The RIGHT WAY To Build AI Agents with CrewAI (BONUS: 100% Local)

Summary

在这个视频中，演讲者展示了如何使用Lightning AI来构建一个高效的Crew AI团队。Lightning AI不仅提供了一个基于云的代码编辑器，允许团队成员在云端协作编写代码，还可以利用开源模型来增强功能。视频中，演讲者首先创建了一个新的Lightning Studio，并在其中建立了一个名为“financial analyst crew”的代码框架。通过YAML文件定义了两个任务——“research company task”和“analyze company task”，以及两个相应的代理——“company researcher”和“company analyst”。接着，演讲者展示了如何使用LangChain和Gro模型来执行这些任务，并最终通过Lightning AI的GPU加速来提高处理速度。此外，还介绍了如何使用Lightning AI的Studio模板来快速启动一个项目，并利用API Builder插件将模型作为API服务暴露出来，从而实现与Crew AI的无缝集成。视频最后，演讲者鼓励观众如果喜欢这个视频就点赞和订阅。

Takeaways

• 🚀 介绍了如何使用Lightning AI搭建一个基于云的协作代码编辑器和开源模型的Crew AI团队。
• 📂 创建一个新的Studio和代码工作室，以便在云端进行项目管理和环境管理。
• 🔧 采用模块化的方式构建Crew AI代码框架，使用yaml定义代理和任务。
• 🛠️ 定义任务和代理，通过在配置文件中编写代码来明确它们的角色和目标。
• 🔄 通过迭代和复制任务定义，快速创建多个任务，每个任务对应一个代理。
• 📈 展示了如何为金融分析团队创建任务，例如研究特定公司和分析公司财务状况。
• 🤖 为每个任务创建代理，如金融研究员和金融分析师，并为它们设置详细的背景故事和目标。
• 📋 编写主文件main.py，将代理和任务组合在一起，创建Crew AI的运行流程。
• 🔗 介绍了如何使用Lang chain和Grock来连接和驱动Crew AI中的代理。
• 🛠️ 演示了如何使用Poetry来打包和管理项目依赖，为项目创建一个虚拟环境。
• 🔧 通过Lightning AI的API Builder插件，将Crew AI与OpenAI兼容的API端点连接。
• 🎉 强调了使用Lightning AI的GPU资源来加速大型模型的推理过程，并展示了实际操作过程。

Q & A

视频介绍了使用什么工具来搭建Crew AI团队？

视频介绍了使用Lightning AI，这是一个云端代码编辑器，可以与他人协作编辑代码，并能够为开源模型提供动力。

Crew AI的代码库在未来将会是什么样子？

Crew AI的代码库在未来将会是非常模块化的，拥有独立的工具区域，使用YAML定义代理和任务，并且所有内容都会汇集到一个简短的main.py文件中。

为什么要使用Lightning AI来管理Python环境？

使用Lightning AI可以避免Python环境管理带来的头痛问题，因为它本质上是一个每次启动都是全新环境的云端代码编辑器。

在Crew AI中，如何定义任务和代理？

在Crew AI中，任务和代理通过在config文件夹中创建相应的.yaml文件来定义，其中包括角色、目标、背景故事等信息。

视频提到了哪些金融分析中的重要指标？

视频提到了盈利能力比率、流动性比率、偿债能力比率等金融分析中的重要指标。

如何使用Lightning AI的GPU来为开源模型提供动力？

通过在Lightning AI的云端环境中配置模型，设置相应的GPU参数，并将模型通过API与Crew AI团队集成，从而使用Lightning AI的GPU为开源模型提供动力。

视频最后提到了哪些步骤来运行和测试Crew AI团队？

视频最后提到了使用Poetry来整合项目、安装依赖、运行项目以及测试使用不同的模型（如Gro和Mixol）来驱动Crew AI团队的步骤。

如何将Crew AI团队的API端点暴露出来？

通过使用Lightning AI的API Builder插件，设置端口和启用认证，然后获取基础URL，最后在Crew AI代码中使用这个URL来替换原来的模型配置。

视频提到了哪些工具来帮助分析公司的财务表现？

视频提到了使用SEC工具来帮助分析公司的财务表现，这是一个已经存在于Crew AI示例库中的工具。

在视频中，为什么要在不同的任务中使用不同的模型？

在不同的任务中使用不同的模型可以根据任务的具体需求选择最合适的模型，例如，某些任务可能需要更专业的财务分析模型，而其他任务可能只需要基本的信息检索功能。

视频提到了哪些步骤来确保Crew AI团队的结构正确？

视频提到了创建源文件夹、定义任务和代理、创建主文件以及使用Poetry进行项目整合等步骤来确保Crew AI团队的结构正确。

Outlines

00:00

🚀 构建最优的Crew AI团队

本段介绍了如何使用Lightning AI这一云端代码编辑器来构建一个Crew AI团队。Lightning AI不仅支持代码协作，还能为开源模型提供动力。视频作者展示了如何创建一个新的工作室，并构建一个模块化的Crew AI代码框架。通过定义任务和代理，使用yaml文件来组织结构，并利用Lightning AI的自动环境管理功能，简化了Python环境配置的复杂性。

05:01

📋 定义任务和代理

在这一段中，视频作者详细阐述了如何定义任务和代理。通过创建不同的文件夹和文件，如config文件夹和agents.yaml、tasks.yaml文件，来明确任务和代理的结构。作者还解释了如何为每个任务和代理编写具体的描述和期望输出，以及如何为每个代理设置角色和背景故事。

10:02

🔧 编写主要的代码文件

视频作者在这一部分开始编写主要的代码文件，包括导入必要的库和创建crew AI项目的基础结构。通过定义agent和task的类，并将它们与之前定义的配置文件关联起来，作者展示了如何将任务和代理整合到一起。此外，还介绍了如何设置任务的处理顺序和详细输出。

15:02

🌟 运行Crew AI团队并测试

在最后一段中，作者运行了构建的Crew AI团队，并进行了测试。首先，使用Gro模型快速有效地完成了任务。然后，作者展示了如何使用Lightning AI的Studio模板来运行更大规模的开源模型，并将其API端点集成到Crew AI中。通过这些步骤，作者成功地展示了Crew AI的强大功能和灵活性。

Keywords

💡crew AI

crew AI是指一种人工智能团队，它能够根据任务需求自动分配和协调不同的AI代理。在视频中，crew AI的创始人展示了如何构建一个crew AI代码库，并使用Lightning AI这个云基础的代码编辑器和开源模型驱动平台来实现。

💡Lightning AI

Lightning AI是一个云基础的代码编辑器，它允许多人在线协作编辑代码，并且可以驱动开源模型。在视频中，Lightning AI被用来构建和运行crew AI团队，提供了一个方便的环境来管理Python环境和依赖。

💡代码框架

代码框架是指在软件开发中提前定义好的代码结构和布局，它为开发提供了一个清晰的指导和模板。在视频中，金融分析团队的代码框架包括了源代码文件夹、配置文件夹以及主程序文件等。

💡任务

任务是指在crew AI中定义的一系列操作或目标，代理需要根据任务描述来执行相应的动作并输出预期结果。在视频中，任务包括研究公司和分析公司两个部分。

💡代理

代理是指在crew AI中执行任务的AI实体，每个代理都有特定的角色和目标。在视频中，代理被定义为金融研究员和金融分析师，它们根据任务需求执行不同的职责。

💡配置文件

配置文件是用来存储软件或系统中设置和参数的文件，它通常包含了系统运行所需的各种配置信息。在视频中，配置文件用于定义crew AI中的代理和任务。

💡主程序文件

主程序文件是软件或项目中的核心文件，它通常包含了程序运行的主要逻辑和流程控制。在视频中，主程序文件是用来将代理和任务整合起来，形成完整的crew AI团队。

💡云环境

云环境是指基于云计算技术提供的虚拟化环境，它允许用户通过网络访问和使用存储、计算等资源。在视频中，Lightning AI提供的云环境使得用户无需管理本地Python环境，可以直接在云端编写和运行代码。

💡环境管理

环境管理是指在软件开发中对开发环境进行配置、维护和更新的过程，包括管理依赖库、软件版本等。在视频中，由于使用了Lightning AI的云环境，避免了本地Python环境管理的繁琐。

💡开源模型

开源模型是指公开发布且允许他人自由使用和修改的预训练AI模型。这类模型通常由社区共同维护，可以被集成到不同的平台和应用中。在视频中，作者提到了使用Lightning AI的GPU资源来驱动开源模型。

💡API

API（应用程序编程接口）是软件之间进行交互的一种约定，它定义了软件组件之间如何通信。在视频中，作者展示了如何通过API Builder将Olama模型的API端点暴露出来，并将其与crew AI团队集成。

Highlights

介绍如何使用Lightning AI云编辑器建立一个Crew AI团队

Lightning AI不仅支持代码云端协作，还能为开源模型提供动力

创建一个新的Studio，简化Python环境管理

使用YAML定义代理和任务，模块化构建Crew AI代码框架

定义任务：研究特定公司的股票信息

定义任务：分析公司的财务状况

创建代理：公司研究员和公司分析师

编写main.py文件，将代理和任务整合在一起

使用Grock开源模型，通过Lightning AI GPU加速

展示如何使用Lightning AI的API构建器公开模型

使用Mistol和Lightning AI GPU运行Crew AI团队

演示了如何结构化Crew AI，使用Lightning AI和开源模型

通过Lightning Studio快速启动和运行项目

展示了如何将Crew AI与不同的模型和工具结合使用

使用Poetry管理项目依赖和运行Crew AI团队

通过Lightning AI的云环境避免了本地环境配置的麻烦

提供了一个完整的示例，从设置到运行Crew AI团队

Transcripts

in this video I'm going to show you the

optimal way to set up your crew AI team

this comes straight from the crew AI

founder and he has shown me the future

of what a crew AI codebase looks like

and we're going to be building all of

this using lightning AI which is not

only a cloud-based code editor which

allows you to collaborate with anybody

on your code in the cloud but you can

also power your open- Source models with

it which I'll also be showing you how to

do so we're going to build a Crea I team

and then we're going to swap out gp4 and

we're going to power it using mixol or

mistol we'll see and I want to thank

lightning. AI for sponsoring this video

I'm super excited to build this team on

lightning and I'm going to publish this

lightning studio in the description

below so you can get access to it you

can clone it and you can play around

with it yourself the first thing we're

going to be doing is building our crew

AI code framework and so the first thing

we need to do is create a new studio so

if you don't already have a lightning

account go ahead head and sign up

they'll give you some free credits to

get started click new studio right there

and we're going to create a code Studio

then we click Start and this entire

environment is in the cloud so it really

saves you a lot of headaches with python

environment management because it's

essentially a fresh environment every

single time and if you've been watching

my videos you know how frustrating

python environment management is all

right so the high level of how we're

going to structure our crew is now very

modular we're going to have a separate

area area for tools we're going to be

using yaml to Define our agents and

tasks and everything is going to pipe

into a very short main.py file so I just

spun up a new lightning studio and

that's really all we have to do we can

already get started so the first thing

we're going to do is create a source

folder so go ahead rightclick click new

folder and then type source and

everything should feel very familiar and

lightning because it is essentially vs

code in the class and the cool thing is

you don't actually need to set up your

environment you don't need to struggle

with your python environments with

dealing with all of these dependencies

it just works so we created our source

folder and then within the source folder

we're going to name our new crew so

we're going to actually create another

folder within that and I want to make a

financial analyst crew and so just like

that I name it financial analyst crew

hit enter then that creates another

folder all right then within that folder

we're going to create a another folder

config and all of this is just to

structure what we need to put everything

together so from here we're going to

write click we're going to create a new

folder we're going to call it config and

this is where we're going to put our

definitions of our tasks and our agents

and as I mentioned they're both going to

be Emo files so let's do that now right

click we're going to say new file and

we're going to call it

agents. for the first one and then we're

going to right click new file and we're

going to call it task. for the second

one and now this is where we're going to

Define our agents and tasks and the nice

thing about doing it this way so if you

start getting into the habit of

structuring your cruise just like this I

believe pretty soon crew AI is going to

give you the ability to actually expose

an endpoint to control your cruise based

on this structure so it'll automatically

be able to create you an API so that's

why it's good to follow the golden path

here so let's start writing our tasks

the first task is going to be

researching a specific company that I

want analyze so I'm going to call it

research company task and then we put a

colon after it and then we need two

parameters within that description and

expected output So within the

description this is where we're going to

describe the task at hand so what we're

going to say is use a Search tool and

we'll come back to this we'll actually

build the Search tool to look up this

Company's stock information and then I

put company name and the fact that I

have curly braces around the company

name now makes it a variable which we're

going to pass in later and this makes it

so that we can basically pass in any

company name the goal is to prepare

enough information to make an informed

analysis of the company's stock

performance the next thing we need is

the expected output and for the expected

output I say all of the relevant

Financial wh I spelled Financial wrong

financial information about the

company's stock performance okay so

that's the first task it's as simple as

that now I'm going to make a second task

so I'm just going to copy that first

task paste it again

and the second task is going to be

analyze company task so I'll delete the

descriptions from before and I'll delete

the expected output from before as well

and we'll put new ones in place so I

switched over to Claude 3 which by the

way is my new go-to model I no longer

prefer GPT 4 Cloud 3 has just been so

good and I asked what metrics should

someone include in a financial analysis

and it gave me a bunch of different

metrics and so I'm simply going to copy

paste these into the description of the

task I want so I say take company names

and again I pass in the variable

financial information analyze it and

provide a financial analysis including

profitability ratio liquidity ratio

solvency ratios and so on and for the

expected output I am going to say a

nicely formatted analysis including all

of the financial metrics necessary for a

thorough financial analysis of a company

and so let's stop there obviously I can

get much more sophisticated than this

but that's not the point I just want to

get something up and working and then I

want to show you powering it using an

open- Source model and powering that by

lightning AI gpus okay so we have two

tasks so research and then analyze and

then we need agents now so let's click

over to the agents. yl file so here is

the structure of an agent definition and

the name is going to be company

researcher the role I'm going to name it

Financial researcher the goal very

similar to the task now one quick thing

that I just added is I explicitly said

using search tools I don't know if I

actually have to do that but I like

explicitly saying it CU it doesn't hurt

and it makes me feel like okay it's

actually going to know to use those

tools okay backstory an expert Financial

researcher who spends all day and night

thinking about financial performance of

different companies now we're going to

set allow delegation to false because we

don't want this agent delegating this

task to anybody else and we're going to

set verose to true because I want to

actually see everything that this agent

is thinking so I'm going to copy that

and I'm going to paste it down here to

create our second agent now typically

what I do is I match every agent to an

individual task so if there are two

tasks there'll be two agents so let's

start changing this now so instead of

company researcher we're going to call

this the company analyst and instead of

financial researcher we'll call it

financial analyst and then we're going

to replace the goal so for the goal I'm

going to say take provided company

financial information and create a

thorough financial report about a given

company the backstory an expert

financial analyst who prides themselves

on creating clear and easily readable

Financial reports of different companies

again allow delegation false there's no

need for that and verbose true so now we

are done with agents now uh one quick

thing I want to mention is you don't

actually have to save any of the files

while you're working with lightning AI

it just saves it automatically as you're

going which is really nice I've

obviously gotten into the habit of

hitting command save all the time so it

doesn't hurt but you don't really have

to do that and you can shut this off you

can shut down the environment and come

back to it later and it's all just going

to be there working ready to go just

like you left it okay next we're going

to create our main file that puts all

the agents in task together so

rightclick on financial analyst crew and

we're going to create a new file and

let's call that crew. Pi so the first

thing we need to do is import all the

relevant libraries so we're going to say

from crew AI import agent crew process

and task and those are the four main

pillars of a crew AI project then we're

going to import crew base agent crew and

task from the crew ai. project and also

because grock is awesome we're going to

be using grock to power all of this and

we could do so through Lang chain so

make sure you have all of these

installed and we can pip install them if

we need to later but right now we just

need Lang chain grock import check Rock

now let's actually create the crew base

so enter enter and we'll say at crew

base and then we're going to create a

class financial analyst crew open PRS

close pars colon we're going to describe

what it is and then we're going to load

up the agents and tasks that we just

created then we need to set the grock

information so we're going to do def

init self none and then self. grock llm

and this is where we pull in the grock

chat information we're going to say

temperature zero model name and we're

going to be using mixol and we'll need

to set the token but we'll do that in a

minute now let's pull our agents and our

task into this file so at agent okay so

I wrote out the first agent definition

so we have at agent def company

researcher self and then we point it at

agent and we're going to be returning a

new agent object config and then we pull

in the config from the agents. yl file

and we pass in the grock L M as it's llm

and remember as I've said a lot of times

you can actually pass in a different

model from different sources into each

individual agent so if you want to use

gp4 for one clad for another Mixel for

another and then an open source model

powered by lightning as another you

could totally do that next let's copy

this and we're going to create our

second agent and this second agent is

company analyst okay so it's going to do

the same thing we're going to pass it as

an agent return the agent op object and

instead of company researcher we're

going to call it company analyst and

that's it th those are our definitions

for the two agents now we need to create

definitions for our tasks so just like

agent we say at task at the top we'll

say def company research task same as

before just like agents except instead

of agent we say task here new line then

we'll return a task object with

definitions config is equal to self.

task config and then we'll pull in the

relevant task and rather than passing in

an llm which wouldn't make sense for a

task we're going to pass in our agent

that we want to do this task so we say

self. compan researcher and that should

be good now let's copy it and create the

second task down here here we'll say def

analyze company task we'll create the

task return the task but we'll say

analyze company task here and we

actually want the company analyst as the

agent and I think we're done I think

this looks good ah no we need one more

thing so we actually have to Define our

crew which puts it all together all the

agents and all the tasks so we say at

crew def crew and we pass it itself make

a new crew object so here we're going to

return a crew and now we need to pass it

all the information to actually create

that crew so we'll do agents is equal to

self. agents and the self. Agents is

automatically created by these

decorators right here just a heads up on

that so that's all we got to do then for

task we do the same thing self. task

then we need to set a process so we're

going to do process. sequential because

we just want it happening one after

another and then verbose is equal to two

because we want it to Output everything

now I think we're done all right now we

need to actually create our main.py so

we already have one right here so let's

go ahead and move that into the

financial analyst crew okay main.py and

all it does is print hello lightning

world so so we're going to import OS and

we're going to import the rest of what

we need to basically use a m file so

Fromm import load. MV load. M then we're

going to import our financial analyst

crew so it's as simple as from

financialist analy financialanalyst

crew. crew import and then import the

financial analyst crew so now we have

everything we need all right and this is

the entire main file so we're going to

Define run we're going to pass in the

inputs and the only input we have is

company name Tesla and the nice thing is

crew automatically interpolates all of

that for us then we're going to say

financial analyst crew. crew kickoff

we'll pass in the inputs and then we put

this kind of main portion down at the

bottom now let's create our n file and

we're almost done so on the very outside

of this entire folder we're going to

rightclick new file and we're going to

call ITV and we're going to pass a grock

API key right there so I grabbed a new

grock API key and yes I am going to

revoke this key before publishing the

video so that's it grock API key and

then pass it in just like that okay so I

am going to create some tools now and

we're not actually going to create them

from scratch we're just going to use

tools that are currently in the crew aai

examples Library already for us so we

have an SEC tools which is kind of

exactly what we need so let's go ahead

and we're just going to copy the whole

file we'll switch back and then in the

financial analyst crew we're going to

right click and we're going to create a

new folder and we're going to call it

tools then we're going to create a new

file in there and we're going to call it

SEC tools. py and we're going to paste

in all that code so all of this

hopefully will just work out of the box

all right now we're going to use poetry

to put it all together and to actually

run the project so to do that we're

going to first do pip install poetry

then we're going to create a new file

and we're going to call it Pi project.

TL and you don't really have to know how

much of this works and to be honest I

don't really know how much of this works

uh I'm just not super familiar with

poetry but this is the way crew is going

to be done going forward so we're going

to paste in all this stuff and it

basically just sets up poetry for us and

the only thing that we're going to need

to customize is the financial analyst

crew name right there next we're going

to lock the dependencies so poetry lock

all right should be done and there it is

the poetry. loock file perfect then

we're going to do poetry install Okay so

this is installing all the dependencies

now all right there it's done now we

should be able to run it from here so

poetry run financial analyst crew hit

enter uh I missed one thing so I forgot

to actually rename it to what we need so

financial analyst crew right there now

that it's there we should be able to run

it properly so I'm going to do poetry

install one more time just in case all

right now we got no warning so now it

should work poetry run financial analyst

crew all right so I got another module

not found issu so let's do pip install

linkchain grock and I'm not sure why we

have to do do that again but let's try

it hopefully this works okay so let's

run it again all right it looks like we

got it working wonderful wow look how

lightning fast that was okay so let's

see what it did so here we go entering

the chain Final Answer Tesla's ticker

symbol current stock price market cap

getting all this great information about

Tesla and recent news as well then it

passes it on to the financial analyst

and it's doing a bunch of different

calculations okay all perfect and then

we have the ratios and the different

metrics that we asked for and then the

final output perfect look at this

beautiful so it all worked and it was

lightning fast so now we have our crew

set up but it's using Gro which is

fantastic but we want to try running it

with an open- Source model powered by

lightning AI so you can plug in any

model that you want so leave this tab

open all right then we're going to go to

the studio templates page within

lightning. so it's lightning. a/ Studios

and this basically comes with a bunch of

preconfigured Studios that are ready to

go and they have a bunch of them a bunch

of really cool ones from running olama

fine-tuning you can build diffusion

models they just have a bunch of them

all right and it already has this run

mistal mixture of experts and all we

need to do is Click open template right

there so we select select what team

space then click open and this takes a

little while to load up because it's

doing a lot in the background but it

sets up everything for us and then I'm

going to show you how to then expose the

API from oama and then we're going to

take that API endpoint and then power

our model with it all right so it's done

now if we look up in the top right and

we hover over this button we can

actually see that we're using a GPU an

a1g and if we click on it then click

into the gpus we can see we have a bunch

available all the way through A1 100s

h100s and yeah they do cost a lot of

money to run but you don't need these

beefy gpus for all the models but we

will likely need something very powerful

for mixol and now they have L4 support

so you can just get another beastly GPU

running your models but I'll just show

you that we get it working I'm not

actually going to let it do all the

inference because that's not the point

let's continue so once we have that

we're going to come down here and we're

going to click this plus button for the

plugins we're going to go down to

serving and then we're going to do API

Builder and click install once that's

done we exit out of here on the right

side we now have this tab API Builder go

ahead and click that now this is one

that's just by default and you can

actually just go ahead and right click

on it and delete this like that and

we're going to do a new API okay so

we'll keep the name the same and for the

port it's 11434 and that is just o

llama's Port we don't need any

authentication although if we're going

to make this into a production

environment yes you definitely want

authentication then let's enable it then

we're going to click on the name and

we're going to grab this base URL right

there so now this is the endpoint that

is exposed with our open AI compatible

API so now I switch back to our code so

now to do it I'll leave the old code in

there but what we need to do is right

here from Lang chain. llms we're going

to import oama and then we're going to

on a new line create a new llm olama

mixol then we're going to call olama

we're going to pass in the model mixol

and then for this base URL we're going

to enter the URL that we copied earlier

so just like that so then we'll copy

this we're going to comment out this

line and we'll do self. I'll call it

grock llm just to save some time but

we'll call it olama mixol just like that

and it should all work now I believe

there might be one more thing we have to

do but let's give it a try and so poetry

run financial analyst all right and

there it goes so it's probably going to

be pretty slow because it is a very big

model Mixr and we're not using a super

beefy GPU but I just want to show you

getting it working and if we switch back

to the oama studio we can actually see

the GPU going bur right there so you can

see it's it's working right now and

there it goes look at that all right so

if I stop it has a bunch of good

information right there go back look at

the GPU and now it goes down to zero

again so we've confirmed it actually

works so now you know how to structure a

crew and do so in the most modern way

you know how to spin up an open- Source

model with lightning Ai and power it

with their gpus and you know how to

expose the endpoint and plug it into

your crew if you liked this video please

consider giving a like And subscribe and

I'll see you in the next one