactly what I did and just sort of show you I've gone back to running this as a free form prompt just because I was trying to get some of the formatting out and I found that this was nicer for that okay first of all I basically take the repo and uploaded it coming back and looking at the repo I'm not really interested in these parts I was interested in the docs not really interested in the tests although that could be useful later on and I'm interested in the actual source code for building crew AI in here I've brought that in as one of the folders that you see here this is the crew AI 30 files and the other one that I've brought in are the docks in here so this is basically just the docs of how to do various things I guess there's some images in there I didn't realize that at the time and then this is basically just sort of like the how-tos the markdown files and stuff like that so it's not really extensive docs it's just I'm not sure if they actually have the whole doc site but anyway what I've done is just bring in those two things so in total we've just got just under 50 files we're looking at around sort of 35 37,000 tokens in total so I've just put in first off tell me about crew AI what it can do so it says okay crew ai orchestrating ai Frameworks for collaborative intelligence and then it basically goes through and tells me a little bit about what it can do here's what Crea can do create and manage AI agents assign and execute tasks Foster collaboration and delegation integrate with various llms enhanced with tools and it's got some benefits it's got some different things like that so I was kind of curious at the start that I didn't see that it actually mentioned Lang chain here so I was curious to ask it okay what's it made with and sure enough J had no problems being able to look at this and say okay it's got python it's got pedantic it's got Lang chain and then by default it's got open Ai and it includes some other additional libraries for different kinds of things as well all right so what I want wanted to do was basically set this up in collab and I wanted it to do all the work and I'm going to try and do as little intervention as possible here so this is the whole idea is to see okay can we get it to basically write some stuff and get it working and stuff like that without me having to make changes I do know how crew AI works I've used it and tried it out for a few different things and stuff like that so I felt when I did this that okay yeah I've got a basic understanding of okay where it might go wrong that kind of thing so I asked it okay what package do I need to install into collab to use this and so it basically tells me that I'm going to need crew AI I'm going to need Lang chain I'm going to need Lang chain Community I'm going to need unstructured Etc and then it just gives me the nice pip installs so I thought okay let's just take those pip installs so you can see the first thing I've done is just copy those over to collab and run them all right next thing it gave me some information about that once it's installed you can do different things and stuff like that and it gave give me some additional tips okay so next up I basically ask it can you write me the code using crew AI for a simple bot that has two agents that talk to each other where one is the customer and one is the salesperson the salesperson is a great saleswoman who is selling air conditioners the customer has a chain of hotels and is thinking about upgrading them but wants to make sure he gets value for his money all right so we can kick this off and start it running and you'll see that the inference times for this actually reasonably short because we're only looking at something that's around 40,000 tokens in here okay so we can see that the code that it's given back it's basically obviously using open AI here by default but it's bringing in from crew AI import agent task crew process it's then basically set up our hotel chain owner with the information the goal for them the back story for them setting them up to be verose and our sales person here now it is funny that I specified that it was a saleswoman that the salesperson was a woman yet when it goes to write it it kind of makes it gender neutral I don't know if that's intentional or just random here so it's basically set up the task so the task is going to be that the customer expresses their need for air conditioner upgrades and their desire for value for money and we can see that the task for the salesperson is to introduce themselves inquire about the customer specific requirements Etc all right and then it's got an explanation of what it actually is doing here as we go through this okay so I've copied this over I've replaced it a little bit so I can actually get the open a I key from here and then I've just pasted in the rest of the code exactly like Gemini wrote out here and we can see if we start looking at the output of it that sure enough it starts working right that basically it decides off it's going to use a tool for this it's going to basically have a coworker action air conditioner salesperson research it's got a whole bunch of things in there it's doing this and we don't have any sort of recursive criticism Loop here going on which probably would make it better output but anyway this is a really sort of basic thing here and we can see the final answer is there are a few options to consider it's got some things about the different kinds of air conditioners and it's gone through and done these it's printed things out in yellow here which is a bit of a challenge anyway for the second task we can see that it's used the tool again that the person's introduced themselves we can see that the salesperson is basically asking for specifics about numbers of rooms numbers of floors that kind of thing and we can see that okay we've ended up with a dialogue or at least extracted out the needs and stuff like that for these particular people in here and in the end we can see that we've basically gotten the sort of thinking process of each of these agents it's it's worked as a clearly the code has worked for the agent framework in here all right so the next thing that I wanted to do is not use open AI so let's get it to use a Gemini model in here okay so you can see that I basically asked it can you make some changes to use Gemini Pro for the llm Via check Google generative AI change this code and tell me what packages I need to install via pip so it's saying okay here is the modified code it's generating out okay so it's finished generating out let's have a look at it okay so we can see that it's basically kept most of the code for this it's told us that the package that we need to install is Lang Ching Google which I don't think is exactly right but then it's also told us okay set the llm up to be chat Google generative AI to set the llm to be for Gemini Pro and stuff in here okay so it's made a number of mistakes in here or it's made a couple of small mistakes in here basically the package that we needed to inst I think is this Lang chain Google gen AI because we want to use the AI Studio One in here and then we need to actually P import that in like this chat Google generative AI in here as opposed to what it originally came out with and then basically just set up the llm like that I've changed the names to be Gemini llm in here the rest of the code is basically the same from what it's done and now we can run it and we can see sure enough it's using Gemini to basically give us outputs for this and the outputs are definitely different than open AI I should point out that the Gemini Pro that it's using here is the 1.0 version and not the 1.5 version in here all right we can see that there's a bunch of things that we can do okay so now I want to try and see can it generate something that uses a tool as well so I you know can you write me the code using PR for a simple bot that has as search agent that uses uses the tool duck Dogo to get information about new AI releases and then another agent that takes that info and rewrites it from a perspective of an AI Doomer and the AI is evil and will end the world I use the Gemini llm model all right so basically I'm trying to recreate the mainstream press for a lot of this sort of stuff they always seem to either hype things up or pull things down all right so let's see can it actually generate the code here so that we will actually get something that's using the tool as we go through this okay so I've taken the code in it seems to be using the old way of doing that so I'm just going to comment that out I'll leave it in for there and I'll update I'll just make this to be llm in there it hasn't given me the PIP for actually running this so let me just get that okay so it didn't give me the PIP to install duck. goo so I'm just installing that myself looking at the rest of the code apart from it sort of getting the sort of old way of doing the Google models the rest seems to be pretty good in this so let's run it and see I like it's basically AI doomsayer warn the world about the dangers of AI the task is search for recent news about AI releases rewrite okay so it's given them the tasks and stuff like that let's run it and see how it goes okay so we got an error out I copied it across to Gemini it looks like that basically here it needed a back story and this had left out the backstory here so let me try again so it gave me one for one of these let me try and get another one okay so it looks like it basically just left out the backstory for both of them I got Gemini to rewrite this one with a backstory and this one I just wrote myself and stuck it in there if we run it now the agent's actually working it goes and does a search for this we can see that it's getting back information about Sora from open AI about a text to speech API from Google a Gemini AI model for generative Ai and Android's new personalization features and then basically you know passes along to the doomsdayer you know who rewrites it these AI releases are just another step towards the inevitable AI takeover and we've got some nice text about how Sora will spread misinformation and propaganda etc etc etc so clearly it's been able to write something now that's using the tool of search and using you know this to actually go along so one of the things I've actually found is that it can be really interesting to basically not include docs just put in the source code and then start asking it questions about the source code and that allows you to sort of get a good understanding of what actually is going on in various code bases so I've tried this for some other code bases and I found that it was actually quite good at working out what were the key classes what were the key inputs and outputs that you would expect from different classes and different functions as you go through this so I think this is definitely a tool that you know is going to get better and better over time so obviously we've seen things like the GitHub co-pilot like duet AI for Google stuff like that these things you know are going to get a lot better and you can actually then go and ask it to write things like you know tests or other things like that and maybe even write some docs or to be honest getting it to WR docs I found not to be you know amazingly great that often you would need to tell it sort of a little bit about what the actual code is doing for it to sort of work out some understanding and then be able to write you know different docs for different parts of the code and stuff like that okay so I think I'll finish up there with this video as you can see that Gemini 1.5 Pro really can do a lot of cool things with code they'll probably lots of other things that we could have done of going through different code bases that kind of thing I will try and do another video in the future of looking at CSV files and looking at how we could do some of those things anyway as always if you've got any comments please put them in the comments below questions anything like that put them in the comments below if you found the video useful please click like And subscribe and I will talk to you in the next video bye for now