YouTube Transcript:
AI Agents from Scratch: A Beginner's LLM Workshop

Skip watching entire videos - get the full transcript, search for keywords, and copy with one click.

AutoDub

Understand YouTube Foreign Videos

Immersive YouTube Voice Translation

Break language barriers, embrace global quality content

Solve Foreign Video Barriers Instantly

Video Transcript

Video Summary

Summary

Core Theme

This content is a workshop session focused on building AI agents using Large Language Models (LLMs) and the LangChain framework, demonstrating practical applications and the underlying concepts.

Mind Map

Click to expand

Click to explore the full interactive mind map • Zoom, pan, and navigate

So we're going to start briefly. So

while we starting we like to launch

a survey

to understand your background for future references.

So please uh answer uh the survey while

So uh my name is uh Murat. I'm Stanford

scientist and Magnamine co-founder

and I worked on uh data science and AI.

Uh I've been working on data science at Stanford.

Stanford.

I also teach in Santa Clara University

um machine learning AI classes.

So uh we've been uh as a magnuman

academy we've been holding our meetup

groups for last five plus years is past

300 events. If you're not member of any

of our meetup groups please welcome to join

join

and then we have a YouTube channel where

we have our recordings from previous events.

events.

We put our most

meetup or events in YouTube channel except

except

um some of those register workshops. So

you have to be register our workshops to

attend it. But most our events there is

very diverse reach content and we share

all other our social media platforms we

provide different contents in each of them.

them.

So there are a few announcements we're

going to have. Um

first uh we're going to have a six weeks

uh LLM mini boot camp. This is a rare

event that we uh spec we specifically

designed for the industry's demands.

And as you filling the survey, uh we

understand from people that most people

look for job market or most job market

openings they require LLM background

and we are going to build AI agents and

you're going to learn the deep

understanding of LLM with Dr. Geran

which I will introduce him and seami

haj he's a engineer

in C3 AI and he's expert in LLMs and he

build LLMs for different big corporates

and also myself uh I will also teach the

LLM aspect how to build products in LLM

or what kind of products you guys can

build with LLM.

So uh that would be good in a intense

six weeks program in LLMs. If you don't

want to miss your spot and get all early

early bird uh it's finishing in 3 days.

While you're registering you have to

finish and complete the purchase

otherwise you won't have a spot and

after 3 days price goes up. That's a

unique and rare opportunity. We don't

know if we're going to offer that again.

So, but we have a regular our mentorship

data science internship programs 15

weeks where uh we assign mentors to students

students

uh four to five mentors working on so so

many projects. Plus we have a company

affiliations we provide where students

work with companies from industry

and then they get this experience but

most recruit recruiters ask on demand

and if you have any questions you can

let us know and then we're going to give

more information on that later. So I

will introduce Dr. Geran

and uh he is

co-founder Magnumine Academy

and he's been u his faculty in San Jose

State University. Prior to that he was

faculty in Kais University in Korea and

Santa Clara University

and he's also working on AI medicine

projects with Stanford University School

of Medicine

and um he's been uh in data science and

computational field in close to 20

years. He's a great person to learn data

science in general specifically LLMs and

AI agents.

So I will let Dr. Jaran start when he's

All right. Thank you. Thank you for the

introduction Dr. Mrat. So hello

everyone. Welcome to another Magnind

workshop session. This time uh we are

going to work together on

creating a simple uh AI agents step by

step. So, we're going to work on four

different uh small uh projects. Uh and

then uh I'm hoping that you guys are

going to work with me together. Uh we're

going to write the code together. You

don't need anything uh except uh opening

your Google Collab. That's it. So, you

don't need anything else. We are going

to build everything from scratch

together. I'll also give you the API for

um for the LLM that we're going to use

today. So all you have to do is just uh

uh type the code with me together and

then see what's going to happen. So

before we start uh uh I would like to

share this tweet with you. So let me see

if I can find it. Uh first let me share

Yeah, I'll share my screen first. Okay.

So this is uh from one of my favorite

machine learning uh book authors

Sebastian Rajka. So the day before

yesterday I just saw his uh tweet. I

really liked what he was saying. So uh

job roles in 2027 2026 I would say from

2025 to 2027.

So data scientist uh is going to turn

into analytics strategist. Before a data

scientist was expected to do cleaning

data sets, structuring queries. uh now

it's going to be about framing questions

and then interpreting outcomes. So

that's that's mostly what I'm interested

in. But if you guys are falling into any

other category like programmer, so

before writing code by line by line

after designing logic and structuring

systems, I really liked the way he put

uh u the change in the job market into

perspective. So that's the first thing I

would I wanted to share with you. Second

thing before we get started, I I wanted

to share with you is a small side

project I'm working on. So this is uh is

very small AI agent running locally

right now on my computer and then I try

to see what kind of tasks uh for jobs

are going to be replaced augmented by

AI. So I would like to just uh look at

data scientist and then data scientists

are related to so many different uh

jobs. So I wanted to select data

scientist. So this is running locally. I

haven't uh what is that uh I didn't

share this with publicly yet. So right

now the an LLM running locally is

analyzing all the tasks related to this

job and then again depending on how I

define different attributes of u

jobs that are going to be replaced or

augmented AI. So this is not directly

coming from LLM but I also gave a lot of

input expert input into LLM to to to

give me feedback on which tasks are

going to be replaced or augmented. So uh

I I I actually did this for many

different jobs. What I see is in the

next again this is not it is not

perfect. I'm not saying this is all 100%

correct but more or less I'm working

with another field expert on this. more

or less most of the tasks we're going to

see for different jobs are either going

to be augmented I'm not saying totally

replaced that's that's I think too

unrealistic but it's going to be

augmented uh by AI in other words we are

going to be competing with our peers who

know more AI so they're going to be the

ones that are going to be stronger in

the near future okay so that is another

thing I wanted to share with you before

we get started now let's let's get

started with our session uh please uh

open your uh browser and then if you

could just uh this is what we're going

to do today. Okay, this is what we're

going to do today. If you could please

open uh a new notebook in your Google

Collab. Okay, so you I'm going to use

Google Collab. Normally uh I did all

this work on my uh local VS code which

is looking much more organized and

professional and it's much easier for me

to push my code. uh but uh for this

session for the workshop I found uh

working on Google Collab will be easier

for all of us uh because we need to uh

install certain uh libraries and then

it's going to be easier if we all work

on Google Collab. So if you guys are

ready, we can start. If you need more

time, please type your questions,

comments, anything regarding this

session to Q&A. Let me know if you need

more time. But if you are ready, I'm

going to start. So just open a new

Google Collab Jupyter notebook. That's

Uh you don't need to title it now. Uh

yeah, it's a blank screen. Just a blank

uh notebook. That's all. That's all you

have to open. You don't see anything on

on on you don't see my screen. Can you

not see my screen?

Says blank screen. Can you can you

screen maybe he has to click on his own

uh computer? Yeah. Oh, okay. So, but you

Okay. Uh yes, everyone see. Yeah. All

right. Okay. Thank you. Can see. Okay. I

can see the screen. All right. Thank

you. All right. So uh the first thing

I'm going to do is I'm going to install

some uh uh libraries. So this will take

uh maybe uh what is it uh uh couple of

uh maybe 30 35 seconds uh depending on

on the speed. Okay. So here is what

we're going to do. We are going to use

pip install.

Uh okay. So lang chain. So today we are

going to use lang chain Google

genai. Okay. So that is first library we

are going to use and then lang chain

community. That's the second one we are

going to use. All right. And then we

going to use uh

a a vector-based

uh database face CPU from Facebook.

Okay, that's the next one. and sentence transformers.

transformers.

That's what we need for one of the uh

small applications we are going to

develop. So please type this. These are

1 2 3 four libraries that we are going

to use. So I'm going to run this cell.

Please run this cell too. So it's going

to install these cells for this session

to your Google Collab. So next time you

come when you run this, you're going to

you have to do the exact same thing. So

it's only going to be installed for this

session. Okay. So I'm running this. So

it'll take a while. So maybe 30 35 seconds.

So you're going to see a lot of uh uh

lines of

like requirements are set. So you're

going to see all these things. So don't

worry. So it's just uh installing the uh

Yeah. So, my sessions was were already

running. So, it's uninstalling and

installing again. So, in your case, this

See,

scroll up, please. Sure. Sure.

These are the I'm going to I'm going to

after I write every piece of code, I'm

going to copy and then share it with

you. So that Okay, I'm going to share it

So please just use Google collect. Okay.

If there are any uh

mismatchings uh if the uh distribution

found for sentence it's it's not found.

So I I have no idea about that error. So

Asia, so I'm sorry I cannot fix it now.

So I cannot I don't have any any any

suggestions how to fix it now. U but uh

on my end this is what I'm going to use.

Maybe maybe you can just follow the rest

of the session and then try to do it

later uh by by watching the recording.

Okay. So I think we are done. So it is

done. So I'm just going to comment this

out. Uh and then I'm ready to start

typing my code for Yeah, it was a type.

Okay. So I'm I'm going to start typing

my code for uh for for this session now.

Okay. Now the first thing is I'm going

to uh uh import Google generative AI.

Okay. From lang chain Google

Google

geni. We are going to use Google's

Gemini uh LLM today. Okay. So I'm going

to because I already installed lang

chain Google geni. So from Google lang

chain Google geni I'm going to import Google

Google generative

generative

AI. Okay. So that's that's that's what

I'm going to import now. So

I'm done with this line. So now I have

Google generative AI imported. All right.

right.

Okay. So, I'm just taking it slow now,

trying to see if everybody is settled

and then ready following me because now

I'm going to give you the API for this

session. This API will only be valid

during this next maybe 1 hour or or or

75 minutes then it's going to be deleted

and it has a cap. So, please uh just use

it during this session for this session.

Yeah, I will post the uh yeah, that's

that's what I'm going to do. So, I'm

going to post everything I write here so

that you don't have to type it if it is

taking too much time. All right? Or or

if you guys have any if you have any any

mistypos. Yeah, this is recorded. The

recording will be shared. Uh but if you

are already here, stay with me, okay?

Let's do it together. Okay. So, now I'm

going to share the API with you guys. So

for that what I'm going to do is I'm

going to write I'm going to say Google G

Google here Google Google

API key. So this is very important. This

one is very important. So with this API

key we are going to get connected to a

large language model. In this case it's

going to be Gemini uh 2.5 flash. So

Gemini has multiple products. We're

going to use uh we're not going to use

the pro version. It's it's it was a

little bit expensive and then we don't

need that now. I thought of running a

free local uh LLM but for this workshop

I think the first workshop I think it's

better if we use a paid uh u LLM so not

running anything locally because we need

to set it up for for for local

connection. I didn't want to spend time

on it. Okay. So what I'm going to do is

I'm going to say input.

So I'm just asking uh I'm just making

this uh more interactive. Okay,

otherwise I can just API

key. So this API key is going to be

stored as a string. Okay, so I'm going

to run this and then it is expecting

from me my Google API key. I'm gonna get

my Google API key uh here and

and

this one. I think I copied this.

Okay. Now I'm going to share this API

key. Let me first enter it here.

Okay. And then I will share it with you.

So you don't need to use any other API

key. You can just use this one for this

session. Again, it's not going to be

valid after this session.

All right, please let me know if you

guys got the API key.

Can we discuss deployment options in the

file? Oh, so Adam that that's a

different uh story. So maybe we can have

another session for how we can deploy

these models. For now, we are just

building uh a simple uh four different

applications. Right. Okay. I think we

are we are good to go. Now what I'm

going to do is I'm just going to see if

this is working. Okay. So I will say Google

Google generative

generative

sorry I'm not typing here. Google

generative AI. Okay. So that's what I

important. Okay. So Google generative AI

has multiple uh parameters. I'm going to

use two of these parameters. First I'm

going to tell which model I'm using. So

here is it. So there are different there

are different models right. So uh where

Gemini API has no here. So I'm using

Gemini 2.5 flash. I'm using this model.

Okay this model. So that's what I'm

going to use. So I'm going to say this

Okay. 2.5

2.5

flash. Okay, this is the product I'm

using. And this product has a API key. Google

Google

API key

is the key that we just

entered. Google API key. Okay. So this

way I am creating an instance from large

language model. Okay. I'm connecting to

large language model through the API.

And this is an instance of I'm going to

call this LLM. You I'm sorry not here.

I'm going to call it LM. You can call it

anything you like.

Going to be LM. I will copy the code and

share it with you in a minute. Let me

make sure that this is working. Okay.

All right. So

I will run this

and then I have a small typo here.

Okay. So I connected to genative AI. I

got llm. So next thing I'm going to do

is I'm just going to

send my first prompt. Okay. So I'm going

to use for that invoke method. So I'm

going to send my prompt and I will get

the response back from Gemini. Okay. So

I'm going to say say hello

hello

to this workshops participants.

participants.

I say this one.

Okay. And then I'm going to get the

response back

from Okay. So let's see.

So I am sending this prompt to LLM and

then waiting for the response. I got the

response. Let's print this response.

If you just say print response,

I will share the code in a moment. Let

me just run this. If you just say print

response, you will see uh hello

everyone. A warm welcome to all our

workshop workshop participants. We're so

glad to have you here today. Let's make

this a productive and insightful

session. So this is what Gemini is

telling us. Okay. So I'll I stop P here.

Let me see if Yeah. Uh so Rupa really I

it's very difficult for me to u check

what is wrong with installing so maybe

you can just follow the rest of the

works workshop and then try it later.

Okay I'm sorry about that. Okay so let

yeah is that a typo J yeah thank you so

I already typed the corrected the typo.

Okay. Now I'm going to paste the code.

So in case you were not able to write

the code.

this one.

chat. So another thing uh another

parameter for LLM is temperature. I

didn't use temperature here. I use

temperature the default temperature.

Temperature is another parameter that we

use to adjust how much of randomness we

are expecting from the response. So

basically what LLM is doing is it's just

predicting the next word. Right? If the

if the temperature is set to uh what

what is that uh one that means uh every

time you send the same prompt you will

receive the same response. uh by

changing the temperature you are

actually creating adding some randomness

to which word is going to be picked next

as LLM is making its decision on which

word it should pick next. So you are

adding some randomness to that decision

making process. So in that case every

time you send the same same prompt the

Yeah. So uh John uh that's what we we

want to do actually. So I I I see your

question why it did not give a length

response. So we are only we are we have

to be satisfied with whatever response

we are getting from this LLM. Now we are

going to write our code so that we are

going to give him some hints on what

kind of a response we are expecting from

him. So maybe we are going to tell him

that can you make this response warmer

and then longer. So that that's what

we're going to do. So this is just the

first step on seeing if our LLM is

actually active. So we now can we are we

being able to communicate with LM?

That's what we want to see now. Okay.

And then let's uh let me copy this one

too and then share with you guys. All

right. So this is the very first thing.

This is just the we're just checking if

our LM is working. Okay. If you guys are

ready, let's turn this into a function

and that will be our first demo. So it's

a row AI. Okay. I'm going to call it row

E AI. We are going to then work on this

to make it more interactive. So we are

going to actually steer AI to give us

responses in the format that we want him

to respond to us. So let me write our

let me turn this into a function and

it's going to look more professional.

Okay. So I'm going to say define my

first function. It's it's very simple

function. I'm going to call it row AI. Row

Row

AI. Okay. That's it. So there is no

parameter for my function. It's a very

simple function. So in this case I will

say I will ask expect from the user a

question. Okay. So what is that

Okay. So we are just going to ask

anything from Gemini and then again we

are just going to use this lm.

Uh let's use response

will be llm invoke. You can already see

the rest of the code right on the screen

but I'm typing it anyways. Okay. And

then I'm just going to print this. Print.

Print.

Okay. I will use this. So let's go to

the next line.

Gemini is saying this. Let me make

another. And this is the response.

So I'm just going to the next line.

Okay. So that is my function. Now I'm

going to uh

move outside. So and then run my function

raw AI.

AI.

Okay. And let's see what's going to

happen. Again the same thing we are just

asking some question. So I'm just

curious about AI agents. Okay. That's

the first thing I'm curious about. So

now we are getting the response and see

what Gemini is going to tell us about AI

agents. Okay. So this a very big

question actually. So uh yeah as you see

it's giving me uh a good explanation on

actually if you took any introduction to

AI course so you would understand you

would go through these terms pretty

easily right so perception action

environment so these are all defining

what an agent in AI agent should be

doing in an environment so then it's

giving us types of AI agents so model

based goal based agents and examples of

AI agents so this I It's not deep but

it's just giving us a very uh short

introduction to what AI agent is. Okay.

So our goal is not learning about AI

agent by the way. We just wanted to see

that we can make this work writing a

function. Okay. So that will be our row

AI function. As I said now we're going

to work on this and then make it work

better. Now let's do this. Let's uh

instead of asking a single question and

by the way let me check

is invoke the standard chat messaging

syntax with these lms. Yes. Yes. Yes.

Adam. So this is actually uh sending the

prompt and then getting the response. So

that's that's that's what we are using

for this is coming from lench chain.

And could you post here the pip install

code as well? I missed the beginning of

this. Sure. Yeah. in and we are going to

be sharing the recordings uh later. So

this is the pip install code I'm sharing

with you.

Okay. Now let's do this. Let's uh rather

than uh having a single question uh

we're expecting a single question from

outside let's put a number of questions

on a list. Okay. So and then we are

going to uh what is that uh uh again ask

from our Gemini to answer those

questions. So we want to do we want to

make a an observation here. Okay. So

this is this is the list of the

questions. So I'm going to put my first

question. Explain

machine learning.

Okay. So I just want him to explain to

me machine learning. So this is going to

be my first question. Tell me briefly. Briefly.

Briefly.

I think LM will understand even though

my pronunciation is not good. tell me

briefly uh about

Okay. So these are all same questions

that I'm

telling in different formats. Okay. What

is this machine learning thing? Okay. So

I'm asking questions in different

styles. Okay. So that's that's what I'm

doing. Next what I'm going to do is I'm

going to make my uh llm answer these

questions one by one. So let's write a

for loop. Okay for questions in questions.

questions.

Okay. So I I'm going to pick each

question one by one. So uh let's first

uh write a print statement to indicate

this is the question. Okay. So that we

know which question we are getting

answered for. Okay. So we are getting to

the next line and then this is the

question. Okay, first let's see the

question then let's get the response to

this question. Invoke

Invoke

question. Okay, let's not print

everything. Let's print just the first

maybe 200 characters because the the

answer is too long, right? So I don't

want to see the whole thing. That's not

the point here. So let's get the response

response

Okay. All right. So what I'm doing is

very simple. I just put the three

questions about the same thing in a

different style and then I'm just uh

sending these questions one by one to my

to the Gemini and then I'm getting the

response. I'm only printing the first

150 characters of these responses. I I

just want to see if there is if I'm

going to get exactly the same responses

or different responses. Okay, that's

what I want to test. Okay, so this is

the first question. Explain machine

learning. So we are waiting for the

response again. We're going to see the

and uh it's taking time. Yeah. So

imagine you want to teach a computer to

tell the difference between a picture of

a cat and a picture of as you see it

just begin to explain

to me and then the next tell me briefly

about machine learning. So machine

learning is a subset of artificial

intelligence that enables computers to

learn. Uh so what is this machine

learning? Imagine you want to teach a

computer to recognize cats and pictures

in traditional product. So I'm I'm

asking three questions about the same

thing with different styles. I'm getting answers.

answers.

You can see more or less the answers are

uh saying same thing but uh you can see

that uh depending on the question right

so the the answers are slightly

different uh so this is just to test

that you may see inconsistencies

about the responses you receive even

though you ask the same question with uh

uh different styles that's the first

thing and second thing is maybe I want

AI to explain to me uh about machine

learning in a certain style. For

example, first explain it briefly, then

give me an example. Uh uh okay. And then

tell me uh what is the role of machine

learning in the future. Okay. So maybe I

want him to explain any topic. It can be

it can be machine learning, it can be uh

uh quantum mechanics, it can be

blockchain. Okay. Any topic. First tell

me briefly what this topic is about.

then give me an example and tell me what

how this is going to be impactful in the

future. So, so if I want the AI to

explain things to me in a certain way,

so I need to tell this. So, I need to

tell this so that I will see consistency

in the responses. Okay, that's just an

observation. Nothing fancy going on

here. Okay. All right. Now, let's uh add

some structure, okay, to our uh what is

that um coding. So this time we are

going to interact with AI but our

interaction will be a little bit more

structured. Let me see if there are any.

Can you paste this in your question?

Yeah, I'm going to

here.

Okay. So what is the goal of this agent?

Right now there is no goal for this

agent. This agent is

a guru sitting on his chair and then

answering any question uh asked to him.

Okay. It's a very now right now this

this this agent knows claims to know

everything about anything. Okay. So then

later we want to specialize it to a

certain topic. All right. So that's what

we're going to do. All right. Provide

the function code. Yeah. I just provided

it. Yeah. Okay. Let's continue. So now

uh I'm going to uh import some

additional libraries uh from Google

Google gen AI import Google

Google

generative AI. Okay. I'm assuming we are

all new to this field. So I'm trying my

best to go slow. Okay. So and then if

you are behind let me know if you need

me share any of the code let me know.

Okay. Once import prompt

prompt

templates. Okay. So you can understand

from the libraries we are importing. We

are still going to use Google generative

AI but this time we are going to create

a prompt template. Okay. So that is

going to give some structure to the the

way we are asking the questions.

And then the last one from lang chain core

core

okay dot output.

I have some other examples on this third

one but I decided not to do it today. So

what we are doing now this this imported

library this third one is going to put

the response

uh which is the output coming from lench

chain coming from llm to a certain

format okay and then you can define this

format yourself here I'm not defining

the format myself I'm just going to use

the format uh coming with the library

okay all right so let me run this and

I provided the function code but uh I

I'll provide it again.

Oh, you mean this one? Sure.

Now, if you guys are ready,

we are here. We are going to

build another application, a small application.

application.

This time we are going to create a

prompt which has a certain structure.

Okay. So let's build a template for our

uh uh prompt which is going to make LLM

respond to us in a consistent way. Okay.

So I'm going to use prompt template.

Okay. So this prompt template has

certain parameters. First of all, what

is the input variables?

So the input variable is going to be let

me get to the next line so you can see

this. The input variable you can you can

do anything here. So the input variable

is topic. So in this prompt one of the

variables is topic. So the topic can be

again AI agents can be blockchain uh or

or can be weather can be a certain city.

So it can be anything. Okay. So that is

a variable in our template. So let me go

here. Okay. Comma. So this is one of the

first parameters of the template. The

second template. Now I'm going to write

my template. Template.

Okay. So here is the template. explain

now this is the prompt that I will give

to the AI tool but I want I don't want

to write a specific topic here so this

topic is a variable okay explain

whatever this topic is so when we create

the prompt that topic will change we are

going to define that topic for now I

don't know what that topic is I'm just

giving a template so you can fill in

empty spaces in this template based on

what you want to say okay explain

whatever this topic

in simple terms with an example. Okay.

So, I'm not going to stop here. I'm

going to even give further instructions.

So, here are the instructions. Okay. The

format is going to be as following. Keep it

under 100 words.

Okay.

And end with a end with a practical

example. This is what we want to see. So

I'm going to give you the topic but and

then here is how you are going to give

me topic is this which is topic.

Okay. This is how I I will see a a

response and then here is explanation.

This is the explanation and I am ending

my template. Okay. So this is my template.

template.

and then let me assign this template to

a variable. Let's use prompt. So I'm I'm

going to use this template for my next.

So prompt

remember prompt is the variable which is

holding the template. Now I'm going to

fill this template by again I have only

one variable here. You may have many

variables. You can add multiple

variables here. I have only one

variable. I just wanted to make it

simple. So now I am going to fill in the

empty spaces. topic is

uh AI agent. Okay. So that is my

template. So this going to be my field prompt.

Okay. Okay. So I just use the template

to create my prompt and then let's see

Okay, it's it looks something like

explain AI agent. So I replaced topic

with AI agent in simple terms with an

example blah blah. Topic is AI agent

explanation will follow.

Okay. Now let's see this is like a mail

merge, right? So there is a template

there are infinite number of variations

and then we are giving AI formatting

instructions. So this is going to ensure

consistency across all responses we

receive from the AI. Now, what I'm going

to do next is

I'm going to build a chain. Okay? I'm

going to call it chain.

All right? So, let me see if you have

any questions.

Uh, can you paste it, please? I'm

Okay. So, I'm pasting this one more

on on the chat. Okay,

All right. So, please let me know when

you have any questions. All right. Now,

I'm going to build a chain. So, now we

are talking about len chain. Okay. So,

I'm building a chain. So, this chain is

going to first take my rounds. So,

imagine that.

Imagine the order of things happening.

Okay. First

give the prompt.

Then connect to LLM and get the response.

response. Response

Response

then format the response

and give it as the out. Okay. So that's

what I want. I want to do three things.

First give the prompt then connect

connect to LLM and get the response and

then I would like to format the response

and give it as the output. That's what I

want to do. So here it is. This is the

prompt. This is the first step. So I'm

going to connect these uh steps uh uh

this way. So prompt line and then lm.

What is lm? lm is what we instantiated

earlier, right? So this one LLM I'm

using Google generative AI Gemini 2.5

flash with this API key. So this LLM is

our guy who is responding to us. Okay,

so we are asking these questions to LM.

Again you can name it anything you like.

My LLM your LLM. Okay, that is the

second step. And the third step is

whatever response is coming it's going

to be parsed output

in a certain format parser.

Okay. So first LLM is going to take the

input. So we are going to fill the

template send it this to Gemini. We're

going to get the response and then the

response is going to be cleaned up and

then we're going to get the output.

That's the order of things that will

happen. So this is our chain. All right.

So let's uh run this.

Nothing is going to happen. So we just

define the chain. Okay. Now what we're

going to do is we are going to give uh

three different topics and then we are

going to get the responses to these

topics. Okay. So let's do that. I'm

going to write a very simple code again.

This this time these are the topics.

I'm just going to give different topics.

Let's say blockchain things that you are

interested in you can give anything what

whatever you are interested in AI agent

okay and then what is that internet model

model

so these are things that I'm interested

in so okay so these are the topics now

let's go through the same uh for loop

that we did earlier for topic in topics

okay again I'm just going to print. I

can see the code here. I'm just typing

it anyways for for us. Okay.

So, I will first print the topic explaining

explaining

Okay. So, this is the topic. Now, this

is what's important, right? So, I'm

going to get the results

by LLM. Sorry, not LLM. I'm not using

LLM anymore because I already have LLM

in this pipeline chain right so I'm

going to use my pipeline chain so I'm

going to use chain

so that is a difference from the

previous case in the previous case we

were invoking we were getting response

directly from lm now I'm going to put my

uh prompt uh through this uh pipeline

Okay, now the chain is going to get an

input from us, right? What is that

input? You look look at the chain. Chain

is receiving the first step in the chain

is prompt. If you look at prompt, prompt

is expecting an input variable from us.

What is that topic? So here here is the

time for us to tell or give the chain

what the topic is. That's what we're

going to do. Okay. So here is what we we

topic. So this is this topic is the

variable this variable. Okay. And this

is the value for that variable that we

are assigning this topic. So this is

maybe maybe let me use a different name

here so that you're not going to be

confused with topic topic

or T in topics

topic is T. Okay. So T is the topic

name. Topic is the variable name. Okay.

Uh I think that's going to be better uh

for us. Uh and then let me take this u

uh what is that uh response first of all

uh u yeah let's just print it. Let's not

make this longer. So I'm just going to

say f response

is the result.

Okay. So let's see what's going to

happen now and then I will also share

the code with you. Let's me first make

sure that uh I have I don't have any

typos and there is typo

explaining oh this should be t2 right.

So t and t are referring to this t which

is coming from this list. So I'm I'm

running for loop uh through this list.

Right? So t is elements of this list.

Okay. Let me run this. Yeah. So it's

working. So now I gave a certain I gave

the instructions right. So first keep it

under 100 or it should be keep it use an

analogy and end with a practical

example. I'm just curious let's see

what's what what is said for blockchain

response blockchain is a shared

unchangeable digital ledger. Imagine a

public diary where each now it's giving

me the analogy right. Imagine a public

diary where each new page block is

cryptographically linked to the last

forming a chain. Once an entry is added

and verified by the network is virtually

impossible to alter. Okay. So this is

the uh what is that the analogy. Now

let's see if it is giving any practical

example. This decentraliz centralization

and transparency make it highly secure

and trusted. A practical example is

bitcoin. Now it's giving us a an

application of blockchain which is

bitcoin. That's perfect. Right? So it

just followed all the instructions and

then it kept with under 100 words and it

we just now we just structured the

response we are receiving from from our

LLM. I think this was a simple but a

good example. So let me copy this and

share with you guys if you are falling

Yeah. Uh so Janine so it's it's it will

be hard for me to to come up with the

version numbers now. Uh

again uh yeah I I had similar issues. So

I I I ran all these things uh on my

other VS codes and then uh I had some

compatibility issues. That's why I chose

to use Google Collab. I thought it would

be this it will be the same for all of

us. It will be much easier for all of

us. Yeah. Uh could could you please

share the prompt? Oh, sure. I'm sharing

Here it is.

Okay. It's a yeah

I'm seeing variance in response structures

Adam what is what do you mean variance

in this yeah so there is there is there

is a yeah there is there is a change in

the bet but this is this is the

structure we are expecting right so it's

exactly following other instructions

that just to see that we can control

this large lm is so nice right so we we

are making it follow a certain structure

yeah kumal You're going to receive an

email from us uh on on on this session

uh where you can see the recordings.

Recordings will be available at YouTube

channel, Magnum's YouTube channel. Okay,

let us continue. Now we going to do uh

uh maybe

uh what is that uh another example.

Okay. So here we learned about uh

creating a chain and that actually uh uh

helped us see a structured response from

the uh AI agent. Now we are going to

take one further step. Okay. Now so so

at first we use the row AI. Second we

structured our query. Third

let's say we want to build an agent

using this LLM. Still LLM is our guy.

Okay. It's waiting there for us uh to

help. So he will always help throughout

the session. He will always help to us,

give help to us. So now what we want to

do is we want to uh work LLM for us for

our company. Okay. Uh but for that we

need to give some information some

knowledge to LLM. So LLM doesn't know

us. He knows uh a lot about in general

about companies but he doesn't know

specifically about our company. Okay. So

it doesn't know about our company but it

knows in general what other companies

are doing. So let's first try to see how

LLM our guy is going to respond to uh to

this question. So I'm going to again uh

not not too what is that uh uh not too

different than what we did earlier. I'm

going to create a couple of questions

for for our LM but these questions are

now specific questions about uh about a

company policy. Okay. So here are some

some questions. Question one, what's

your return policy? What is your return

policy? Okay, it's a very valid

question, right? So, question mark.

Okay, that's the first question. Let's

give the second question. What are your

Okay, that's the second question. And

third, make sure there is a comma here.

And I'm going to give a short break.

Okay, after this one. And then what are

your support hours? You see these are

some valid questions that you can ask

from uh from any company right when you

are uh making any transaction with them. Okay.

Okay.

Now what I'm going to do is and I'm

going to go through these questions again

again

for Q in questions in company questions.

You are already familiar with this

format right? So we did this uh many

times now. So I'm going to say response is

I'm going to invoke

what is that uh my chain remember the

chain that we used

right so this time topic is these questions

questions

okay topic is this questions but uh I

mean u yeah let's see let's see what

what what how how it's going to work

because now I actually I don't want to

use an analogy it's not going to be

Let's just let's rather than using the

chain, let's just use LLM itself. Okay,

LM I'm just going to use LLM itself

because it doesn't make sense to have

any uh analogy for these questions,

right? So, let me use NLM.

Okay, walk. Uh what is that?

Q, right?

Q. All right. And then I will just print

the question. question

question

50 characters

again. Okay. So, let's see if this is

All right, let's see what what we what

we are going to receive.

Oh, we should have used uh

Yeah, it's not it's it's it's not

actually answering our our question. So

first of all, let me make this

Oh, I'm just printing 150th character of

the response. So that's why. Okay, I'll

I'll fix it. I'm going to print the

first 150 characters. So that's why you

see you we need to know how to deal with

um in this case strings, right? So I was

just printing the 150th character. So okay.

All right. Let's see. Yeah. What is your

return policy? Imagine you want to teach

a computer recognized GM in traditional

company questions?

It is still taking the queue from

previous case. Okay, let me fix this.

Let's take a break now. I'm going to fix

this. I'm not going to use Q here. So,

it's it's a good time to take a break.

It's 10:02 during the break time. I'm

going to fix the code. So, let's meet at

10 12. Okay. In 10 minutes, it's 10:02

Okay. Uh we're going to start in 3 minutes. So during this time, let me go

minutes. So during this time, let me go through the questions. First of all, I

through the questions. First of all, I just fixed that was a very simple

just fixed that was a very simple mistake I was making. I just fixed the

mistake I was making. I just fixed the uh the problem here. So, so we get the

uh the problem here. So, so we get the answers. But let's talk about this in 3

answers. But let's talk about this in 3 minutes. Uh getting an error on the

minutes. Uh getting an error on the first row. LM is not defined. So, we

first row. LM is not defined. So, we defined LLM at the very beginning

defined LLM at the very beginning literally. So, here LLM this one. So,

literally. So, here LLM this one. So, you need the Google API key for that.

you need the Google API key for that. So, this is the API key. I'm going to

So, this is the API key. I'm going to share the API key one more time.

This is the API key. you can use for this session. It's not going to be there

this session. It's not going to be there in maybe 30 minutes. And this is LLM.

This is how we define LLM. And uh so yeah, notebooks. I I can share this

yeah, notebooks. I I can share this notebook. It's not a problem. Uh let me

notebook. It's not a problem. Uh let me see what how we can do. So I can I can

see what how we can do. So I can I can put this one to my uh to my uh what is

put this one to my uh to my uh what is that?

that? uh GitHub account. So I think that's

uh GitHub account. So I think that's going to be the best. Okay. So I can

going to be the best. Okay. So I can upload this to the to my GitHub account

upload this to the to my GitHub account and then make it public so you can have

and then make it public so you can have access to this. But again we are going

access to this. But again we are going to have the recording anyways. So you

to have the recording anyways. So you can follow the recording as well. Let me

can follow the recording as well. Let me see if we'll be recording. Yeah. Could

see if we'll be recording. Yeah. Could you explain briefly what link chain is

you explain briefly what link chain is and what its relationship with LLM is?

and what its relationship with LLM is? Thanks. So link chain is uh as a

Thanks. So link chain is uh as a framework here. So what what it is

framework here. So what what it is allowing us is creating structured

allowing us is creating structured prompts. So again it's it's a lot more

prompts. So again it's it's a lot more than that but uh for now that's that's

than that but uh for now that's that's the use the use of lench chain for us.

the use the use of lench chain for us. We are creating structured prompts so

We are creating structured prompts so that we can communicate with our large

that we can communicate with our large language model. So it is a a way for us

language model. So it is a a way for us to communicate with LLM in a structured

to communicate with LLM in a structured manner. Okay, let's put it that way.

manner. Okay, let's put it that way. It's it's a lot more than this by the

It's it's a lot more than this by the way.

way. uh how would you pass out parameters

uh how would you pass out parameters from one step in the chain to the next?

from one step in the chain to the next? For example, LM has more than one

For example, LM has more than one variables in the response. Uh can I'm

variables in the response. Uh can I'm trying to understand. So how would you

trying to understand. So how would you pass output parameters from one step? Oh

pass output parameters from one step? Oh yeah, you can you can assign your output

yeah, you can you can assign your output to one of the variables to another

to one of the variables to another prompt and then create a chain like you

prompt and then create a chain like you get the response from the previous

get the response from the previous question but then the response to the

question but then the response to the first question will be an input to the

first question will be an input to the next prompt. So you can create a chain

next prompt. So you can create a chain of prompts and each prompt is using the

of prompts and each prompt is using the response to the previous prompt. Yeah.

response to the previous prompt. Yeah. But again uh this is the first time we

But again uh this is the first time we are uh I'm assuming that all most of us

are uh I'm assuming that all most of us are uh learning about this topic for the

are uh learning about this topic for the first time. So we are just uh navigating

first time. So we are just uh navigating uh entrance okay uh entrance door to to

uh entrance okay uh entrance door to to AI agents. Maybe uh in other workshops

AI agents. Maybe uh in other workshops we can go deeper and then create more

we can go deeper and then create more sophisticated ones. But this these are

sophisticated ones. But this these are very simple uh agents. We just want to

very simple uh agents. We just want to see the difference between just

see the difference between just communicating with chat GBT versus

communicating with chat GBT versus communicating with GPT with in a

communicating with GPT with in a structured manner.

Okay. Yeah. I think we are what extra does

Yeah. I think we are what extra does prompt give over simply invoking LM?

prompt give over simply invoking LM? What extra does prompt give over simply

What extra does prompt give over simply invoking?

invoking? I mean uh the prompt is set of the

I mean uh the prompt is set of the instructions right in addition to

instructions right in addition to instructions uh we can use the same

instructions uh we can use the same template for many different topics. So

template for many different topics. So for each question I don't I'm not I

for each question I don't I'm not I don't want to write a prompt. What I

don't want to write a prompt. What I want to do is I want to use in this case

want to do is I want to use in this case I'm just explaining regarding this case

I'm just explaining regarding this case I want to use a template which is going

I want to use a template which is going to only replace in this case only topic

to only replace in this case only topic so that I can send the prompt to to the

so that I can send the prompt to to the LM

LM could you put the link to the collab

could you put the link to the collab here please? JJ uh for that I need to

here please? JJ uh for that I need to share this one. Sure I can do that.

share this one. Sure I can do that. I need to find the Okay, so let's do

I need to find the Okay, so let's do this. Just I'm going to stop sharing for

this. Just I'm going to stop sharing for a minute and then uh in 30 seconds we

a minute and then uh in 30 seconds we are starting again. So let me find the

are starting again. So let me find the file in my Google Drive. Okay, just one

file in my Google Drive. Okay, just one second please

workshop called

called June 28,

June 28, 2025.

Okay, [Music]

[Music] just one second, please.

Okay, I got it. So, this is the link to this

So, this is the link to this file. I'll keep this file in my Google

file. I'll keep this file in my Google Drive so you guys can have access to

Drive so you guys can have access to this file. Okay. All right. Let me share

this file. Okay. All right. Let me share my screen now.

All right, let's continue guys. Let's continue. Uh there is uh there is more

continue. Uh there is uh there is more to do now. All right, now what we're

to do now. All right, now what we're going to do is first of all let's see

going to do is first of all let's see what happened here. Okay, so I asked

what happened here. Okay, so I asked three questions. What is your return

three questions. What is your return policy? What are your shipping costs?

policy? What are your shipping costs? What are So LM the guy that is

What are So LM the guy that is responding to us as you see gave us very

responding to us as you see gave us very gener general answers, right? So as an

gener general answers, right? So as an AI, I don't sell physical product or

AI, I don't sell physical product or services. I don't have a return policy.

services. I don't have a return policy. I'm here to provide. So there is no

I'm here to provide. So there is no return put because this guy is not

return put because this guy is not selling stuff as an AI I don't sell or

selling stuff as an AI I don't sell or ship physical. So I don't have any

ship physical. So I don't have any shipping cost. So yeah we are asking

shipping cost. So yeah we are asking some questions but these questions uh

some questions but these questions uh LLM cannot answer them because I mean uh

LLM cannot answer them because I mean uh even what are the what are the support

even what are the what are the support hours? I'm available 24 hours 7 days a

hours? I'm available 24 hours 7 days a week but that is only the LLM itself.

week but that is only the LLM itself. Okay. So but we want to ask these

Okay. So but we want to ask these questions we want to be able to answer

questions we want to be able to answer these questions for our own company.

these questions for our own company. Okay. So that's the next thing. What we

Okay. So that's the next thing. What we want to do is we want to provide some

want to do is we want to provide some input to our LLM so that LLM can use

input to our LLM so that LLM can use that input when responding to these

that input when responding to these questions. So that's the next thing

questions. So that's the next thing we're going to do now. So that is called

we're going to do now. So that is called rack solution retrieval augmented

rack solution retrieval augmented generation. So we are giving AI access

generation. So we are giving AI access to specific knowledge and then uh uh

to specific knowledge and then uh uh it's like giving giving AI a handbook

it's like giving giving AI a handbook about our company. So AI is going to

about our company. So AI is going to check that handbook before answering the

check that handbook before answering the questions.

questions. All right. So that's what we want to do.

All right. So that's what we want to do. So first of all that handbook again

So first of all that handbook again we're going to keep it very simple. This

we're going to keep it very simple. This can be very very sophisticated. We're

can be very very sophisticated. We're going to keep things very simple. This

going to keep things very simple. This here is our handbook. Okay. So we are

here is our handbook. Okay. So we are going to create documents.

going to create documents. These documents are this extra knowledge

These documents are this extra knowledge we are providing to LLM. So this is

we are providing to LLM. So this is about our company. Okay. So I'm just

about our company. Okay. So I'm just going to write this. We offer 30-day

going to write this. We offer 30-day returns

returns with receipts. Okay, that's our policy.

with receipts. Okay, that's our policy. Okay, second,

Okay, second, free shipping

free shipping on orders over $50. Okay, again, that's

on orders over $50. Okay, again, that's our second policy.

our second policy. Third, support available

9:00 a.m. to 5:00 p.m. Friday, Monday to Friday, Monday

Friday, Monday to Friday. So, I'll talk about this in a

to Friday. So, I'll talk about this in a moment. And then we

moment. And then we accept all major

accept all major credit cards. Okay, not paper. All

credit cards. Okay, not paper. All right. So these are our company

right. So these are our company policies. Now let's think how we can use

policies. Now let's think how we can use these documents. I can build an

these documents. I can build an application.

application. In this application uh if a customer

In this application uh if a customer arrives at my application at my website

arrives at my application at my website and then wants to ask a question the

and then wants to ask a question the application will receive that question.

application will receive that question. So it's almost like input right? So we

So it's almost like input right? So we are asking the customer input give me

are asking the customer input give me your question

your question of course we're going to say please give

of course we're going to say please give me your question your question

me your question your question and then what we can do is given this

and then what we can do is given this question

question we can check the words in the question

we can check the words in the question let's say if in the words it's about

let's say if in the words it's about returns

returns okay I can simply if for example if

okay I can simply if for example if input is a question

take this question and then split it into words and then check each word in

into words and then check each word in that question and then if you catch any

that question and then if you catch any words that are also in one of these

words that are also in one of these documents for example the question is

documents for example the question is what is your return policy right so one

what is your return policy right so one of the words return is matching with

of the words return is matching with this returns that means the question is

this returns that means the question is most likely about this document I can

most likely about this document I can simply report the first document so what

simply report the first document so what is your return policy the answer will we

is your return policy the answer will we offer further returns with receipts.

offer further returns with receipts. So to some extent that would solve the

So to some extent that would solve the problem to some extent. Okay. Again I

problem to some extent. Okay. Again I take the question I break the question

take the question I break the question into words. I go through the words in

into words. I go through the words in the question and then I count how many

the question and then I count how many of those words are actually I do the

of those words are actually I do the same thing about these documents. I also

same thing about these documents. I also break these documents down into words

break these documents down into words and then I compare what is the overlap

and then I compare what is the overlap between the question and these documents

between the question and these documents and the document that gives the most

and the document that gives the most overlap I'm going to pick that document

overlap I'm going to pick that document and then uh the posted as my response.

and then uh the posted as my response. Okay, I can use an if else statement for

Okay, I can use an if else statement for that. It's going to be a rulebased

that. It's going to be a rulebased system and to some extent that might

system and to some extent that might work. But let me ask you this.

work. But let me ask you this. Rather than doing this, how about I tell

Rather than doing this, how about I tell you that take the question

you that take the question and then find the closest document to

and then find the closest document to that question. Again, we are the logic

that question. Again, we are the logic is still the same, get the question like

is still the same, get the question like what is your return policy, but this

what is your return policy, but this time rather than going through each word

time rather than going through each word one by one. So if there was a way for me

one by one. So if there was a way for me to compare this question with these

to compare this question with these documents

documents not not word by word but uh if I could

not not word by word but uh if I could represent these documents and the

represent these documents and the questions in terms of numbers

questions in terms of numbers and then let's say one of the four let's

and then let's say one of the four let's say the first documents. So I create

say the first documents. So I create let's say uh numbers representing these

let's say uh numbers representing these documents. Now my documents will look

documents. Now my documents will look something like this. Okay I'm just I'm

something like this. Okay I'm just I'm just making up. Okay. 4 5 1 6.

just making up. Okay. 4 5 1 6. Four is representing the first document.

Four is representing the first document. Five is representing the second

Five is representing the second document. One is representing the third.

document. One is representing the third. Six is representing the fourth document.

Six is representing the fourth document. And the let's say question whatever the

And the let's say question whatever the question is. So what is your return

question is. So what is your return policy? Question is also represented as

policy? Question is also represented as a number. Let's say and then that number

a number. Let's say and then that number is I'm just making it up. I am just

is I'm just making it up. I am just making up. Okay. It is 3.9.

making up. Okay. It is 3.9. Okay. again how the number is coming

Okay. again how the number is coming from how why suddenly we are using

from how why suddenly we are using number for a question that we'll see in

number for a question that we'll see in a moment but what I'm doing is I take

a moment but what I'm doing is I take the question what is your return policy

the question what is your return policy and that is mapped to a some number that

and that is mapped to a some number that is 3.9 now I'm going to compare I'm

is 3.9 now I'm going to compare I'm going to measure the distance between

going to measure the distance between this number and all other numbers so

this number and all other numbers so this time the distance is very easy

this time the distance is very easy right so what is the distance between

right so what is the distance between 3.9 and four it's 0.1 what is the

3.9 and four it's 0.1 what is the distance between 3.9 and five it's 1.1.

distance between 3.9 and five it's 1.1. So I'm just taking the difference

So I'm just taking the difference between these numbers. So the closest

between these numbers. So the closest the smallest distance is four. That

the smallest distance is four. That means four is corresponding to this. We

means four is corresponding to this. We offer third terms with receipts. Okay,

offer third terms with receipts. Okay, that's the first uh novelty I'm bringing

that's the first uh novelty I'm bringing rather than counting the words. What I'm

rather than counting the words. What I'm doing is I am mapping all the documents

doing is I am mapping all the documents and the questions all the strings into

and the questions all the strings into numbers. Here I used only one number

numbers. Here I used only one number single numbers a scaler but I can use

single numbers a scaler but I can use bunch of numbers actually I can use

bunch of numbers actually I can use let's say 150 numbers for any document

let's say 150 numbers for any document and if you ask me how that's a different

and if you ask me how that's a different story okay but let's assume that it did

story okay but let's assume that it did that is doable that's the first novelty

that is doable that's the first novelty I introduced to this this very simple uh

I introduced to this this very simple uh what is that uh application now the

what is that uh application now the second one rather than just saying we

second one rather than just saying we offer third rate turn with receipts yeah

offer third rate turn with receipts yeah we know that the question is what is

we know that the question is what is your return policy? The answer is we

your return policy? The answer is we offer 30 returns with receipts. I also

offer 30 returns with receipts. I also want now my chatbot let's call it

want now my chatbot let's call it chatbot respond more humanely like uh

chatbot respond more humanely like uh thank you for asking this question. Uh

thank you for asking this question. Uh so we understand that you want to return

so we understand that you want to return your product and uh uh we just want to

your product and uh uh we just want to let you know that our we offer a

let you know that our we offer a thirdary return with receipts. Please

thirdary return with receipts. Please make sure that you bring your receipt

make sure that you bring your receipt with you when you want to return your

with you when you want to return your product. So what we are doing is we are

product. So what we are doing is we are actually answering the question but what

actually answering the question but what when we answer we are answering using

when we answer we are answering using this policy but in a in a more more what

this policy but in a in a more more what is that humanly way so that the other

is that humanly way so that the other person is going to be pleased to hear

person is going to be pleased to hear that response. So that is the second

that response. So that is the second novelty we want to bring. Okay. The

novelty we want to bring. Okay. The first novelty is the way we are

first novelty is the way we are searching through the documents finding

searching through the documents finding the best document which we can use in

the best document which we can use in our in our response to this question.

our in our response to this question. Actually that document will be that

Actually that document will be that document and the question together will

document and the question together will be a prompt we are sending to the LLM

be a prompt we are sending to the LLM and then LLM is going to response using

and then LLM is going to response using those two inputs. What is that the

those two inputs. What is that the document that is most related to the

document that is most related to the question and question itself and then

question and question itself and then LLM is going to use these two to create

LLM is going to use these two to create a better response rather than simply

a better response rather than simply saying we offer further returns with

saying we offer further returns with receipts. it's going to be more uh again

receipts. it's going to be more uh again we can even check control that how the

we can even check control that how the response is going to be how long it's

response is going to be how long it's going to be how it should maybe start

going to be how it should maybe start with a greeting to the customer anyways

with a greeting to the customer anyways so that's what we want to do now okay

so that's what we want to do now okay that's the next thing we want to do all

that's the next thing we want to do all right so let's get started then

right so let's get started then the first thing is let's talk about

the first thing is let's talk about these numbers we're going to call them

these numbers we're going to call them embeddings I'm going to write the word

embeddings I'm going to write the word embedding again this is a large topic if

embedding again this is a large topic if you has learned large language models.

you has learned large language models. Uh so this is a very important topic

Uh so this is a very important topic embedded. Uh so how we create

embedded. Uh so how we create embeddings? Embedding is again coming up

embeddings? Embedding is again coming up bunch of numbers which is representing a

bunch of numbers which is representing a document. Okay. And then we want to make

document. Okay. And then we want to make sure that these numbers are for any

sure that these numbers are for any document at equal length. Let's say I'm

document at equal length. Let's say I'm going to use 134 numbers or 300 numbers

going to use 134 numbers or 300 numbers for any document. How we come up with

for any document. How we come up with numbers as I said that's a different

numbers as I said that's a different story. Okay. for coming up with

story. Okay. for coming up with embeddings I'm going to use hugging face

embeddings I'm going to use hugging face are these what we call vectors yeah

are these what we call vectors yeah embedding vectors Adam is that's what we

embedding vectors Adam is that's what we call yeah I I don't want to use two so I

call yeah I I don't want to use two so I want to keep it as uh what is it again

want to keep it as uh what is it again we are I'm assuming that we are totally

we are I'm assuming that we are totally new to this domain all right so I'm

new to this domain all right so I'm going to use hugging face embeddings

going to use hugging face embeddings there are many different libraries we

there are many different libraries we can use or tools we can use to create

can use or tools we can use to create embeddings for this one I'm going to use

embeddings for this one I'm going to use hugging phase embeddings. First of all,

hugging phase embeddings. First of all, I'm not going to I'm not going to run

I'm not going to I'm not going to run this one. Okay, so let's um yeah, let me

this one. Okay, so let's um yeah, let me get to the next one. I'm going to use

get to the next one. I'm going to use embeddings. Embedings.

embeddings. Embedings. I will use hugging

I will use hugging face

face embeddings. Did I? But before I use

embeddings. Did I? But before I use hugging face embeddings,

hugging face embeddings, I think I need to import it from

I think I need to import it from lang chain embeddings

lang chain embeddings import

import hugging

hugging face

face embeddings again embeddings will create

embeddings again embeddings will create these numbers for the documents that's

these numbers for the documents that's all it's going to do okay that's what we

all it's going to do okay that's what we expect and then later I'm going to use

expect and then later I'm going to use this uh database this database is going

this uh database this database is going to hold these embeddings this database

to hold these embeddings this database is going to so for example when I

is going to so for example when I receive the question. How am I going to

receive the question. How am I going to go and dig and find the closest response

go and dig and find the closest response to that question? So, I need to keep

to that question? So, I need to keep these documents. After I find the

these documents. After I find the embeddings for these documents, I need

embeddings for these documents, I need to store them in a in a in a database so

to store them in a in a in a database so that it's going to be fast. So, I expect

that it's going to be fast. So, I expect a database that is going to be very

a database that is going to be very efficient in terms of finding finding

efficient in terms of finding finding the closest embedding really fast. Okay?

the closest embedding really fast. Okay? Because I don't want people wait when

Because I don't want people wait when talking with a chatbot. you ask a

talking with a chatbot. you ask a question, you you don't want to wait for

question, you you don't want to wait for 10 seconds to get the answer, right? You

10 seconds to get the answer, right? You want to get the answer right away. So, I

want to get the answer right away. So, I need something which works really fast

need something which works really fast so that uh when I receive a question, I

so that uh when I receive a question, I find the embedding for that question and

find the embedding for that question and I find the closest record from all the

I find the closest record from all the documents in my database. So, for that

documents in my database. So, for that I'm going to use a database, but not for

I'm going to use a database, but not for now. Okay, not for now. Okay,

now. Okay, not for now. Okay, embeddings. Let me use this word

embeddings. Let me use this word correctly. Embeddings. Hugging face.

correctly. Embeddings. Hugging face. It's a small A.

It's a small A. embeddings and uh I'm going to use a

embeddings and uh I'm going to use a specific model. Again, these model names

specific model. Again, these model names can change by time. Okay, this is all

can change by time. Okay, this is all mini L ln L6

mini L ln L6 D2. Okay.

D2. Okay. All right. So, let me run this and

All right. So, let me run this and hoping that it's going to work. So, we

hoping that it's going to work. So, we we're just going to wait for loading it.

we're just going to wait for loading it. Okay. And then during this time, let me

Okay. And then during this time, let me see if you guys have any questions at

see if you guys have any questions at the end. Can you explain how to get our

the end. Can you explain how to get our own API if you if you are expired? Yes,

own API if you if you are expired? Yes, John, it's very easy. I got this Gemini

John, it's very easy. I got this Gemini API. Go to your Google, go to just type

API. Go to your Google, go to just type Gemini API. Uh it's it's like 10

Gemini API. Uh it's it's like 10 seconds. So, you just going to use your

seconds. So, you just going to use your Google account to get API. Actually, you

Google account to get API. Actually, you can get API for free. Gemini gives you

can get API for free. Gemini gives you free API. Use that for your exercise.

free API. Use that for your exercise. Just keep in mind that it has a it has a

Just keep in mind that it has a it has a limit on I think per minute. Uh there

limit on I think per minute. Uh there are there is the limit on number of

are there is the limit on number of prompts you can send. Uh I didn't want

prompts you can send. Uh I didn't want to use the free one for this workshop

to use the free one for this workshop because uh I don't want to you know stop

because uh I don't want to you know stop in the middle right now. The API you are

in the middle right now. The API you are using I I'll be paying for it but it's

using I I'll be paying for it but it's it's not going to be there in in 30

it's not going to be there in in 30 minutes. But you can use Gemini API free

minutes. But you can use Gemini API free API uh for learning purposes and it's

API uh for learning purposes and it's not going to cost too much if you even

not going to cost too much if you even pay. I think it's going to cost maybe

pay. I think it's going to cost maybe $1, $2 uh for for your entire week of

$1, $2 uh for for your entire week of learning process. Okay,

let me see. Paste documents signing into with a different device. Okay.

with a different device. Okay. Paste documents to sell documents. I

Paste documents to sell documents. I didn't get this. Oh, this one. Sure,

didn't get this. Oh, this one. Sure, sure, sure. I will post it.

sure, sure. I will post it. Edwards, I got I understood your

Edwards, I got I understood your question. Okay.

question. Okay. Should we use the updated model due to

Should we use the updated model due to deprecation warning? Uh yeah, you can

deprecation warning? Uh yeah, you can but I I'm not worried about it now. So

but I I'm not worried about it now. So you can always go back and then update

you can always go back and then update your code. Uh that is polishing your

your code. Uh that is polishing your code for now. I'm okay with this one.

code for now. I'm okay with this one. It's it's working. I checked this one

It's it's working. I checked this one before the I checked each one of these

before the I checked each one of these things before the workshop. It's

things before the workshop. It's working. Okay. So we got the embedding.

working. Okay. So we got the embedding. Okay. We got the uh hugging face

Okay. We got the uh hugging face embeddings. I uh I have it here. So now

embeddings. I uh I have it here. So now now let's see what embedding means

now let's see what embedding means actually. Okay. So what I'm going to do

actually. Okay. So what I'm going to do is I'm just going to do a very simple

is I'm just going to do a very simple example. So I'm going to give some

example. So I'm going to give some sample text. Sample text.

sample text. Sample text. I'm sorry. Sample text.

I'm sorry. Sample text. These are the sample text. Return

These are the sample text. Return policy.

policy. Okay. Now we are getting into

Okay. Now we are getting into application, right? Refund policy.

application, right? Refund policy. something concrete and believe me it's

something concrete and believe me it's not going to take too much time for you

not going to take too much time for you to learn and

to learn and implement this locally. Now deploying it

implement this locally. Now deploying it as an application for business you need

as an application for business you need to you need to spend some more time on

to you need to spend some more time on that. Okay. So this is sample text. So

that. Okay. So this is sample text. So what I'm going to do now is now I

what I'm going to do now is now I already have embeddings right. So I'm

already have embeddings right. So I'm going to use embeddings. Okay. So this

going to use embeddings. Okay. So this instance again you can you can give any

instance again you can you can give any name here my embeddings your embeddings.

name here my embeddings your embeddings. I'm going to use embeddings. embedings

I'm going to use embeddings. embedings do embed documents

sample text. Okay. So, and then I'm going to assign this to let me sample

going to assign this to let me sample embeds. Okay. Let's see. Let me run

embeds. Okay. Let's see. Let me run this. And then uh sample text.

this. And then uh sample text. Okay. All right. It's it's very fast. I

Okay. All right. It's it's very fast. I got the embeddings. Now I'm just going

got the embeddings. Now I'm just going to show you uh the text length and

to show you uh the text length and embedding length. Okay. So let's write a

embedding length. Okay. So let's write a list comprehension for this one. For

list comprehension for this one. For text and sample text

text and sample text okay for text and sample text I'm just

okay for text and sample text I'm just going to have length of text. Okay. So

going to have length of text. Okay. So this is I'm just counting number of

this is I'm just counting number of characters. So this one is what is 5 6

characters. So this one is what is 5 6 11 right? I can see that. Okay, let me

11 right? I can see that. Okay, let me print this.

print this. This is the length of the text. So,

This is the length of the text. So, return policy is 13. Refund policy is

return policy is 13. Refund policy is 13. Leather forecast is 16. Okay. Now,

13. Leather forecast is 16. Okay. Now, I'm going to do the same thing for

I'm going to do the same thing for sample. Let me copy this.

sample. Let me copy this. Paste it here. So for I'm going to use

Paste it here. So for I'm going to use EMD embedding

EMD embedding in sample embeddings. And by the way,

in sample embeddings. And by the way, there is there's a typo here.

there is there's a typo here. Embeddings.

Embeddings. Okay.

Okay. Print the length of each embedding.

Print the length of each embedding. You got it right. So that's what I was

You got it right. So that's what I was trying to say. So the text no matter how

trying to say. So the text no matter how long. So now I'm going to return policy.

long. So now I'm going to return policy. You see the the length of the text is 13

You see the the length of the text is 13 13 16 but I'm using a standardized

13 16 but I'm using a standardized uh size of numbers representing these

uh size of numbers representing these text. So the first text is represented

text. So the first text is represented by 384 numbers. It's a vector of

by 384 numbers. It's a vector of numbers. The second one refund policy is

numbers. The second one refund policy is represented by 384. The third one is

represented by 384. The third one is represented by 384. Each of these are

represented by 384. Each of these are represented by a vector which is which

represented by a vector which is which has 384 numbers elements in it. So I can

has 384 numbers elements in it. So I can return policy for

return policy for purchased goods

purchased goods from our store. You see I made it even

from our store. You see I made it even longer. So let's see

longer. So let's see now I have 48 characters in the first

now I have 48 characters in the first text but still embeddings will remain

text but still embeddings will remain the so embedding changed. embedding for

the so embedding changed. embedding for the first text already changed but its

the first text already changed but its length is standardized it is so that's

length is standardized it is so that's the beauty no matter how long the

the beauty no matter how long the document is no matter how long the

document is no matter how long the document is I will always represent

document is I will always represent every document with the same length of

every document with the same length of vector in this case 384 so these vectors

vector in this case 384 so these vectors are not random numbers that's why it is

are not random numbers that's why it is beyond the scope of this workshop the

beyond the scope of this workshop the way we come up with these numbers is

way we come up with these numbers is also special so that for example one

also special so that for example one embedding ing if it is representing

embedding ing if it is representing let's say king.

let's say king. Okay. Another embedding is let's say

Okay. Another embedding is let's say representing queen. Okay. So queen not

representing queen. Okay. So queen not queen.

queen. Let's say these are two words king and

Let's say these are two words king and queen. I have embedding one for king.

queen. I have embedding one for king. Embedding one or king. Let's say I have

Embedding one or king. Let's say I have embedding two or queen. Let's say I also

embedding two or queen. Let's say I also have embedding three for men.

have embedding three for men. men and embedding for for woman. Okay.

men and embedding for for woman. Okay. So here is the nice thing about these

So here is the nice thing about these embeddings. The way they are generated

embeddings. The way they are generated is not random. Again this is uh an

is not random. Again this is uh an example which we need to check but more

example which we need to check but more or less this is what's going to happen.

or less this is what's going to happen. So what I can do is what I can do is

So what I can do is what I can do is I can take E3 men and then subtract

I can take E3 men and then subtract woman from men.

woman from men. Okay. E4. So this is giving me the

Okay. E4. So this is giving me the difference between men and woman. Okay.

difference between men and woman. Okay. And then when I add this to queen E2 you

And then when I add this to queen E2 you see queen and king woman and men I'm

see queen and king woman and men I'm taking difference between men and woman

taking difference between men and woman adding that to queen. So this is giving

adding that to queen. So this is giving going to give me a number which is very

going to give me a number which is very close to E1. You got it right? So what's

close to E1. You got it right? So what's happening here is these embeddings in

happening here is these embeddings in the space they are vectors

the space they are vectors are computed such that they are also

are computed such that they are also representing

representing the distance between the meanings of the

the distance between the meanings of the words and it's not just words it's

words and it's not just words it's meanings of bunch of words meanings of

meanings of bunch of words meanings of documents okay the the content of

documents okay the the content of documents so that's how I'm going to use

documents so that's how I'm going to use embeddings when I receive a question I'm

embeddings when I receive a question I'm going to find the closest answer to that

going to find the closest answer to that closest document to that question by

closest document to that question by comparing their embeddings. Okay, let me

comparing their embeddings. Okay, let me stop uh about embeddings. Okay, let me

stop uh about embeddings. Okay, let me continue with with the with the code.

continue with with the with the code. Okay,

Okay, [Music]

[Music] uh yeah and then I really so sorry. So

uh yeah and then I really so sorry. So yeah, you might receive these uh

yeah, you might receive these uh exceptions. Uh I I don't know what what

exceptions. Uh I I don't know what what is the cause of the error now. So uh

is the cause of the error now. So uh just follow the please follow the

just follow the please follow the workshop and then maybe you can try

workshop and then maybe you can try later on the yeah you can yeah I shared

later on the yeah you can yeah I shared the collab with you guys so you can you

the collab with you guys so you can you can check it okay let's finish this we

can check it okay let's finish this we don't have too much time how to know

don't have too much time how to know where to embed and give it number so

where to embed and give it number so that's John that's that's that's beyond

that's John that's that's that's beyond the scope of scope of this workshop okay

the scope of scope of this workshop okay how we can get these embeddings you can

how we can get these embeddings you can go back and then study a little bit on u

go back and then study a little bit on u how we create embeddings for for words.

how we create embeddings for for words. It's it's one of the first initial

It's it's one of the first initial topics when you learn NLP or large

topics when you learn NLP or large language models, right?

language models, right? Okay.

Okay. Now, let's let's really write the

Now, let's let's really write the application now. Okay. All right.

So this time I'm going to use as I said I I need to use a database lench chain

I I need to use a database lench chain vector stores

per stores I'm going to stop in 15 minutes okay

I'm going to stop in 15 minutes okay so base is uh what is that uh

so base is uh what is that uh a a database which is going to hold our

a a database which is going to hold our documents another uh

documents another uh um library we're going to use from lang

um library we're going to use from lang chain

chain lang chain

lang chain as schema actually for documents. So we

as schema actually for documents. So we are going to create documents in a

are going to create documents in a certain way and then we're going to use

certain way and then we're going to use this documents uh in our database. Okay.

this documents uh in our database. Okay. So let's

So let's this all right

this all right now remember our documents documents

now remember our documents documents remember the documents we created we

remember the documents we created we offer third rate return with receives so

offer third rate return with receives so we are still there okay we are still

we are still there okay we are still there so now I'm going to turn these

there so now I'm going to turn these documents to a specific format so for

documents to a specific format so for I'm going to use list comprehension for

I'm going to use list comprehension for documents in documents

okay So I'm going to take each document and turn it into a document format.

and turn it into a document format. Okay. So documents

page content. Basically this is creating this document is just a string. It is

this document is just a string. It is going to put that string into a document

going to put that string into a document template. Okay. So there is a metadata

template. Okay. So there is a metadata defining how a document is going to look

defining how a document is going to look like. Okay. we are we are actually

like. Okay. we are we are actually creating a document filling a document

creating a document filling a document template where we're going to use it to

template where we're going to use it to push it into our uh our database. So our

push it into our uh our database. So our database is expecting documents in a

database is expecting documents in a certain format. That's what we are doing

certain format. That's what we are doing here

here documents.

documents. So I'm getting each and every document

So I'm getting each and every document from my document list and turning it

from my document list and turning it into a document format and then I'm

into a document format and then I'm going to assign this list of documents

going to assign this list of documents to this variable. Let me print documents

to this variable. Let me print documents on the screen. See how they look like

on the screen. See how they look like you see? So I just changed I didn't

you see? So I just changed I didn't change too much, right? So I just took

change too much, right? So I just took this string but now I have a specific

this string but now I have a specific format. So document metadata is empty.

format. So document metadata is empty. Again this is a very basic uh what we

Again this is a very basic uh what we are doing is very basic and this is the

are doing is very basic and this is the content of the document. So you can you

content of the document. So you can you can also define the metadata for the

can also define the metadata for the document. So we are not there yet. All

document. So we are not there yet. All right. So now now this is a format which

right. So now now this is a format which our database is going to accept. Okay.

our database is going to accept. Okay. Now I'm going to take my database

Now I'm going to take my database and then it's going to take from these

and then it's going to take from these documents.

documents. Okay.

Okay. All right. And then documents,

All right. And then documents, comma. But before these documents, I

comma. But before these documents, I need to create the what is that the I

need to create the what is that the I need to also give the embeddings. This

need to also give the embeddings. This one what is embeddings? You see

one what is embeddings? You see embeddings is the name of the

embeddings is the name of the this embeddings object. You remember

this embeddings object. You remember this hugging face embeddings. So this is

this hugging face embeddings. So this is this is think of it as a box. From one

this is think of it as a box. From one side you put your documents from the

side you put your documents from the other side you get the embedding for

other side you get the embedding for that document. So this embedding is that

that document. So this embedding is that box. So what we are doing is to our

box. So what we are doing is to our database we are giving the documents and

database we are giving the documents and the box.

the box. Okay. By the way I should be consistent

Okay. By the way I should be consistent with the name embeddings. Embeddings.

with the name embeddings. Embeddings. Yeah. Correct. Okay. So the the the

Yeah. Correct. Okay. So the the the database is going to take our documents

database is going to take our documents in a certain format. It's going to take

in a certain format. It's going to take the embeddings box which is going to

the embeddings box which is going to create embeddings for the documents.

create embeddings for the documents. It's going to push the documents through

It's going to push the documents through that box. It will get the embeddings and

that box. It will get the embeddings and then there it's going to store those

then there it's going to store those embeddings for those documents nicely.

embeddings for those documents nicely. Okay, that let's put it that way. So

Okay, that let's put it that way. So that later on we can have access to

that later on we can have access to these embeddings easily and that let's

these embeddings easily and that let's call this vector store.

Okay. Now let me run this. Okay. Now if uh you want to see uh how this is

uh you want to see uh how this is working first of all vector store let's

working first of all vector store let's see what what we're going to see it's

see what what we're going to see it's not going to give us anything it's just

not going to give us anything it's just saying that vector store is is an object

saying that vector store is is an object something that is stored at this address

something that is stored at this address okay so we're not seeing the content of

okay so we're not seeing the content of that database all right so but that that

that database all right so but that that database is created and it is somewhere

database is created and it is somewhere stored in the memory now let's test

stored in the memory now let's test this. Let's test this. For example, I'm

this. Let's test this. For example, I'm going to run a query. Test query.

going to run a query. Test query. Okay. And refund

Okay. And refund policy.

policy. Remember, one of the documents is one of

Remember, one of the documents is one of the documents is uh what is that? We

the documents is uh what is that? We offer 30-day returns with receipts.

offer 30-day returns with receipts. So, this is the question refund policy.

So, this is the question refund policy. What is your refund policy? Now, here is

What is your refund policy? Now, here is how it's going to work. I'm going to

how it's going to work. I'm going to take my

take my database. Remember our database already

database. Remember our database already has this box embeddings. Okay. So, I'm

has this box embeddings. Okay. So, I'm going to take and then I'm going to ask

going to take and then I'm going to ask my database, can you search

my database, can you search for similar

search? Can you f search in your database? Can you search in your records

database? Can you search in your records for similar things to this test query

for similar things to this test query refund policy? Okay, let me put this

refund policy? Okay, let me put this inside test

inside test query. So what this vector store is

query. So what this vector store is going to do is it's going to take in

going to do is it's going to take in this refund policy query. It's going to

this refund policy query. It's going to push that query through this embeddings

push that query through this embeddings box and it will receive the embeddings

box and it will receive the embeddings for this query refund policy. It will

for this query refund policy. It will find the embeddings value embeddings

find the embeddings value embeddings vector for refund policy and then it's

vector for refund policy and then it's going to compare that embedding with all

going to compare that embedding with all the embeddings in the record. Okay. So,

the embeddings in the record. Okay. So, how many embeddings we have? 1 2 3 4.

how many embeddings we have? 1 2 3 4. There are four embeddings in the

There are four embeddings in the database. It's going to find the

database. It's going to find the embedding for refund policy. It'll

embedding for refund policy. It'll compare. It'll find the most similar

compare. It'll find the most similar here. We're going to say, "Give me the

here. We're going to say, "Give me the most similar one document." If I say two

most similar one document." If I say two document, it's going to give me the most

document, it's going to give me the most similar two documents in order. I'm

similar two documents in order. I'm going to say the give me the most

going to say the give me the most similar document to this embedding.

similar document to this embedding. Let's see what it's going to return.

Let's see what it's going to return. Okay. So, let me run this.

Okay. So, let me run this. You see? So it already returned the ID

You see? So it already returned the ID of that document,

of that document, its metadata and its page content. Okay.

its metadata and its page content. Okay. So let's make this let's let's make this

So let's make this let's let's make this more uh put in a more nice nicer format.

more uh put in a more nice nicer format. So I'm going to say this is my relevant.

So I'm going to say this is my relevant. You got the you got you got what's

You got the you got you got what's happening here, right? Relevant relevant

happening here, right? Relevant relevant documents. Okay. So you you got the rest

documents. Okay. So you you got the rest of I asked refund policy. It easily

of I asked refund policy. It easily found the most relevant document to this

found the most relevant document to this question. I didn't search for words.

question. I didn't search for words. Remember that was our if as rulebased

Remember that was our if as rulebased system. We would break down the

system. We would break down the queries and the documents into words and

queries and the documents into words and then find the word frequency map

then find the word frequency map matchings. We didn't do any of that. We

matchings. We didn't do any of that. We turned everything into a a set length of

turned everything into a a set length of vector 384 long vector and then we just

vector 384 long vector and then we just compared those vectors and then found

compared those vectors and then found the closest vector to refund policy

the closest vector to refund policy vector. That's what we did. So let's

vector. That's what we did. So let's make this a little bit nicer. So I'm

make this a little bit nicer. So I'm going to print

going to print uh uh print.

[Music] Yeah, you guys must be tired now. So I'm

Yeah, you guys must be tired now. So I'm I will get the first element and then

I will get the first element and then from that I'm going to get page

from that I'm going to get page contents. So this is a list. As you see

contents. So this is a list. As you see this is a list. This is a list. So there

this is a list. This is a list. So there is only one element here because I asked

is only one element here because I asked only the closest one document. I'm

only the closest one document. I'm taking the first document in that list

taking the first document in that list and the first documents page content.

and the first documents page content. Okay. Page content.

Okay. Page content. Okay.

What happens? page content relevant documents

relevant documents zero

zero and then

and then page content

page content relevant.

Oh, I need to put F in here. Sorry. Yeah. So, we offer 30-day returns with

Yeah. So, we offer 30-day returns with receipts. Perfect. So we got the most

receipts. Perfect. So we got the most relevant document uh to our now let us

relevant document uh to our now let us create our what is that u agent AI

create our what is that u agent AI agent. So this time so what we going to

agent. So this time so what we going to do is now we understood that we can

do is now we understood that we can have a list of documents that is only

have a list of documents that is only known to us for our own organization.

known to us for our own organization. LLM doesn't know those documents. LLM

LLM doesn't know those documents. LLM has a generic idea about the world.

has a generic idea about the world. Okay, but we want to

Okay, but we want to create prompts

create prompts for large language model that is going

for large language model that is going to be about our own company uh

to be about our own company uh documents. Okay. And then we want to

documents. Okay. And then we want to when we receive a question, we want to

when we receive a question, we want to find the most relevant document from our

find the most relevant document from our database. And then we're going to put

database. And then we're going to put this question and document together, put

this question and document together, put into a prompt, send it to LLM and ask

into a prompt, send it to LLM and ask LLM to give a nice answer using these

LLM to give a nice answer using these two documents. That's what we're going

two documents. That's what we're going to do now. And then I'll finish with

to do now. And then I'll finish with that. I have one more uh application,

that. I have one more uh application, but I think we are out of time. So, and

but I think we are out of time. So, and then I see that you guys also are

then I see that you guys also are getting tired. Relevance of K plus what

getting tired. Relevance of K plus what is K= 1 referring to? So, if I make this

is K= 1 referring to? So, if I make this K= 2, I'm finding from this you have I

K= 2, I'm finding from this you have I have four documents. when k equals 1 I'm

have four documents. when k equals 1 I'm finding the out of these four documents

finding the out of these four documents I'm finding one document which is the

I'm finding one document which is the most which is the closest in terms of

most which is the closest in terms of relevance to refund policy string if I

relevance to refund policy string if I make this two then I'm going to have uh

make this two then I'm going to have uh so let me

Actually, let me let me print here. Relevant

Relevant documents.

documents. Let's not change this one. Okay, let run

Let's not change this one. Okay, let run this. Let run this relevant document.

this. Let run this relevant document. Now you see I have two documents because

Now you see I have two documents because I changed this. So these are the two

I changed this. So these are the two documents that are in terms of relevance

documents that are in terms of relevance closest to this qu question query.

closest to this qu question query. refund policy. Okay, that's what it

refund policy. Okay, that's what it means. Highest similarity, shortest

means. Highest similarity, shortest vector path. Yes, the shortest. We are

vector path. Yes, the shortest. We are just measuring the distance between two

just measuring the distance between two vectors actually here. It's got to be

vectors actually here. It's got to be cosine similarity. Adam, uh more

cosine similarity. Adam, uh more technically it's a it's a cosine

technically it's a it's a cosine similarity. We are looking at the angle

similarity. We are looking at the angle between two vectors. So if two vectors

between two vectors. So if two vectors are closer to each other, cosine of that

are closer to each other, cosine of that angle will be the smallest.

angle will be the smallest. You're welcome.

You're welcome. Not tired. Okay. Why? So wow. All right.

Not tired. Okay. Why? So wow. All right. So let's let's finish this guys. Let's

So let's let's finish this guys. Let's finish this one. Now let's uh create our

finish this one. Now let's uh create our uh new application. All right. Okay.

uh new application. All right. Okay. Let's see. So let's write this as a as a

Let's see. So let's write this as a as a function together. Uh so this is the R A

function together. Uh so this is the R A rag. So Dr. Jam, maybe before you start,

rag. So Dr. Jam, maybe before you start, I make the last announcement. Oh, sure.

I make the last announcement. Oh, sure. Please go ahead. Okay. Sorry. Just give

Please go ahead. Okay. Sorry. Just give me time. So I hope you guys enjoying I

me time. So I hope you guys enjoying I see that participants level is almost

see that participants level is almost 98% from starting it means that people

98% from starting it means that people really enjoy we are over 100 people

really enjoy we are over 100 people here. So uh if you want to continue and

here. So uh if you want to continue and excel in LLMs which most jobs are really

excel in LLMs which most jobs are really looking that skill set LLMs and AI

looking that skill set LLMs and AI agents through that six weeks program we

agents through that six weeks program we mentioned in the beginning we have

mentioned in the beginning we have actually a early bird discount till end

actually a early bird discount till end of the month in couple days two days and

of the month in couple days two days and it is $500 original price $800.

it is $500 original price $800. recommend not to mess it. Not on early

recommend not to mess it. Not on early bird discount, we have a limited space.

bird discount, we have a limited space. We may close the uh uh we may close the

We may close the uh uh we may close the registrations in after maybe two or

registrations in after maybe two or three days. Uh just encourage to

three days. Uh just encourage to register if you want to get more in

depth and you will be also joining Magniman alumni where we also offer

Magniman alumni where we also offer special uh classes to just alumni only.

special uh classes to just alumni only. So thank you Dr. Jang. So you continue.

So thank you Dr. Jang. So you continue. Thank you. I just want to add my my

Thank you. I just want to add my my colleague uh Sei is an MIT graduate.

colleague uh Sei is an MIT graduate. He's really good what at what he's

He's really good what at what he's doing. He's actually working in the

doing. He's actually working in the Silicon Valley. Uh he he's he's an

Silicon Valley. Uh he he's he's an expert on this. So he will be giving

expert on this. So he will be giving half of that workshop and then I'll give

half of that workshop and then I'll give the other half. So just for information.

the other half. So just for information. Okay. So let's continue. Let's now write

Okay. So let's continue. Let's now write the function which is going to be our

the function which is going to be our application. I'm going to call this ask.

application. I'm going to call this ask. I'm sorry here.

I'm sorry here. ask rag. So this time I'm going to

ask rag. So this time I'm going to create an application which is going to

create an application which is going to receive a question okay from outside and

receive a question okay from outside and then uh exactly we are going to look for

then uh exactly we are going to look for the most relevant document to that

the most relevant document to that question and then we're going to send

question and then we're going to send question and that most relevant document

question and that most relevant document to our LLM and then we are going to get

to our LLM and then we are going to get the response to that. So let's see how

the response to that. So let's see how it's going to work. Let's see how it's

it's going to work. Let's see how it's going to work. Okay, first uh so we are

going to work. Okay, first uh so we are going to find the relevant documents. So

going to find the relevant documents. So relevance

relevance document.

document. So we already document I'm sorry

So we already document I'm sorry document we already have our vector

document we already have our vector space here. Vector store. I'm going to

space here. Vector store. I'm going to use the same vector store. Otherwise we

use the same vector store. Otherwise we should have put this inside the

should have put this inside the function. I'm not doing that. So and

function. I'm not doing that. So and then I am searching for similarity

then I am searching for similarity search. Okay. And then uh I will say I'm

search. Okay. And then uh I will say I'm looking for similarity search for this

looking for similarity search for this question. Let's set K to two this time.

question. Let's set K to two this time. Okay. So earlier it was one. Now I'm

Okay. So earlier it was one. Now I'm looking for the most relevant two

looking for the most relevant two documents to this question. Okay. So let

documents to this question. Okay. So let now

now next what I'm going to do is sorry next

next what I'm going to do is sorry next if you like you can print those

if you like you can print those documents on screen if you like just to

documents on screen if you like just to if you want to see how they look like

if you want to see how they look like but let's not do it. We already did that

but let's not do it. We already did that part. Uh printing the content of the

part. Uh printing the content of the documents. We already did that. Now I'm

documents. We already did that. Now I'm going to put these documents, join them

going to put these documents, join them together and then create one context

together and then create one context from that. Okay. So for that I'm going

from that. Okay. So for that I'm going to first create a uh use what is that uh

to first create a uh use what is that uh uh list comprehension for documents in

uh list comprehension for documents in relevant document.

relevant document. Okay, I'm going to put these documents

Okay, I'm going to put these documents together. So for that uh I will get the

together. So for that uh I will get the first I will get the page content of

first I will get the page content of each document. I'm sorry page content.

each document. I'm sorry page content. You know what this page content does

You know what this page content does right here. Page content it's just from

right here. Page content it's just from this document. So this is the document.

this document. So this is the document. Okay let's say this is the document the

Okay let's say this is the document the most relevant document. I'm only taking

most relevant document. I'm only taking the page content part of that document.

the page content part of that document. This 30-day we offer we offer 30-day

This 30-day we offer we offer 30-day return with receipt. I just take that

return with receipt. I just take that part. Let's say I have two documents. I

part. Let's say I have two documents. I take that part of these two documents

take that part of these two documents and create a list of that. Okay. So

and create a list of that. Okay. So actually let me write let me run this

actually let me write let me run this here

here for you this one. So let me first print

for you this one. So let me first print this. I just want you guys see what I'm

this. I just want you guys see what I'm doing inside the function.

doing inside the function. This one

This one that's exactly what the look look this

that's exactly what the look look this is the this is the two documents. The

is the this is the two documents. The first list is the two documents. Second

first list is the two documents. Second is only the list of the document content

is only the list of the document content of those two documents. this one. This

of those two documents. this one. This one, the second, that's what we did

one, the second, that's what we did here. Now, next what I'm going to do is

here. Now, next what I'm going to do is I'm going to join them. So, let's let's

I'm going to join them. So, let's let's see how it's going to work. So, I will

see how it's going to work. So, I will join them uh uh

join them uh uh like I I I'll put first document content

like I I I'll put first document content and I'll put the second one right below

and I'll put the second one right below that one. For that, you know, you know

that one. For that, you know, you know the join uh function, right? So what I'm

the join uh function, right? So what I'm going to do is first I will write a

going to do is first I will write a string that string is slashn / slashn is

string that string is slashn / slashn is going to go to the next line. Right? So

going to go to the next line. Right? So I will join these two documents using

I will join these two documents using slashn. That's what I'm doing. That

slashn. That's what I'm doing. That means in between these two documents I

means in between these two documents I will put slashn. So let let's show what

will put slashn. So let let's show what what

what I do. I need this one. Okay here you see

I do. I need this one. Okay here you see this I put in between two documents I

this I put in between two documents I put this slashn. Basically first I'm

put this slashn. Basically first I'm just telling

just telling that this this is the first document

that this this is the first document content and this is the second document

content and this is the second document content. I'm just creating the context

content. I'm just creating the context for that. So let's do this

slashn dot join. Perfect.

dot join. Perfect. We don't know to do this. We don't need

We don't know to do this. We don't need to do this but it's going to be easier

to do this but it's going to be easier if we join all the documents and then

if we join all the documents and then call it a context. Okay. Now let's

call it a context. Okay. Now let's create our prompt. We are experts of

create our prompt. We are experts of creating prompt, right? We already did

creating prompt, right? We already did that earlier. I'm going to we're going

that earlier. I'm going to we're going to follow the same steps. So this is my

to follow the same steps. So this is my prompt template. We know how to create a

prompt template. We know how to create a template. Now

in between this I'm going to write my template. Okay. So based on this company

template. Okay. So based on this company information

information okay so I'm now this is my instructions

okay so I'm now this is my instructions to LLM take this context

to LLM take this context now this context is coming from here

now this context is coming from here right that is the input to our prompt

right that is the input to our prompt and then take this question

okay take this question and then this is my instruction ction. Please, please

my instruction ction. Please, please provide

provide a helpful

a helpful answer

answer based on the company information

based on the company information provided.

provided. Okay. And then answer. You see this is

Okay. And then answer. You see this is my template. Perfect. Next.

Now this template is going to receive context from here question as the

context from here question as the function parameter. Okay. All right. All

function parameter. Okay. All right. All right. Now I'm going to get the

right. Now I'm going to get the response. Response

response. Response sorry

sorry response is llm.invoke.

response is llm.invoke. Here if you like you can create a chain

Here if you like you can create a chain too. Remember we created a chain. Now

too. Remember we created a chain. Now let's keep things simple for now. Let's

let's keep things simple for now. Let's see what is the response. Okay. And then

see what is the response. Okay. And then I'll get the front rag.

I'll get the front rag. Nothing stops us create from creating a

Nothing stops us create from creating a chain here. I just don't do it here for

chain here. I just don't do it here for now. And then return

now. And then return response. That's it. So I just created a

response. That's it. So I just created a very simple rag application. Okay. So

very simple rag application. Okay. So let's test it. Question.

let's test it. Question. This is what the person is going to ask.

This is what the person is going to ask. What is

What is your

your return policy?

return policy? Okay, now let's get the response answer.

Okay, now let's get the response answer. I will ask my

I will ask my AI agent. Hey, give me the answer for

AI agent. Hey, give me the answer for this question. Now, let's print this

this question. Now, let's print this answer.

answer. Okay, let's see if this is going to

Okay, let's see if this is going to work, guys. If it works, we are fine.

work, guys. If it works, we are fine. Okay, let's see. And most of the time it

Okay, let's see. And most of the time it doesn't work the first try. Oh, it it

doesn't work the first try. Oh, it it did. You see it says we offer 30 returns

did. You see it says we offer 30 returns with a receipt. So it is we offer 30

with a receipt. So it is we offer 30 returns with receipts. It just added a

returns with receipts. It just added a receipt. So that's it was a very uh so

receipt. So that's it was a very uh so we can add please provide helpful answer

we can add please provide helpful answer provided uh please.

like that. And then first

first greet the customer.

greet the customer. Say a few nice words. Okay.

Hello there. We're happy to help you with your question. Our return policy is

with your question. Our return policy is quite straightforward. We offer 30

quite straightforward. We offer 30 returns for items when accompanied by

returns for items when accompanied by receipt. We also accept all major credit

receipt. We also accept all major credit for a convenience. Perfect. Right.

for a convenience. Perfect. Right. Perfect. So uh just imagine that these

Perfect. So uh just imagine that these uh these documents are not just simple

uh these documents are not just simple list of sentences. You can have actual

list of sentences. You can have actual documents really long documents, right?

documents really long documents, right? And each of these documents are going to

And each of these documents are going to be squeezed into a fixed length of

be squeezed into a fixed length of vectors and then they will be stored in

vectors and then they will be stored in a database which is going to allow you

a database which is going to allow you to search for them very quickly. And

to search for them very quickly. And then you can turn any question into a

then you can turn any question into a vector and you can find the most similar

vector and you can find the most similar vectors from your vector space and then

vectors from your vector space and then you can send those documents with the

you can send those documents with the question to your LLM and then LLM is

question to your LLM and then LLM is going to get back to you based on your

going to get back to you based on your instructions and and we also learned how

instructions and and we also learned how to create a prompt uh a template for a

to create a prompt uh a template for a prompt. I think this was very useful. So

prompt. I think this was very useful. So I had one more application that I was

I had one more application that I was planning to share with you today. But I

planning to share with you today. But I think I think we did we did good today.

think I think we did we did good today. Uh maybe next time. All right. Thank you

Uh maybe next time. All right. Thank you so much for listening. Uh our time is up

so much for listening. Uh our time is up and u hope to see you guys in another

and u hope to see you guys in another workshop. All right. So I shared the

workshop. All right. So I shared the document with you. I will close API

document with you. I will close API right away. So that means you need to

right away. So that means you need to come up with your own API to make the

come up with your own API to make the code work. Otherwise you're going to

code work. Otherwise you're going to receive that the API key is is no longer

receive that the API key is is no longer valid. you're going to receive that that

valid. you're going to receive that that kind of a response. Okay, thank you and

kind of a response. Okay, thank you and take care. Uh see you next time.

Click on any text or timestamp to jump to that moment in the video

Most transcripts ready in under 5 seconds

One-Click Copy125+ LanguagesSearch ContentJump to Timestamps

Paste YouTube URL

Enter any YouTube video link to get the full transcript

Most transcripts ready in under 5 seconds

Get Our Chrome Extension

Get transcripts instantly without leaving YouTube. Install our Chrome extension for one-click access to any video's transcript directly on the watch page.

Add to Chrome — Free

Works with YouTube, Coursera, Udemy and more educational platforms

Get Instant Transcripts: Just Edit the Domain in Your Address Bar!

YouTube

←

→

↻

https://www.youtube.com/watch?v=UF8uR6Z6KLc

YoutubeToText

←

→

↻

https://youtubetotext.net/watch?v=UF8uR6Z6KLc

YouTube TranscriptPreparing your results…

YouTube Transcript:AI Agents from Scratch: A Beginner's LLM Workshop

AutoDub

Video Transcript

Summary

Core Theme

Paste YouTube URL

Transcript Extraction Form

Get Our Chrome Extension

Get Instant Transcripts: Just Edit the Domain in Your Address Bar!

YouTube Transcript:
AI Agents from Scratch: A Beginner's LLM Workshop