Build AI chatbot with custom knowledge base using OpenAI API and GPT Index

hi this is Serena Nick from Tech
foundation and in this video I'm going
to show you how to create your own AI
that can answer the questions based on
your custom knowledge base so why do you
have to create custom AI if you already
have charge repeating well you can ask
chargepc something about the general
knowledge on the internet like for
example what is ux design
or what is product management and charge
GPT will give you a decent answer based
on something it has seen before on the
internet and there is no way that
charger PC can answer questions about
the specific knowledge for example about
my specific user research because there
is no way it can access this data and
that's why I've decided to create my own
AI
it's pretty easy just few lines of code
we are going to use open AI API so still
we are going to use GPT 3 model and we
just upgraded in that way that it knows
about our custom context how are we
going to let this model know about our
custom context well we just give it in
the prompt you might ask why cannot we
then use just check GPC for that because
we can give the context in the prompt
also in charge repeating the problem is
that usually user research databases are
quite big and if you want to implement
any other custom knowledge database it's
usually big so you cannot put everything
just in one prompt and that's why we
have to create a little bit more
flexible way to give the context in the
prompt and we are going to do this by
breaking our knowledge database on small
chunks and indexing these small chunks
so when we query our AI it will be able
to find the relevant context in all our
database and based on only this relevant
context give us the answer so we are not
going to give all our knowledge in one
prompt in one query but we are going to
find the relevant knowledge and give
only these pieces of information to our
Ai and that's how it works pretty easy
so instead of going and giving all the
content effects we are we need to find a
way to find a relevant context and
that's why we need an external library
to do this and that's why we cannot do
it manually but we have to do it in code
indexing that data is not an easy task
but luckily for us there is already an
open source library that we can use so
we don't have to go into the technical
details how to do that but we just need
to figure out how to use this Library
the library is called GPT index and it
is also using gpt3 API in order to break
your data and make an index out of it
so we just need to figure out how to use
it and it's pretty simple I'm going to
show you now how I've assembled a small
project in Google notebook it is called
collab environment so you can also
follow this tutorial and try it on your
own so the first thing that we need to
do is to upload our data that we want to
Index right and if you want to use your
custom knowledge database then you need
to create a folder here named
context data I've called it context data
and then slash data but you can call it
whatever you want and then just change
the name here so you create a folder
with all the files from your knowledge
database and it will be there
uh I've prepared already a folder for
you that we are going to use in this
tutorial and these are uh just fake
interviews so I cannot share with you
the real interviews because it's a
sensitive information and it's private
information I cannot share it so I've
decided to create fake interviews also
with the chargerp team I've actually
asked jgpt to create an interview script
for me about cooking habits and the use
of domestic appliances it created not
very good interview script but for the
sake of experiment I think that's fine
so after that I've asked the GPT to give
me an interview script for these
questions
and I gave it just example of what an
interviewee is so I created just a small
use case like business professional
living alone who tries to control on his
diet and for the this small use case
2gbt was able to give me an interview
script that later I was using as the
custom knowledge base so in this example
we are going to use a fake user research
data unfortunately it also means that we
cannot be we cannot get any really
insightful information out of it because
still tgpt provides you
uh quite blank data so it's more like a
boilerplate than um real insights but
unfortunately I cannot share with you
the real information so that's what we
have so I've created a repository for
you that you can copy into your project
and there is already a piece of code for
that we are going just to run it and
what it does it clones our git
repository that is here
and add the information here so now it's
not there I just need to update it yes
reload please
um yeah I don't know why it doesn't show
it from the beginning but anyway now we
have our context data and we have a data
folder inside where we have all our
interviews
you can check this interviews
interviewer answers and interviewee
answers and interviewer questions
so for our examples we are going to use
just for interviews of course in the
real scenario you're going to have much
more
then we need to install the dependencies
so we are going to use the GPT index
library and this library
in turn uses the line Channel Library so
we are going to use actually two
libraries that we need to install now
we run the code it is installing all
right we install down dependencies and
now we are going to create our functions
that we are going to use and it's only a
few lines of code just two functions for
us so first we need to import everything
that we need from DPT index and line
chain and if you want to learn more
about it you can just go to the
documentation and it is written here
what you need or you just copy it's the
same and then we are going to write just
two functions first one is construct
index and these functions this function
take the directory path in our case it
will be context data select data and it
creates an index out of it so instead of
having just the row data like we have
here it will create a Json index here
that will index all the information that
it finds here and that's all the first
thing that we are doing here is just
setting parameters for the large
language model that we are going to use
so it's just the parameters from open AI
API and you can also find the parameters
here so maximum length the temperature
that is responsible to how the answer of
AI is different from one another so if
you ask the same question will it be
always the same transfer and then the
temperature will be zero or will it give
for um
different answers and then the
temperature will be higher so you can
read more about the settings for lash
language model in the open AI API
documentation and you can then play in
the playground just asking questions to
I for example
what is ux design
and
then
checking what answer it will get
and that's what we are doing here we set
the parameters like maximum input size
we set that temperature the model name
because there are different models and
you can learn more in the open AI
documentation and so on so after we set
all the parameters we are going to use
the functions that are defined in the
GPT index Library so we just have to
Define our path
that we are going to use to create our
index and that's basically all after
defining the path and loading the data
we are just creating the index and
saving it to the local folder and that's
basically all what we do just loading
the data
running the index and then saving the
index in Json format here and then we
have the second function and what it
does is only querying our AI model with
the question and context information so
we first identify the index that we are
going to use and then just query
the AI with the user input so here we
ask a user to input something what we uh
what we want to ask our Ai and then we
just query our AI with the context
information and displaying everything in
markdown format so pretty easy we just
have to run this code and for now you
won't see anything because we just
we just let this notebook know that
these are our two functions we didn't
run them yet then we need to set our
open API keys so uh to in order to use
openai API we need a key and you can get
it by signing up on their website so
after you sign up you'll have here API
keys so if I go here view API Keys it
will be there if you don't have anything
you can create a new secret API it will
create an API for you that you can copy
I copy this one press OK and
just go here and run the code it will
ask me to put my open AI key here so
that's what I will do and you will do as
well I won't show you my key because you
need to use your own and the reason for
this is because using openai API is not
free but it's relatively cheap so for
example for all my experiments I've
spent only 14 cents and you'll get 18
dollars for free so they call it credit
but whatever now it's time to construct
our index so remember we defined the
function here construct index and now we
just need to run it with the path of the
data that we want to index so in our
case it's context data data so let's run
it all right so we have created our
index and it will appear here I don't
know why this Co-op environment doesn't
show it immediately but if you reload it
will show it so let's just do it
after we reload we can see that
index.json is now here it wasn't there
before so it all worked and now what we
need to do is just start asking
questions so quite exciting let's try
uh remember our data was about
interviews about cooking and using
domestic appliances and the information
from these interviews is what our AI
will use to answer our questions for
example I can ask what people like about
cooking at home because the interviews
were about cooking at home and we were
asking this question to our interviews
let's try to see what it will answer
so it transfers that people like cooking
at home because it allows them to relax
and unwind after a long day it's Health
it healthier and more affordable than
its and out and so on so all this
information was mentioned in our
interviews unfortunately this time it
didn't break it down but we actually can
ask our AI also to break it down
into clusters and then it will be easier
for us or for example we can even ask to
brainstorm something for example
marketing campaign
based on our interviews and already use
AI to brainstorm something about our
project and it's very exciting because
now we can actually use AI to brainstorm
with us to brainstorm about our project
based on our information so it won't use
just the general information from the
internet but it will use the information
relevant to us in order to brainstorm
more ideas after you played enough with
your AI you can just stop this
it will show you the error but this is
because we just stopped this code from
happening so it's not there just the
keyboard interrupt so don't be worried
about it and that's it just few lines of
code actually the only thing that we do
this is the install independencies let's
close it
uh it's just this one so we just Define
the functions and that's all and then we
just run the functions I hope this video
was helpful for you and now you are able
to create your own AI that will use your
custom knowledge base don't forget to
like this video And subscribe to the
channel because here you could find a
lot of relevant and free information for
designers bye
foreign