Autogen Agents with RAG - Upgrading Agents with a Knowledge Base

Summary

视频介绍了如何使用自动生成框架Autogen进行知识检索，以提高工作流程的效率。通过向GPT模型提供特定上下文，可以生成更个性化和针对性的内容。以向QI团队发送电子邮件为例，展示了在提供公司背景信息后，GPT能够生成更具体和相关的邮件内容。视频还提到了使用Chrome扩展和AI代码助手来提高编程效率，以及如何通过Autogen框架进行自动化服务的推广。

Takeaways

• 📚 使用Autogen框架进行知识检索是提升工作流程的关键更新。
• 🔍 知识驱动（RAG）通过在提示中添加更多数据或上下文，帮助利用新知识或特定知识完成任务。
• 📧 在写给特定公司的电子邮件中，提供公司相关信息可以避免内容过于通用。
• 🤖 GPT-4由于知识截止日期的限制，可能无法准确了解新出现的公司或平台。
• 📈 通过Autogen的检索代理，可以从QI的GitHub仓库中提取数据，以改善生成的电子邮件内容。
• 🔗 利用Autogen的Gab仓库中的笔记本和示例，可以启发不同的用例。
• 🛠️ 使用Chrome扩展和AI代码助手可以提高代码理解和问题解答的效率。
• 📝 在自动化服务的电子邮件中，个性化和展示对项目的理解是关键。
• 🔄 通过对比使用和不使用RAG的电子邮件，可以看出知识检索对内容质量的影响。
• 📊 知识检索（RAG）可以显著提高AI框架在特定任务中的性能。
• 📌 通过优化提示和系统消息，可以进一步提升自动化框架在创建个性化内容方面的潜力。

Q & A

什么是RAG（Retrieval-Augmented Generation）？

RAG是一种结合了检索和生成的技术，它通过在生成过程中引入外部知识源（如数据库或网页）来增强模型的知识库和生成质量。

在脚本中，如何使用RAG来改进GPT-4的输出？

通过在提示中添加关于特定公司或主题的额外内容，RAG可以帮助GPT-4生成更具体、更准确的信息，而不是仅仅基于其内部知识库生成通用内容。

脚本中提到的Autogen是什么？

Autogen是一个框架，用于构建和部署基于GPT-3的应用程序。它提供了一套工具和库，使得开发者能够更容易地利用GPT-3的能力。

脚本中提到的QI是什么？

QI是一个平台，旨在通过利用AI来提高销售和客户服务团队的生产力。它通过自动化数据录入和分析等任务，帮助团队更专注于与客户互动和完成交易。

脚本中提到的GPT-4在没有额外上下文的情况下对QI的了解如何？

GPT-4在没有额外上下文的情况下，对QI的了解非常有限，可能会基于其内部知识库生成一些基于猜测的内容。

脚本中提到的Chrome扩展是什么？

脚本中提到的Chrome扩展是一个名为“继续”（Continue）的AI代码助手，它可以帮助用户在编写代码时提供智能建议和帮助。

脚本中提到的自动化服务是什么？

自动化服务指的是使用技术（如AI和RPA工具）来简化和自动化业务流程，以提高效率和生产力。

脚本中提到的个性化电子邮件的目的是什么？

个性化电子邮件的目的是向收件人展示发件人对他们的了解和关注，从而提高回应率和建立更好的客户关系。

脚本中提到的CMO在团队中的角色是什么？

CMO（首席营销官）在团队中负责市场营销策略的制定和执行，专注于市场定位、品牌建设和推广活动。

脚本中提到的自动化服务如何与QI团队的工作集成？

自动化服务可以与QI团队的工作集成，通过提供先进的自动化解决方案，帮助他们优化开发和生产流程，提高AI代理的工作效率。

Outlines

00:00

🔍 知识检索与自动生成的结合

介绍了在特定工作流中使用自动生成（Autogen）框架的重要性。通过向提示中添加更多数据或上下文，可以利用新知识或特定知识来完成特定任务。例如，向GPT模型提供关于新公司的额外信息，以便生成针对性的电子邮件。视频展示了如何使用知识检索在Autogen中，以及如何通过提供更多上下文来改善GPT模型的输出。

05:02

📝 代码实现与应用

讨论了如何通过代码实现知识检索，以及如何将检索到的信息应用于自动化工作流。提到了一个名为“继续”的Chrome扩展，它是一个AI代码助手，可以帮助理解和测试代码。视频还展示了如何使用Autogen框架来构建一个销售和营销用例，以及如何通过提供GitHub仓库的URL来让检索代理获取信息。

10:04

📧 自动化服务的电子邮件撰写

展示了如何使用代理框架来撰写一封针对QI团队的自动化服务推销电子邮件。邮件的撰写过程包括了多个角色，如老板、助理、高级撰稿人和首席营销官，他们各自有不同的任务和目标。视频还比较了使用和不使用知识检索（RAG）时的电子邮件撰写效果，以及如何通过提供正确的提示和指导来改善输出。

15:05

📩 电子邮件内容的比较与分析

分析了在没有使用知识检索（RAG）的情况下，GPT模型生成的电子邮件内容。指出了邮件内容的不足之处，如缺乏个性化和对QI项目的了解。然后，展示了在提供QI项目上下文后，GPT模型如何生成更准确和个性化的电子邮件。视频还提到了如何通过比较QI与其他代理框架来增强邮件的说服力。

20:06

🚀 利用RAG提升自动化服务

强调了在自动化服务中利用知识检索（RAG）的重要性。通过实际案例，展示了如何通过RAG来提升电子邮件的个性化程度和效果。视频还提到了如何通过更好的提示和系统消息来优化代理框架，以及如何通过这种方式来实现大规模的个性化。

25:08

📢 订阅频道与持续自动化

视频最后，鼓励观众订阅频道以获取更多关于自动化和代理框架的视频。同时，提出了对视频内容的反馈和建议，以及如何通过观众的反馈来优化和改进视频内容。

Keywords

💡知识检索

知识检索是指在特定框架中，如Autogen，通过提供额外的数据或上下文信息来增强模型的理解和响应能力。在视频中，这允许模型更准确地了解特定任务的背景，如向特定公司发送电子邮件。

💡Autogen

Autogen是一个框架，它允许用户通过检索代理来获取和利用外部数据，以增强生成内容的相关性和准确性。在视频中，Autogen用于从QI的GitHub仓库中提取数据，以便更好地理解QI公司。

💡上下文

上下文在这里指的是与特定任务相关的额外信息，这些信息可以帮助模型更好地理解任务的背景和需求。在视频中，上下文的提供是通过在提示中添加关于QI公司的信息来实现的。

💡GPT

GPT（Generative Pre-trained Transformer）是一种自然语言处理模型，它能够生成连贯的文本。在视频中，GPT用于生成电子邮件内容，但其准确性受到其知识截止日期的限制。

💡自动化服务

自动化服务指的是使用技术来简化和优化业务流程，减少人工操作。在视频中，自动化服务被用来提升QI团队的工作效率和扩展其影响力。

💡个性化

个性化是指根据特定用户或群体的需求和偏好来定制产品或服务。在视频中，个性化是通过在电子邮件中包含关于QI公司的具体信息来实现的，以展示对公司的了解。

💡角色扮演

角色扮演在这里指的是在自动化流程中，不同的AI代理扮演特定的角色，以完成特定的任务。在视频中，有老板助理、文案撰写者和高级文案撰写者等角色。

💡开源开发

开源开发是一种软件开发模式，它允许社区成员共同参与和改进软件项目。在视频中，QI公司对开源开发的承诺被视为其项目成功的关键因素之一。

💡AI代理

AI代理是指在自动化流程中执行特定任务的人工智能实体。在视频中，AI代理被用来模拟团队成员，以提高工作效率和处理复杂任务。

💡工作流程自动化

工作流程自动化是指使用技术工具来自动执行一系列任务，以提高效率和减少错误。在视频中，工作流程自动化是通过AI框架和代理来实现的，目的是简化和加速业务流程。

Highlights

知识检索与自动生成（Autogen）是任何工作流程中的关键且可行的更新或升级。

RAG（Retrieval-Augmented Generation）或知识驱动，基本上是将更多数据或上下文引入提示中，以便利用新知识或特定知识完成特定任务。

例如，如果你想给一个新公司写电子邮件，而这个公司在C GPT中找不到，你可以在提示中添加关于公司的内容。

如果不提供关于Qi的更多上下文，生成的电子邮件将会非常通用。

视频展示了如何在Autogen中使用知识检索。

GPT-4由于知识截止日期在Qi推出之前，所以不知道Qi是什么，可能会基于上下文写出一些不准确的内容。

通过Autogen的检索代理从Qi的GitHub仓库中拉取数据，可以生成更好的电子邮件，因为它能够理解上下文。

Autogen gab仓库提供了许多不同用例的笔记本，这些笔记本包含了许多示例。

展示了如何使用Autogen中的检索代理来拉取Qi GitHub仓库的数据。

代码示例展示了如何构建一个团队，包括老板、检索代理、文案和高级文案，以及CMO（首席营销官）。

目标是向Qi团队发送一封个性化的冷邮件，展示对他们所做工作的了解。

在没有RAG的情况下，团队可能会生成一个通用的电子邮件。

使用RAG时，电子邮件会更加个性化，因为它包含了从Qi GitHub仓库中检索到的上下文。

展示了如何使用Autogen框架来创建个性化的电子邮件，而不是使用RPA（Robotic Process Automation）工具。

强调测试代理框架的潜力，而不使用RPA工具，以查看仅使用代理框架的潜力。

展示了如何使用Autogen框架来创建个性化的电子邮件，而不是使用RPA工具。

讨论了如何通过更好的提示和系统消息来优化代理框架。

视频最后邀请观众订阅频道，以便获取有关自动化和代理的更多视频。

Transcripts

knowledge retrieval with autogen this is

a crucial and viable update or upgrade

for any of your workflows when you're

using a gentic Frameworks specifically

autogen for those of you who don't know

what is rag or knowledge drival is

basically pulling more data or more

context into the prompt allowing you to

leverage mostly a new knowledge or

specific specific knowledge for the

specific tasks so for example let's say

I would like to

write an

email to a specific company and this

company um is a new company that you

can't find in C GPT so what you can do

you can add content about the company in

the

prompt and let's say you write provide

to me a c email for to the company cre a

I so you provide it in the prompt and

then CH GPT will write the specific C

email to this company in this instance

Qi but if you didn't provide more

context about

Qi the email would be pretty generic so

in this video I'm going to show you how

I'm using knowledge knowledge retrieval

in autogen wow I feel Rusty I didn't

upload a video a few days ago when my

speaking kind of sucks but bear with me

because this is a very interesting use

case so let's just

cover a few things so before we dive in

I want to show you that CH GPT doesn't

know at the moment what is Q AI can you

write for me an email

about crew

AI do you know what is clue AI so it

will probably just write

BS so Qi let's see Qi is a platform

designed to enhance sales and customer

service teams productivity by leveraging

AI it assists in automating various

tasks such as data entry and Analysis

enabling teams to focus more on engaging

with customers and closing deals Qi

typically

integrates with existing crm's

communication tools to provide

actionable insights Coach sales rep

Etc so you can see um the llm in this

specific case GPT

4 since its knowledge cut off was before

the launching of Qi it doesn't know what

Q is it does H write some BS based on

its understanding from the context

because he sees okay a name which is Qi

so it thinks okay it's Pro it's possibly

a platform leveraging artificial

intelligence which is correct in this

case but it's not good enough so let's

say I want to write an email to the quii

team I have to feed CH GPT or anting an

agentic framework that I'm leveraging I

have to feed it with the

knowledge of what is Qi so basically

what I did and I'll show with I'll show

you the code in a moment I used the

retrieval agent in autogen to pull data

from the QI GitHub repository and then

see

if the email that I generated was better

because it was able to understand the

context by the way the whole idea is a

spin-off from the

autogen gab repository and they have

here over here they have many notebooks

and these notebooks contain many

examples of different use

cases obviously I'm not just copying and

pasting this and I'm using these

examples as in inspiration to what I'm

trying to build and this use case is

focused is more like a sales SL

marketing use case now let's dive into

the

code uh one more quick update by the way

I told you a few weeks ago I I showed

you to those of you who are

following the channel I showed you this

Chrome extension which is called

continue not a Chrome extension it's a

visual

code extension which is basically an AI

code assistant I plugged into it dolphin

mistal which is a model that is

supported by orama and it has been

working pretty well so far so I'm using

it to ask it many questions about the

repository about the code just testing

it around to see exactly what go what's

what it's understanding what doesn't

understand you can check out the video

specifically that I did about this uh

visual code extension a few days ago but

it's pretty cool so far so good I'm

pretty happy about

it now let's uh take a look into the

code I won't cover the installation you

know my Moto this channel isn't about

installations it's about applications

and

implementations but in general we

have a crew that consists of a boss the

boss assistant which is a retrieval

agent an assistant who has extra cont

content retrieval power for solving

difficult

problems and we are providing him with

the URL in this case the

URL the knowledge base is the the GitHub

repository of Qi let me show you the the

content you can see over here you guys

are probably familiar with it if you've

been following the agent space so you

probably know about Qi so they have

information over here Qi is a Cutting

Edge framework for all Orest rating role

playing autonomous a autonomous AI by

fostering coll collaborative

intelligence Qi empowers agents to work

together seamlessly tackling complex

tasks and they have a ton of information

about y AI why is it better than Chad

Dev which I don't know nobody's talking

about I mean based on my observation

nobody's talking that much about Chad

Dev lately but I need to if you guys

seen anything else so let me know but I

I'll definitely check out the repository

afterwards and see if I'm correct or not

and this is a comparison to autogen

never mind so basically this code is

going to pull the

content of this page so let's go back to

visual studio code so we have the boss

we have the boss assistant which is a

retrieval agent and as you can see it is

pulling let's see if I can make this

larger

let's close this

guy yes

and so we have the retrieval agent we

have the

copywriter and I gave it the system

message you are a witty and bold

copywriter specializing in writing short

and perspective called

emails we have the senior

copywriter and the system messages you

are a senior copywriter you specialize

in providing constructive feedback to to

copy of copy maybe and we have the CMO

you are a CMO specializing in marketing

copyrighting and understanding the

prospect mindset and the whole goal over

in this instance

is to send a Cod email to the QI team

show them that we know what they are

doing so personalizing the Reach

Out

you can see over here so the

problem basically the prompt for the

group is write a short called email to

the QI team pitching Automation Services

so I'm providing Automation Services and

let's assume that I want to pitch the QI

team start with a hook then make sure to

show them your knowledge of the project

and then add a CTA suggesting suggest

thing a quick call also write two

follow-ups in case they don't answer the

third follow-up should be a standalone

email that will be sent to the founder

of the project offering him to come to

be a podcast a guest on on the

podcast now the reason the reason why I

wanted to build this workflow which we

using

rag because in the past if you guys have

been following me and by the way this

may be a

good um opportunity to tell you that I

invite you to subscribe to the channel

if you enjoy videos about Automation and

agent and you haven't subscribed yet

please make sure to

subscribe I usually drop this H plug or

this request to subscribe at the end of

the video but my wife told me I should

do it at the beginning this is not the

beginning it's like halfway through or

maybe even more but next time I'll

probably remember to do it um at the

beginning anyway

so the whole idea here

in opposing to other videos in which I

automated a

workflow of personalizing a DM in

Facebook or LinkedIn based on what I

scraped

from the prospects profile which is also

a viable um

Modi like you can do that H what I

wanted to do to see here if I can use

only the agentic framework in opposing

to what I did in the past

which was basically using an RPA tool

Microsoft power automate to scrape the

data and then fit it as

context to

gp4 and in the prompt it had the context

about the prospect so this is slightly

different workflow and the main reasons

for doing this is first of all

I'm I want to create videos and

tutorials that are not using necessary

RPA tools and like doing the whole

workflow from a to zed using Python and

the second reason is I want to stress

test um the agentic Frameworks because I

want to see I'm pretty pretty

straightforward I want to see the

potential of using agentic Frameworks

without using RPA

tools which yeah they do they do allow

us more control but at end of day the

whole goal of using an AIC framework and

my vision and hopefully other people

Vision in the space is using agents to

streamline the whole process and

automate more parts of the process

instead of using an RPA tool which

basically we create and control the

workflow and just use the ating

framework or the GPT in only certain

instances well I've been talking a lot

let me take a sip of water

Okay cool

so I showed you the The Prompt for the

team we've

created and the boss the boss assistant

the copywriter the senior copywriter and

the

CMO now we have two um different

instances

one is an instance with no rag we will

just send the prompt and these are the

agents and there is another chat another

instance in which we also share um the

context so as you can see over here we

have the boss Aid which is the retrieval

agent the CMO the copywriter and the

senior

copywriter

and there is a distinction between both

chats because one we have print line

message rag chat and the other one

contains the no rag tag so I already run

this code let me go up to show you the

results

okay let's do this again

it will be better let's see if I I can

also let me just adjust the prompt to

make it more

interesting

um

special let's see what we can

adjust also WR two follow-ups in this

case write a short call email to quii

team pitching Automation Services start

with a to start with a hook then make

sure to show them your knowledge of the

project try to

comp lement them on

their

unique

this compare

them to a

competitor okay comparing them to a

competitor and this is something that

only if if you take a look in this

page you will know the difference

between Qi and the other agentic

Frameworks so when we are not going to

use the rag there is no

way that uh the output is going to be

correct now let's run this code and see

what happen in in both instances again

one instance we have no

rag the chat team has no idea what is Qi

they're probably going to BS and the

other team does know what is Qi because

we have the retrieval agent and he's

pulling

data okay so as you can see over

here write a short Cod email to the quii

team pitching Automation Services start

with a hook then make sure to show them

your knowledge of the project try to

complement them on their

uniqueness comp compare them to the

competitor and then add a c ta

suggesting suggesting wow a quick

call

okay this is the first one email One

initial called email to the qu AI team

transforming the future of work with CI

hey CI team hope this finds you well

I've been closely following your journey

and I'm genuinely impressed by how

you're redefining productivity tools in

the workspace okay you can see it's it's

incorrect but still it is somewhat

compelling and can be relevant no

probably not relevant but it is pretty

generic but not so far off so it's clear

you're not just another player in the

game you're setting new standards much

like how Tesla reimagined the electric

car okay nice analogy your unique

approach to integrating AI into everyday

work tool is nothing short of

revolutionary it's evident that Qi is on

the path on becoming an indispensable

asset for business aiming to stay ahead

of the curve okay that's good that said

I believe there's an untapped potential

in further automating your processes

enhancing efficiencies and scaling your

impact our team specializes in B spoke

automation solutions that could

seamlessly integrate with your existing

Frameworks propelling kui to new heights

how about a quick call next week to

explore how we can collaborate sorry

about that how we can collaborate to

make work even smarter

I'm confident we can add significant

value to your already impressive Suite

of

tools looking forward now follow up if

no response in three or four days let's

make magic happen at Qi hey Qi team just

floating to the top of your inbox in

case my previous email got buried Buried

buried I'm still super excited about the

possibility of bringing our automation

expertise to complement kui Innovative

platform

can we find 15 minutes this week for a

chat I have a few ideas that I believe

could really amplify your

impact second followup quick ping hey qu

team I understand you you're swamped

with making a work the workspace a

better place so I'll keep this brief

steing on blah blah blah let's move

on this is the Standalone email to the

founder dear Founder's name I've been an

admirer of Qi and your Visionary

leadership in transforming how we

approach prod activity tools your

journey is a beacon of for Innovation

much like how pioneers reshape

Industries this is correct I host a

podcast that spotlights thought leaders

and disruptors who are making

significant waves in their sectors

okay now um so this was the no rag chat

obviously not not obviously but as you

can see over here and first of all the

pumpt for the cold email wasn't good

enough so I didn't Pro

provide

um enough examples perhaps and also good

enough guidance to the

copywriter and so it was too long but in

general I do like this is way too long

but in general I like the I somewhat

like the content but so this is one

thing that obviously needs to be

improved like the specific promp prompt

and as you can see over here the group

chat basically doesn't exist so only the

copywriter

answered now let's move on to the

rug so as you can see over here in

opping to the

first um process that has no rug over

here

automatically autogen is

creating the knowledge base so it is

fetch fetching the data from the URL

that we provided which is the GitHub

repository of

crei and now it it created a few

documents and now you can see the boss

assistant to the chat manager so you

retrieve augmented coding assistant your

answers blah blah blah this is a a bit

of

um

information now the user question is

write a short IM to the qu I team

pitching Automation Services start with

the hook then make sure to show your

knowledge try to complement them you

know this

already now he is adding and this is the

rag the context is and it is basically

scraped the page from Qi let's open

it

scraped the context roles of agents in a

crew this just looks like a lot of text

to you guys but as you can see over here

let's see what

happened okay elevate your AI crew with

tailor Automation Services this is the

message tailor to the QI team that

already is supposed to know the context

let's see what is the

difference I hope this message finds you

well I've been closely following your

journey in orchestrating role playing

autonomous AI agents and I must say your

framework is nothing short of

revolutionary the way Qi empowers agents

to collaborate and Tackle complex task

is a game changer setting you apart from

competitors like autogen and ched Dev

yes baby this is exactly what we wanted

so you can see took this

context

and he it also took this how I compar it

name dropped the

correct competitors now let's see

because we are using GPT 4 let's see

just making sure to see if it knows what

our autogen and ched Dev so do you know

do you know the

names of the

QI of qu

eyes

competitors Ki operates in the r of AI

power la la la so it

is offering stuff that I don't even know

but it's obviously incorrect great now

let's move back to the email I hope this

message find you well we covered that

already while your approach offers

unparalleled flexibility and ad an

adaptability I believe our Automation

Services can further enhance your

project efficiency and scalability our

expertise in integrating Advanced

automations solutions could streamline

your development and production

workflows ensuring your AI crew Crews

operate as their Peak

potential I'm particularly impressed

with your commitment to open source

development and the community-driven

enhancement that have prop propelled Qi

forward it's clear that your project is

not just about technology

not just about technology it's about

building a future where AI can work

alongside humans

seamlessly would you be open to a brief

chat okay call to action let's move on

to the next email quick followup hi I

just wanted to quickly follow up on my

previous email I understand you're

incredibly busy but I truly believe

thato let's see if there is anything

specific to

Qi nothing specific okay but because I

didn't um guide the second followup the

first followup to be specific to

kui now another email last try can we

help KU I saw I

realize also over here nothing

specific now let's see the message to

the

founder I've been deeply inspired by

your work with Quan the groundbreaking

strides you're making in the realm of

autonomous AI agents your vision for a

future where Ai and humans collaborate

seamlessly is not only Innovative but

essential I host a podcast blah blah

blah and that is pretty much

it

so let's go back to the wide screen

because it just looks

better I let's go back here as

background Qi although it's it's kind of

funny because I've used

autogen in the whole video and the video

was about approaching the QI a team or

founder but it was just a a simple a

simple example of how we can leverage

rag in h an ating framework in this

specific case it was

autogen I very happy about the results I

believe that with slightly better

prompts and better system messages for

each one of the agents this could be

very valuable because it basically um

will allow us to create

personalization at

scale obviously there are many other use

cases and instances in which you can

leverage uh retrieval Rags or knowledge

bases but this is what just a a simple

example of how we can leverage it for

creating

personalization um I guess that's it for

today guys if you enjoy the video like

always please leave a comment below I'm

happy to hear any feedback any

suggestions for more videos any

suggestions how these videos can become

more interesting more impactful please

let me know in the comment section

always happy to learn and optimize and

improve um regardless again if you

haven't subscribed yet please do and I

guess that's it until next time keep on

automating