How to Create AI Sales Agent using Open AI Assistant API

knowing how to create an insanely powerful gbt right now puts you in the top 1% of those who at will a fully leverage opportunity presented by the new release of the open AI API assistant you're not yet familar to technology you're not alone I was in the same position when I first learned about it however by the end of this video you not only understand how to create incredibly powerful gbts while you also learn how to develop your own sales assistant GPT this will give you an idea on how to cre

ate specialized AI tool designed to L substantial value to your company business or clients anyway in case you're still confus meet Emma our AI personal sales assistant in just 20 minutes I'll show you how to create Emma from scratch using pine and open AI assistant API she can analyze your company St dat and calls find information on new leads by searching the web Prov answers to certain questions by Levering the information store your CRM and even make recommendations to close more deals now b

efore I go more to Emma let's go over a fast explanation on what the open AI assistant in its API is the OPI assist is simply a collection of tools and functionalities that developers can use to build their own AI assistance an API gives you the ability to L three powerful tools and its capabilities the first tool being retrieval is simply being a tool that acts as an open AI paretic search engine allowing developers to easily create assistance that Cur information from uploaded documents or ext

ernal knowledge bases for answering domain specific questions about the hassle of managing database the second tool being the code interpreter which grants AI assistance the ability to generate and execute python code directly from the selfhosted stand boox this unique capability empowers assistance to generate responses based on the outputs of the executed code the last tool being function calling which pretty much allows developers to expand their applications and have it complete specialized

tasks beyond the core assist capabilities by utilizing these custom functions AKA specialized classes or models that are created by the devs themselves developers can use these functions to find unique processes or calculations or actions like for example they can create functions for conflex data manipulations or to integrate external service to distract certain information and so on then you can use this an API to seemlessly integrate these custom functions into the conversation flow and ex th

ese Define action based on user input now that we've broken down the basic components let's go into a quick little demo showing what Emma can do and what the final results of what we'll build will look like before we going into the full coding demonstration yes we have the emap but install in Microsoft teams I'm going to go over all the functionality that Emma provides that I'll go over in the video the first functionality go over is the retrieval functionality we use this functionality to have

Emma answer questions and make analysis on all the sales calls a company has which recorded in the PDF documentation so we can ask something like based on all the sales based on all the lead SI special customers you talk to in the sales call rank them one to four which would be the best leads AKA most likely buyer Services pay the most amount of money or will seem like they'll continue to work us work us in the future and give us a reasoning and also rate those leads from one 10 score so I have

that question right here copy and pce these are the exact things I mentioned [Music] enter and now you can see it gives us response after reviewing all the available data from the sales call here is analysis and ranking of the leads which each being rated on lead score from 1 10 Kim lead score of 9 out of 10 Kim is highly gag eager to move forward the proposal and shows R this review and proceed with the project they are Sid side with the presented price and payment structure and are ready to in

itiate the project and gain strong byy 10 potential for a longterm partnership car Le score of 810 expressly interest in customization integration valid teal insights offered for the project Patrick Le score of 9 7 to 10 s 10 so you can pause this and kind of read through all the breakdowns that the chat but gives us now the next functionality I'm supposed to go over deals with a function called calling functionality so we have two cases where we use this functionality first case is where we hav

e ml information from a database we have created an air table based on what the lead user inures about so this information will have stuff like when is the lead sex meeting meeting notes how the lead was found and so on so we might ask a follow-up question like we'll kind of ask a question like this so we can get details on what more information our lead so when our next meeting Carlo is what we should prepare with Carlo so this will let us um quickly CER information and find information on on o

n our leads and pick Carlos since it was one of the guess the highest rated this leads that we have the previous question so ENT so you response next me car scheduled for November 25th 9:00 Easter time for this meeting Loy the salesperson sign will focus on getting the pricing and discussing the security access to their chat Bo incling how we should ensure the is secure car company has 800,000 employees we found Carl at the AI conference if you need to get in touch with him his email is car cons

ole gives you all information our leads now the second functionality we have or the second case where we use the um function call functionality we ask Emma to pull information from the Internet or web unless we don't have our database so for example let's say I met a dude called Li Ley in a conference and a discussion on business opportunities with him after I want to find some information on him see how good lead is I can ask a a chat a question like this during AI conference attended last week

I had an opportunity to meet up with a gentleman called Liam Aly mentioned he runs the AI automation agency and express interest in potential partnership can you please conduct some research for me find out more about this his company especially I want to know the name of this company it's official website and a brief overview of the service product he offers so get that see in a sense says I found information about Liam and his company M AI Liam Ali is the founder of Mer AI which is a full cyc

le AI development automation primary they're special in AI development aiming to take business to the next level incing AI into their systems their official website is more say Ai and for the com offers um they offer com AI service that help STS Enterprise build CL products that use intelligence to optimize effici andary Grove it says this they guide through the whole process process it tells you when it's founded where it's headquarter is where they specialize in and so on so gives you kind of

overview of the Liam Oley and his company so you get more you have more information or more idea of who he is and how good it of he a lead and then the final question is the final functionality we have is is a code functionality this allows em to create items like graph based on the numerical data we have stored in Excel she by generate own code so you want a visual ofan we ask them in like crit bar graph comparing potential deal amount for each C each lead so and now you can see it gave you a b

ar graph comparing the has potential de amount deal amount each lead is will bring in so pretty much how much money will get off from each lead you see David is around 7,000 Patrick is around 30,000 so you can see all the deal amount each lead could um could bring in so now since I showed you a short demo of how Emma Works let's go into showing you how I built this thing now the first assistant that I'm going to go over is what we're going to use to answer questions and inquiries of our madeup c

ompany sales calls i r information docs that have all the sales information written into them let's have a quick look at what's in that sales doc you see that it has seven sales conversations and all the sales conversations are sent around a Salesman trying to sell a chap to a lead each sales conversation having its unique elements when it comes to price objections timelines comment objections feature requests and so on now if we go to the code you see that in that first line we have their API k

ey which we have decided in our environment file so I can don't nobody can see it so I don't share it with you and then we have in the next part code we have the client equals open. client code this code is pretty much what we're using to create a pip object called client that interacts with open as API object simply being a container that holds the data and instructions on what object is and what object can do for example you might have an object called login that has classes that allow you to

set a username and password and change that password through ACC em verication and if you look at the back at the open eye client expression the Open Eye part simply refers to the library using the clent client refers to class in that library now if we go down we have their create assistant function and pretty much it's we using the create assistant since open AI builds you per assistant you create in the first couple lines we'll have an instatement that checks if a text file C assistant ID exis

ts if it does we leverage capab of already pre-existing um assistant by reading using the assist ID that's written assistant text file and if it doesn't exist we'll use the client. beta. assistant create method the call methods in the client class in open a library that practically act as a wrap around one of open A's endpoints that will CIT assistant with the attribute described in the client. beta method and if we actually go to the open ey documentation you can see the open ey end point the c

ode need to utilize the open ey end point and all the parameters you can use or send to the request and if we go look back our code you see the parameters we have we have the name sales um sales call Knowledge bot we have an instruction which is the our prompt that we're going to feed it and we cite that over here your sales chap by responsible for re information and so on we have our tool which is the retrieval tool which is going to give us the functionality to extract data from our docs and t

hen we gave it a file from the file create and attach the assistant using the client. files create method and then I'm going to run the code soon but before I run the code I want you to look at this left hand side you see that there you can see all the files over there I also want you to look at my open Assistant tab you see right now we don't have assistant at all now I'm going to run it okay and the first line you can see that it's create we're creating assistant since we don't have assistant

right now um it's going to it's not of using a prese assist it's going to create assistant and you see on the left hand side it creates a text file used to that and this is the file we use to pull the pull the um the guest the assistant ID from anytime anytime we want to use the pre assistant assistant instead of Crea a new assistant and you go back to my opena account and refresh you can see a newly created assistant that wasn't there before and you even see the promp over here in the assistant

ID now let's go back to the next prer code so next step is create a thread we do this by using the thread. create method thre simply being the conversation session between the assistant and the user AK the context window that has all the message stored once we create thread we'll create a user message user message being what are the most common objections we face in our sales call and then we'll add that message and thr thread we'll also have a user Ro attached to it to signify signify that mes

sage sent by a human user and then next we use the client. bath. thread. runr to invoke the thread in other words we're pretty much sending the request open as API Tri assistant to process the current threats of conversation history which right now has already used message in and then generate a response to that used message and while we're waiting we'll use run that retrieve to check on the progress or the status of that of that run or the in invocation of that assistant and you can see in our

console we have a bunch of waiting for assistance waiting for assistance we complete statements printed out and when that thread is finally completed it will capture response but until then it'll be kind of in the Pro process process status and when it does complete we'll we have we'll list all the mesages using thread. messes list and then we'll eventually Loop through all the messages and get the latest message in thread which will contain the assistant response and if we need if you need any

more insight on what threads or runs are this is all in the open open eyes documentation now going back to our code you can see that we have some added logic to create a message object into Json format and right into a file name message Json um the inspect or look at and we go back to our print statement oh we're cool we also have this for a run run we have we for a run object we write into adjacent file too then look at look out later also and then we go back to our print statement you see that

we have message. R capitalize capitalize which is pretty much code simply capitalize the first level of the string and then the role is going to print out where the message come from like who's sending message and you see here it says assistant send the message so that's the person that's send the message assistant and then we also have message. content that's kind of how we're going to Loop through and find the latest message we have message out content zero so we're going to get the latest me

ssage in there and text about is where the exact message is so if we go and actually look at message Json we can see value and we actually see the actual response Sy gave us and we see the value is inside the text bracket which is inside the content bracket and if you look at the run run St um Json file you see stuff like the Run ID the status the two calls you use the third ID and so on um yeah and then going we go back to the console we actually see the system response we see that it's answer

to what the most common injections we face or our sales call is stuff like higher Point payment flexible payment structure concern about return of investment long support and maintenance cost and so on now let's go over how you can actually integrate the system functionality into a chatbot so if I go over the rep here you can see the sales calls. pile and you see a lot of code we already went over before just like before we have the client object and we have the credit assistant function that he

lps us create assistant or user pre assisting assistant when leverage and ID we also have the prompt P file which will have her prompt or assistant instructions and then we have the secrets tab which will have her API ke in now if we go to the main.py file this is where we'll set up our API endpoints that we call when we want to start up our system or grabit responses I'll go over more why after we go over the code you can see in the first couple lines we're trying to create a flash application

then create a client object that we use to interact with open API and we finally will call that create assistant function and load the that's been created in the create assistant function in the sales e file and if we go back you can see that we're returning that assistant ID so that's the assistant ID that the main. file will use then we created two two endpoints one is a start end point which will create a thread when we hit it with request and the other end point being the chat end point whic

h will pull the thread and the message from that request the message being the message that user sends to the assistant from that from the request that hits endpoint then we'll use that information to run the thread and return the system response so now let's run the code now after we run it you can see the web view Tab and after it loads after a while you see it openads a web page called not found and here we're going to go here and we're going to copy URL and you see the little show you the th

at we copied here's Ur that we copied now this will go to the next part which will explain the whole reason why we move the Cod to reing Great end points to hit so I'm going open up voice flow voice flow is a software that helps us build a AI Ag and chat Bo through low code and FL designers right away you can see that we have a box called Crea a thread which has a Euro we saw before plus the Endo that we mentioned in the start SL start so when I click it you can see that we're saying get quest t

o it you see the full you see that SL start end point that we mentioned in our code and as I showed before we're trying to extract the thread idea vals values from that response so that start end point it's going to create a third ID when it creates a third ID you want toct that thir ID from from that start endpoint so kind of show you what that endpoint will send I'm going to hit it in post man real quick give you example so send and see when I send that um get request you can see down here bel

ow you see that thir thir ID value that's the value that we're going to be C capturing when we hit the start end points then if we succeed we'll send two messages one is high I'm your company interal sales Bo and we'll send we'll send a response to that and we will capture that reply that reply that user sends so that prob probably be the question that the user wants to ask chatot and it set equal to the last utterance value this pretty much will be the user message that we want to give to assis

tant so it can generate response to it then those two values the third ID and the guess response that we capture that guess or AK the last utterance are going to be values that we send to the chat endpoint as a request and as before the chat endpoint is specifically used to run threads and send the response assant gives and let me show you example by sending that that send that request you're to get the chat endpoint through a crow command so let me copy and paste this thread [Music] and now you

can see that sender response so that's gives you example of what kind kind of the body that would look like you see the message that we sent you see the thread that we sent as the body and then you can see the response that it sends back to us one of the leads that we have sales call with is somebody called named Patrick now going back once the assistant C response the next pause called s response will first clean up the response and make sure it's sent in the good format for example we see the

response Crea in vs code before you see it has a bunch of brackets and stuff inside it that code will kind of clean that up remove all the brackets from it so when we actually run it it will just be clean format and clean text no no extra stuff that we don't need extra syntax that we don't need the next step in there is pretty much going to print out a show the message in the chat but or chat window so after it generates response it we can see we can visually see the response that it gives the

user and after that will'll loop back to the capture user reply so that you can repeatedly asks that chat Bo questions and then we also have I guess kind of a failover so if you see um these failed chap chap um guess box steps anytime that there's a issue to the back end it will route to these failure boxes and say something like fail to gener new response um ask your question ask your question again or contact support so you can customize that and have guess responses in case that in the back e

nd you're chat fails or something and now let's have a quick demo let's run this run test and a ask tap up what are the what are most objections pleas during sales calls and can see sent a similar response that I sent before cost concerns return to investment flexibility and payments log for enhancements that's kind of similar to what it said before in Venus code flexible payment structures High fund payments concern about return of investment Nance cost and so on so that's kind of the example t

he first assistant that I created the first functionality that I showed you in the user before in the demo before now that we've seen the basis of how systems work let's go for some of the other functionality that sys have which I mentioned earlier video let me show you my other project in voice flow which I use to build chat Bots with not only retrieval functionality but the function calling and the code interpreter functionality also now the biggest difference here is that it has a classificat

ion block that will route you to a chat point of a float to another component component for so write you to another component and all this component does is it uses the AI Tas which will categorize the user messages into four different categories air table if the user asks for a specific question about the clients in their database sales if the user asks question about recorded sales calls data of user message ACC for Data Insights a graph Based on data collected and recorded by the company and

info if user asks questions about information about potential lead or partner is not recording database yet and based on where the AI categories the user question route the response specific endpoint using IF conditional statements each endpoint h a assistant that specialize to answer the category user messages in so you see all these um blocks and each one uses its own assistant with the last respon being its own block that we'll use to display the response in chat UI and loop the flow back to

the original point or back to the user Capt block allowing the chat flow to repeat the process after you asked another question now let's quickly go for the code blocks and all the different assistants we have let's go over to the code for a table functionality like described in the classification block this is used for answering questions on when is the next meeting with the client what notes are to know for the meeting what is the size of client's company and so on this information pulled from

a database in air table low code software now let's quickly take take a look at what we have in air table if you can see in our database we have stuff like lead name lead email next meeting meeting notes company size and where lead was found and now if we look at our code you see we have the first thing you see is that we have a function called get records by lead name all this doing is Tak whatever value the lead name is equal to which is our argument and filtering through the row your lead na

me occupies using error table API so it sends all the information need to answer the user message to the assistant so how does the assistant know to run this function how does know what value to pass to the function first if you look at my code you can see that I have a variable called two list this is where I will Define my function for my open Assistant now I put here is the function name we're using description of the function which will is used to help the assistant determine when to run the

function and the property section where we Define the argument that will pass to the function this is what we're going to use to take the string AK the user input we sent to assistant pull the lead name from the user input and turn to a Json format so we can pass that value into our function and you see that we have twool list defined in our guess in our assistant Rec creating and if we go to the main.py and we go to our run status for function call there's a period where function status will b

e equal to required action this is state is where the system will determine the name of the function and argument to pass and while it's in this required sayate it will check the function name which tool list and then if the name is set to the right function it will get the AR argument that we tool list which we're converting from Json to dictionary then we'll take the function we set up in function pi and call that with the same argument and like I mentioned in the description the argument will

be the name of the lead we're trying to get information from after the function runs we're run we run the thread using the Smit to Output method with the output value as one of the parameters so the Sant can use the value from returns to help answer the user questions I write the Run s into the file and print the argument output to give you a deeper look at what's happening so let's go to Voice flow and test this first let's run this first let start up our server Let's test this let's functiona

lity this air functionality so let's go here let's go to start let's go to run let's go to run [Music] test and let's ask what should I prepare for in my next meeting with [Music] David and this what we get your next meeting with David schedle you should prepare for a basic demo of one of the Chaplet to previously built sptings be aware that company so it it lits all the data that we we we have on David make sure to have the demo tell it to company siid and PR to answering inqueries he might hav

e ration skill be in support so and then if we go back first go back and look at our run. tools jon5 and we see that we can have we have a required action bracket we have the submit two Clause bracket and then we have a function bracket and this is where we retrieve our argument from and the function name from and then to the outputs we have we have the the output that we retrieved when we sent um API request to a table so this is the response a table give us back and we also have the argument s

o we if I copy copy this and take this copy and then go test this in the UI also next um what do I need to pre for in our next meeting with David what do I need to prepare for in our next meeting with David run that it access the SM output so it doesn't have access to the function that our endpoint had access to so it's going to ask us to submit an output and it's going to be in that required action State until we give it output so I'm going to give the same output that air table gave us um when

we're r code so copy and paste submit continues to run and we'll just wait it's in progress we'll wait until it gives the output and it gives us the output and we see that outut gives us for the accompany r d there are a few things you need to repair a little different but pretty similar overall so you say be ready presentent basic de the chatbot the size company read interaction David particularly so output but the same overall meaning now I have another function that allows people to find inf

ormation on leads not in their database by searching their name Google and finding their LinkedIn profile and scraping data from their profile and their company profile so they answer any speciic question on these leads so you look at my code here I have a j Json readable summary function have a search LinkedIn profile function I have a get LinkedIn profile company data function and the main one the main one is the LinkedIn insance company dat um function that's what my function call is going to

be calling and then the rest of the functions are pretty much kind of sub functions in that function that be called so I won't go into too much more detail on this but I'll show you example to kind of show you what it can do so our test subject is going to be this guy he's the COO of voice floor the software we're using so I'm copy and paste this [Music] name now I'm going to run run test and before I do that after number I need to run this now I say and now he gave us a huge huge huge response

so it says BR re is a primal figure in commcial F of AI he's currently SE of voice FL that specializ in Democrats creation digal assistance under his leadership company blah blah blah experience as a blah blah blah undertaking te including raising BL so you also learn more their website.com we looking for D content so it gives his direct contact and his website and talks more about his company size by so it gives you a bunch of information track record so so that's that's just example of what m

y my LinkedIn assistant can do the final assistant we are using is going to leverage Cod interpretor to create graphs based on sale driven data we feed the assistant we feed the assistant the s data through an Excel file take a quick letter Excel file you see that we have have the deal owners company status contract deal owner future client clients want implemented and the most important part statistical part is deal amount expenses property deal for referral potential and so on now we take a lo

ok at our code that we have some extra code there so assistance does allow you to um read Excel files or taking Excel files like kind like how it can read and take in sales P PDF files but but instead I'm going to use um code to turn the Excel sheet into more readable format for the assist instead so pretty much what this code is doing here is that it's reading processing from the Excel file converting it from a more usable and readable Json format while also clean and remapping any unnamed colu

mns you'll write all that data to a text file called output then we'll add the retrieval function as an option along with code interpreter so the assistant can read that file and we look at our made. pile [Music] file you see that it checks if an image is created if an image is created for that file D and read that file D then it'll use that to take the data from image read it and write into a new file that will say as a PNG file and once it gets that file it will take that file and host it in a

nother API endpoint right here and what what will do is is that we're going to take that imel from that um from our chat endpoint so you look at you look at our um voice flow you see the what you see what it takes you see it takes the image URL and then if it gets the image URL it's going to go to this box here and then what what it this what allows you to do is if you give it image URL it'll visually show that image in that in the chat UI now we're going to run a test and see [Music] so you kin

d of see it you see the deal amount and you can see if you go to the Excel sheet it matches up so 3,000 6,000 2000 3,000 6000 2,000 so it matches up and if you also look at here another cool thing look here we can also see it's making I guess get requests to open AI endpoint that we brought up earlier before so it just shows you that code is making a call to the endpoint and overview open AI assistant API functionality if you thought this video brought you value drop a like And subscribe and if

this video still left you a lot of answer questions you can join my telegram group down below it's the best place to reach me and ask me any questions if you still need help on this topic also let me know if you're still interested in a video to show you how to integrate your chatbot into applications like Microsoft team and slack like I showed you in the demo at the beginning of this video