Okay, so in this video, I'm going to look at using Gemini 1.5 Pro for doing a number of tasks around code. And the first thing that I thought I'd take a look at is something that people were asking in the comments: 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 a much longer form tutorial for, over time. So, basically, crewAI is a framework for multi-agent apps or bots. And it uses LangChain underneath in here. You can see here there's some docs about it in the repo, which I'll talk about, which we'll bring in. I'm telling you what it is, telling you how to install it, telling you 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 freeform 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 crewAI 30 files. And the other one that I've brought in are docs 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. I'm not sure if they actually have a 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 35,000 - 37,000 tokens in total. So I've just put in first off, tell me about crewAI, what it can do. So it says, okay, crewAI, 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 crewAI can do.

Create and manage AI agents, assign and execute tasks, foster collaboration and delegation, integrate with various LLMs, enhance with tools, and then 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 LangChain here. So I was curious to ask it, Okay, what's it made with? and sure enough, Gemini had no problems being able to look at this and say, Okay, it's got Python, it's got Pydantic, it's got LangChain, and then by default, it's got OpenAI. And it includes some other additional libraries for different kinds of things as well. All right, so what I wanted to do was basically set this up in Colab, 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: 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 crewAI 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 packages do I need to install into Colab to use this? And so it basically tells me that I'm going to need crewAI, I'm going to need LangChain, I'm going to need LangChain 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. And so you can see the first thing I've done is just copy

those over to Colab and run them. Alright, 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 me some additional tips. Okay. So next up, I basically ask it, Can you write me the code using crewAI 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. Alright. So we can kick this off and start it running. And you'll see that the inference times for this are 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 is basically

obviously using OpenAI here by default. but it's bringing in, from crewAI import agent, task, crew process. it's then basically set up our hotel chain owner, with the information, the goal for them, the backstory for them, setting them up to be verbose. 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. Alright, 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 OpenAI 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 that'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 that's doing this. And we don't have any sort of recursive criticism loop here going on, which probably would make it better outputs. 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, 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 worked as a clearly the code has worked for the agent framework in here. Alright, so the next thing that I wanted to do is not use OpenAI. So let's get it to use a Gemini model in here. Okay, so you can see that I've basically asked it, Can you make some

changes to use Gemini Pro for the LLM via chat 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 that'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 LangChain 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 install, I think, is this.

langchain_google_genai, because we want to use the AI studio, one in here. And then we need to actually put, import that in, like this ChatGoogleGenerativeAI, in here as opposed to what it originally came out with. I 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 OpenAI. 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. Alright? We can see that there's a bunch of things that we can do. Okay, so now I'm going to try and see, can it generate something

that uses a tool as well. So I've asked it, you know, can you write me the code using crewAI for a simple bot that has a search agent, that uses the tool, duck, duck go 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. 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 doctor goes, 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 it. 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 the Gemini. It looks like that basically here it needed a backstory.

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. We run it. Now, the agent's actually working at goes and does a search. For this, we can see that it's getting back information about Sora from OpenAI. About a text-to-speech API from Google's Gemini AI model for generative AI and Android's new personalization features. And then basically, you know, passes along to the Doomsayer, 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.

et cetera, et cetera, et cetera. 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 putting 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've 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 Copilot, 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. Although to be honest, getting it to write 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. There are 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 find the video useful, please click like and subscribe. And I will talk to you in the next video. Bye for now.

Gemini 1.5 Pro for Code - Part 01

HI my name is Sam Witteveen, I have worked with Deep Learning for 9 years and with Transformers and LLM for 5+ years.  I was appointed a Google Developer Expert for Machine Learning in 2017 and I currently work on LLMs and and since earlier in 2023 on Autonomous Agents. 

Sam Witteveen

Gemini 1.5 Pro를 사용하여 코드 관련 작업을 수행하는 방법을 보여주고, 특히 crewAI 저장소를 기반으로 예제를 빌드하는 방법을 설명합니다.

crewAI는 다중 에이전트 앱 프레임워크이며, Gemini Pro를 사용하여 에이전트 간의 협업, 도구 사용, 정보 검색 및 재작성과 같은 다양한 작업을 수행하는 코드를 생성하고 실행하는 과정을 시연합니다.

crewAI 저장소의 코드와 문서를 Gemini 1.5 Pro에 제공하여 저장소에 대한 이해를 높이고, 필요한 패키지 설치 및 코드 수정 과정을 통해 Gemini Pro의 코드 생성 및 문제 해결 능력을 평가합니다.

Gemini 1.5 Pro를 사용한 코드 작업 소개

Gemini 1.5 Pro를 사용하여 리포지토리를 분석하고 예제를 생성하는 방법을 살펴본다.

crew.ai 리포지토리를 업로드하고 문서 및 소스 코드에 집중한다.

crewAI에 대한 정보를 요청하고 기능에 대한 설명을 얻는다.

crewAI 설정 및 필요한 패키지 설치

Colab에서 crewAI를 설정하고 최소한의 개입으로 작업을 수행하려고 한다.

crewAI 사용에 필요한 패키지 목록을 Gemini로부터 얻고 pip를 사용하여 설치한다.

두 에이전트(고객 및 판매원)가 상호 작용하는 간단한 봇을 위한 crewAI 코드를 생성한다.

호텔 체인 소유자와 에어컨 판매원의 역할, 목표 및 배경을 설정한다.

생성된 코드를 실행하고 에이전트의 작동 방식을 확인한다. 에어컨 종류에 대한 답변을 얻는다.

Gemini 모델을 사용하도록 코드를 수정하고 Gemini를 사용하여 출력을 생성한다.

Gemini Pro를 사용하여 LLM을 변경하는 과정

Gemini Pro를 사용하여 LLM을 변경하고 필요한 패키지를 설치하는 과정을 설명한다. 오류가 있음을 언급한다.

Gemini LLM을 설정하고 코드를 실행하여 Gemini를 사용하여 출력을 얻는다.

도구를 사용하는 코드 생성 시도

검색 에이전트와 AI 종말론자 에이전트를 사용하여 AI 릴리스에 대한 정보를 얻고 재작성하는 코드를 생성하려고 시도한다.

코드에서 오류가 발생하여 수정하고, 배경 설명이 누락되어 수정한다.

수정된 코드를 실행하고 AI 릴리스에 대한 정보를 검색하고 종말론자가 재작성하는 과정을 확인한다.

문서를 포함하지 않고 소스 코드에 대한 질문을 통해 코드베이스를 이해하는 것이 유용함을 설명한다.

Gemini 1.5 Pro가 코드와 관련하여 다양한 작업을 수행할 수 있음을 강조하며 마무리한다.

향후 CSV 파일 관련 영상 제작 계획을 언급하고, 댓글, 좋아요, 구독을 요청하며 마무리한다.

🤖 Gemini 1.5 Pro를 활용한 코드 작업

이번 영상에서는 Gemini 1.5 Pro를 사용하여 코드 관련 작업을 수행하는 방법을 살펴봅니다. 특히, **레포지토리를 분석하고 이를 기반으로 예제를 만드는 것이 가능한지**에 초점을 맞춥니다.

선택된 레포지토리는 crew.ai이며, 이는 다중 에이전트 앱 또는 봇을 위한 프레임워크입니다. **crew.ai는 LangChain을 기반으로 구축**되었으며, 레포지토리에는 관련 문서가 포함되어 있습니다.

⚙️ crewAI 레포지토리 분석 및 설정

crewAI 레포지토리에 대해 물어본 결과, Gemini는 crewAI가 무엇을 할 수 있는지 설명했습니다. **AI 에이전트 생성 및 관리, 작업 할당 및 실행, 협업 및 위임 촉진, 다양한 LLM과의 통합, 도구로 기능 향상** 등의 기능이 있다고 합니다.

crewAI를 사용하기 위해 필요한 패키지를 Colab에 설치하는 방법을 물었더니, crewAI, LangChain, LangChain community, unstructured 등이 필요하다고 답했습니다. **Gemini가 제공한 pip install 명령어를 사용하여 필요한 패키지를 설치**했습니다.

👩‍💼 에이전트 기반 봇 생성 실험

두 에이전트가 서로 대화하는 간단한 봇을 만들기 위해 crewAI 코드를 작성하도록 요청했습니다. **한 에이전트는 고객, 다른 에이전트는 에어컨을 판매하는 뛰어난 영업사원** 역할을 합니다.

영업사원이 여성임을 명시했음에도 불구하고, **Gemini는 성 중립적으로 작성**하는 경향을 보였습니다. 의도적인 것인지, 아니면 단순히 무작위적인 것인지는 알 수 없습니다.

OpenAI를 사용하지 않고 Gemini 모델을 사용하도록 변경했습니다. **ChatGoogleGenerativeAI를 가져오고 LLM을 Gemini LLM으로 설정**했습니다.

도구를 사용하는 간단한 봇을 만들도록 요청했습니다. **검색 에이전트가 DuckDuckGo를 사용하여 AI 릴리스에 대한 정보를 얻고, 다른 에이전트는 해당 정보를 AI 종말론자의 관점에서 다시 작성**합니다.

코드를 실행한 결과, 에이전트가 실제로 작동하는 것을 확인했습니다. **Sora에 대한 정보, Google의 Gemini AI 모델의 텍스트 음성 변환 API 등에 대한 정보를 가져와서 종말론자가 재작성**합니다.

🔑 코드 분석 및 문서 작성 능력

문서 없이 소스 코드만 제공하고 질문을 하면, **Gemini가 코드 기반에서 주요 클래스, 입력 및 출력을 파악하는 데 능숙**하다는 것을 발견했습니다.

Gemini 1.5 Pro는 코드와 관련하여 다양한 작업을 수행할 수 있습니다. **테스트 또는 문서 작성도 가능하지만, 문서 작성의 경우 코드에 대한 추가적인 정보가 필요**할 수 있습니다.

다음 영상에서는 CSV 파일을 분석하고 관련 작업을 수행하는 방법을 살펴보겠습니다. 의견이나 질문이 있다면 댓글로 남겨주세요. 영상이 유용했다면 좋아요와 구독 부탁드립니다.