So, this rag agent took me less than 5

minutes to set up. This is the easiest

method I've ever tried. I'm going to

open up this chat and paste in this

message, which has three different

queries. The first one is, "What was

Tesla's total revenue Q2 2025?" The

second one is, "What was Nvidia's total

revenue Q1 fiscal year 25?" And the

third one is, "What were Nike's revenues

for Q4 fiscal year 2025?" You can see

the agent hit its pine cone tool three

times. And if we look at its answers,

not only is it giving us the correct

answers, but it's giving us the exact

document, the page numbers, and the

exact quote that it found in the PDF.

And that's how we can trust the answers

are correct. So, starting off with

Tesla, if I take the exact quote that it

found in the PDF, and I go into the

Tesla document, and we do a quick

search, we can see that it pulled it

from page 4. And back in Naden, our

assistant said that it found it from

page 3 to 7 on this exact document.

Next, if we go to the Nvidia answer and

we pull the exact quote and then I

control F that in this document, you can

see that it finds it right here. And

this is page one of the PDF and we go

back into Nit and it says that it found

it on page one of the NVIDIA annual

report document. And then finally, if we

go to the Nike answer and we grab this

quote that it found, switch into the

Nike document and do another control F,

we can see the exact quote was found

right here on page one. And in edit, the

assistant said that it found it on page

one. So, those of you that have built

rag agents before, you know that it's

not super easy to be able to get your

agent to actually site exactly where

it's pulling everything from. It's

doable, but you have to set up a

pipeline to do things like metadata

tagging. And like I said, all I did here

was drop in a file and then chat with

it. So, that's exactly what I'm going to

be showing you guys how we do that

today. Okay. So, now that you've seen

that quick demo, I'm just going to

basically build that exact system in

front of you guys step by step. So, feel

free to open up your computer and follow

along with what I'm doing. but also the

entire template that you'll see today

will be available for download for free.

All you have to do is join my free

school community. That way you can just

plug and play and test it out for

yourself. So what we're doing today is

we are using a Pine Cone assistant as

our rag search. So if you've never used

Pine Cone before, it is a vector store,

but in here you can see that we have a

section right over here called

assistant. So if I click on that, you

can see that I have a test assistant,

which is the one we were just using. But

what I'm going to do today real quick is

create a new one in front of you guys.

I'm just going to call this one demo.

and I'm going to create the assistant.

Now, real quick, what you want to pay

attention to is down here at the bottom.

It says active assistants have a fee of

5 cents per hour. So, it's not too bad.

And you can also click right here for

more details about the pricing. So, now

that we got that out of the way, I'm

going to go ahead and create this

assistant. And what you can see is this

kind of looks like a custom GPT

interface where you've got a chat right

here and then you have the ability to

drag in files. So, what I'm going to do

is drag in those three files that we

were looking at earlier, the Tesla,

Nike, and Nvidia earnings reports. And

I'm just going to import those right

away into our Pine Cone Assistant. And

now that those are uploaded in Pine Cone

Assistant, I can paste in that exact

same question that I asked our NADN

agent in that demo. So you can see it

comes back with the correct answers. And

we could hover over each of these links

because we're in the playground mode to

see what PDF it pulled from and then the

page numbers that it got that

information from. So the power of this

comes through the actual API that we can

use to talk to this Pine Cone Assistant.

And like I said, don't worry if this

seems confusing. I'm going to show you

guys exactly what we have to do. And

it's actually super easy. So you can see

over API, there's two things we can do.

We can upload files or we can chat with

our assistant. And in today's video,

we're just going to be focusing on

chatting with our assistant because we

already uploaded our files in this

interface right here. So I'm going to go

back into NADN and we're just going to

real quick grab an AI agent. And now

we're going to be able to hook it up to

this Pine Cone Assistant tool. So when

we give our agent a tool, you can see

that there's no native tool here for a

Pine Cone Assistant, but there is for

Pine Cone Vector Store. So, what we need

to do is grab an HTTP request so we can

talk to our assistant. So, all you're

going to do is go back into Pine Cone

over here. You're going to open up this

little connect thing to get to the API.

And then down here where you see chat

with your assistant, I'm basically going

to copy everything except for the top

two rows of Pine Cone API key and this

blank row. So, I'm going to copy this,

go back into edit, hit import curl, and

then paste that in there. And when I hit

import, it basically fills in the entire

HTTP request that we need except for a

few things that we'll have to tweak. So

the first thing is a Pine Cone API key.

The way you get that is you go back into

Pine Cone. You'll click over here on API

keys and then you'll just create a new

key. So here's what that looks like.

You'd click create API key. I'm just

going to call this one demo 4. We're

going to create that key. And now it's

going to give us this secret value. And

I have to copy this. And after you save

it and you click close, it's going to be

gone. So you can always create a new

one, but just keep in mind that that key

will be gone. So then you'd come back

into Pine Cone and just paste that key

right in there and you'll be all set. So

it's super simple. And one pro tip that

I want to show you guys is if you

already have connected to Pine Cone

through a native node, you could come to

authentication, you can go to predefined

and then you could come in here and just

type in Pine Cone. And that would pull

in your credential of your native Pine

Cone integration. So you can either do

that or you can do this. They're both

the same. So because I'm doing this, I'm

just going to turn off my headers. But

don't let that confuse you. All that is

is basically just putting in our

password so we can access that Pine Cone

Assistant. So now what I'm going to do

is change the description and I'm just

going to say use this tool to talk to

your knowledge base. So we're going to

paste in that as a description. I'm

going to name this tool Pine Cone just

to keep everything simple. And then the

last thing we have to set up is the

actual request that we're making to our

Pine Cone assistant. So right now if we

triggered this tool, no matter what we

asked our assistant for help with, it

would be sending off this query to the

assistant. What is the inciting incident

of Pride and Prejudice? And so we don't

want to have that be the query every

single time. So what we have to do is

change this to an expression, meaning

that these values can be dynamic. I'm

going to get rid of this content right

here. And I'm going to type in two curly

braces that face each way. We're going

to do a dollar sign, and we're going to

grab this from AI function. And this is

just going to let our AI agent determine

what query do I send to the Pine Cone

Assistant. And so all I'm going to do is

type in search query within that little

quotation mark right there. And now this

is basically set up to be dynamic. And

I'll show you guys exactly what I mean

by that if that isn't yet making sense.

Cool. So now all that's left to do is

test out the agent. Of course, we first

have to give it a brain. So I'm just

going to go grab my open router API key

and we're just going to stick with

GBT4.1 Mini for now. All right. So now

the agent's basically set up. All that

will be left to do is test it, tweak the

system prompt, and keep playing around a

little bit. So that's exactly what we're

going to do. First, I'm going to drop in

this query asking about the Tesla

document. How many vehicles did Tesla

deliver in Q2 2025? And how did that

compare to the same quarter in 2024? So

once again, what's happening is it's

hitting our Pine Cone Assistant and then

it's going to come back with some

answers for us. So, you can see it said

that Tesla delivered a total of 384,000

vehicles. And then it gives us a

breakdown of the different models that

it actually delivered. And it also said

that it saw a 13% decrease in total

deliveries year-over-year. I just went

ahead and fact checked all this. That is

correct. Now, keep in mind there's no

system prompt in this AI agent, which is

why we're not getting any source

information like what PDF it pulled from

and what page numbers it found it from.

So, what I'm going to do real quick is

give it a quick system prompt right

here, which is what defines the behavior

of our AI agent. So, I'm going to open

this up full screen, and I'm just going

to paste in this prompt that I was using

earlier. It's really, really simple. I

said, "You are an AI agent specialized

in analyzing earnings reports data. Use

your Pine Cone tool to search through

earnings reports from Tesla, Nike, and

Nvidia." When answering the user's

question, always site your sources as

far as what document you got it from,

what page it was from, what section, and

an exact textbased quote from the

original source. because when we're

doing rag and vector search, it's really

important that we're seeing where our

agent got the information from so that

we can trust it. So, this is just a

great example of defining the agents

behavior. Because what's interesting is

you'll notice when it searched the Pine

Cone tool, we actually were getting that

information. If I was to scroll down

over here, you could see that what we

were getting back from Pine Cone is the

exact pages that it pulled this

information from, as well as the name of

the PDF it found it from. the AI agent

just didn't know to actually tell us

that information. Real quick, guys,

excuse the lighting. I'm sitting here at

nighttime editing this video and I

realized that I didn't show what I meant

by the agent filling in the query down

here by itself. So, if you remember this

question we asked was how many vehicles

did Tesla deliver in Q2 2025 and how did

that compare to the same quarter in

2024. So, what happened here is we can

see the agent called the Pine Cone tool

twice. We can see there were two items.

And if we click in, we can see the two

different searches that it made to our

Pine Cone Assistant. Up in the top left,

if we go to run one out of two, it

searched Pine Cone Assistant for Tesla

vehicle delivers Q2 2025. And then it

knew it had to compare that to 2024. So

then it searched it again and it said

Tesla vehicle deliveries Q2 2024. And so

this search query is the variable that

we have down here because the agent

basically decides what do I want this

search query to be and how many times do

I need to make a search query. And

that's why on this right hand side, we

also have two separate runs and two

separate results. So for all the future

examples, just keep that in mind. The

agent decides the search query and

that's how it's able to get the right

information every time. Again, excuse

the lighting. I'm working on getting a

professional light back here. So if I do

record in the dark, I'm not terrifying

looking like this. Let's get back to the

video. So just to show you guys that

that prompt worked, I'm just going to

repost this exact same message. And what

we're going to see is it gives us the

correct answer once again, but this time

it's going to have a PDF name and pages.

So there you go. We got the same answer

and this time we got our source of Tesla

Q2 and we have our page numbers down

here. But what you may notice now is

that we're not getting the textbased

quote from the PDF. And the reason why

we're not is because what happens in

this Pine Cone node, and I'm not going

to try to get too technical here, but

the Pine Cone assistant on that server

searches through the knowledge base and

then it gives us basically like a short

summary with the correct answer. This

right here is not an exact textbased

quote, but this is what our NAN AI agent

uses to give us an answer. And just to

highlight that point, I'm going to show

you guys a better example of that. I'm

asking, what were Nike's revenues for Q4

fiscal year 2025? And how did they

compare to last year? And what you're

going to see happen is it gives us the

correct answer, but now it gives us the

exact quote, which is wrong. So if I was

to copy this quote and go into the Nike

doc, which was this one, and if I paste

this in, we don't get any hits. We get

zero matches. And that's because once

again, what happens is the Pine Cone

content that comes back is a summary.

This is the Pine Cone Assistant answer

that it made based on the exact text. So

the way that we fix this is we have to

understand the type of request we're

making to the Pine Cone Assistant. And

in order to do that, we have to look at

the API documentation, which I will drag

over right here. So this is basically

the documentation that tells us how to

use the assistant over API. We can see

things like streaming response,

extracting the response content,

choosing a model, all this kind of

stuff. But what I'm interested in is

this bottom section that says include

citation highlights in the response. So,

all we have to do to get an exact site

is add this thing down here that says

include highlights. This is basically

just a little lever that we can pull to

change how the assistant works. So, keep

in mind how this quote didn't work. I'm

going to go into the HTTP request and

all I'm going to do is add this little

section include highlights equals true.

And now I'm going to save this and run

the exact same question. So, we're

running the exact same question. the

answer is going to come back correct

again because the Pine cone assistant is

going to be pretty good at getting it

correct. But now we're going to see an

exact quote come back as well. So right

here we get this exact quote. I'm going

to copy this, go back into the Nike doc,

and now if I paste this in, we can see

it's pulling the exact quote now. And

the reason it was able to do that is

because in Pine Cone, once again, this

section called content is not the exact

quote. But if I scroll all the way down,

you can see we have a new section that

got added called content. And this is

basically what's going to pull exactly

what was found in the document. So this

is just a really great example of not

only being able to understand what's

happening, but being able to understand

that if I want to change the behavior, I

can go read through API documentation.

And this is how I can also change things

like the sampling temperature or I can

change the model because one thing that

you will notice is that if you're in

your assistant playground right here and

you change the chat model to interact

with, you can see I switched the model.

If I hover over this, it says select the

model for the assistant to use in this

conversation to persist this choice in

your application. Remember, you have to

set your model in your API calls. So,

going back to NADN in the HTTP request

down here, we have a model. And so, if I

wanted to change what model we're using

to interact on the Pine Cone Assistant

side, we would have to change that right

here. All right. So, sorry if I was

getting a little technical there. I just

think it's really important that you

guys understand how this is actually

working. If you're looking to understand

this stuff a little bit better, I have

an API video that you can watch. I'll

tag it right up here. Or you can also

check out my plus community. The link

for that's down in the description.

That's where you'll receive more

structured guidance and a community of

people that are building every day.

Anyways, real quick now, I want to show

you guys the difference because I'm sure

you're wondering, why would I not just

use a Pine Cone vector store or a

Superbase vector store? Why is this

quicker and easier and maybe even

better? So, the reason why I'm a big fan

of this right now is because what's

going on here is Pine Cone on the back

end is handling all of the indexing,

embedding, chunking, it's handling all

of that in order to make your job easier

so that you don't have to manage all of

that to get your accurate sources back.

Cuz like I said, if you're taking the

typical vector approach, there's a lot

more work that goes in up front. You

can't just drop in a file and be good to

go. You have to set up a pipeline to

have like metadata filtering and maybe

even different types of splitting and

chunking. I'm not going to dive super

deep into this right now because I don't

want the video to go too long, but I

want to show you guys a real quick

comparison. So, all three of these

agents have the exact same prompt and

the exact same documents except for

they're just being stored somewhere

differently. So, I'm going to drop in

this question right now to our Pine Cone

assistant that asks, "What was Tesla's

operating margin in Q2 2025?" So, you'll

see it comes back with the operating

margin was 4.1%, which is correct. You

can see it gives us the document. It

gives us the page numbers as well as an

exact quote. And also keep in mind that

we got this answer back right here. It

says 1,277

tokens. So keep that number in mind. Now

if we move over to this middle agent

that's using a Pine Cone Vector store

rather than the assistant, we're going

to ask the exact same question. And

remember, we're looking for the answer

4.1%.

So what we get back is I don't really

know what the answer is. And you can see

that it took almost 30,000 tokens. So we

got an incorrect answer. And it cost us

like 15 times more. More than 15 times

more. Yeah, that was horrible math, but

you get the point. And because we took a

typical vector chunk approach where we

were in charge of all of the

pre-processing of those chunks, I'm

assuming it's going to do the exact same

thing when we try this with our

Superbase vector store. So, you can see

here pretty much the same answer. It

wasn't able to find the exact figure. We

were looking for 4.1%. And this one only

took 5,000 tokens, which is still like

three times more expensive than the Pine

Cone Assistant. So, that's all I'm going

to talk about today. I'm not saying that

the Pine Cone Assistant is always the

best option because you also have that

running cost of 5 cents every hour. But

the point I'm trying to make is if you

want to spin something up quickly and

play around with how it's working, this

Pine Cone Assistant is a game changer,

especially if you're a beginner and you

want to experiment with stuff like rag

agents. And also wanted to give a quick

shout out to the legend Mark Cashef.

He's the one that showed me these Pine

Cone Assistants and like I said, I think

they're super cool. So hopefully now you

understand how you can test out the Pine

Cone Assistant for yourself. And if you

want to download this template so you

can just kind of play around with the

differences, then you can get it for

free by joining my free school

community. And if you like seeing stuff

like this and you want some more

structured learning, then definitely

check out my plus community. The link

for that is also down in the

description. We've got a great community

of over 200 members. Everyone every day

is building and earning with NADN. And

we have a full classroom section with

three full courses. Agent Zero is the

foundations for beginners. 10 hours to

10 seconds dives into NAND. And then our

new oneperson AI agency course is

available for our annual members and it

goes over how to lay the foundation to

build an AI automation business. So I'd

love to see you guys in these

communities. But that's going to do it

for the video. Hope you guys enjoyed. If

you learned something new, please give

it a like. It definitely helps me out a

ton. And as always, I appreciate you

guys making it to the end of the video.

I'll see you all in the next one. Thanks

guys.