Build India's First ChatGPT like App for Politics: BJP-GPT

Key Takeaways

hello everyone welcome to AI anytime Channel so in today's video we are going to work on a very interesting project okay we are going to you know build a gpt3 powered application or a chat bot with memory and context and we have named it bjpgpt so if you are not aware about uh BJP it is it's an acronym of parties and the party so it is a political party you know in India and our prime minister Narendra Modi is you know a member of this party and we are going to build this uh chat bot okay uh powered by gpt3 that will have this you know memory and context and we will build it from uh let's say scratch but we will write the code we will create this memory we'll see how we can save this history with the context and how we do the prompt uh with a set of prompts and how we use some of the NLP techniques when we are you know considering this history and context we look at at cosine similarity and all those techniques of natural language processing to build this bot so I'll first show you a demo of this okay uh now see this is an interface uh a very minimal interface that you can see on your screen it says bjpgpt okay which is an acronym as I said party okay uh and before we go into the demo I would like to you know first show you some of the tools that has been already developed and you know I you know I have drawn some inspiration from those tools now if you see Bloomberg has released this something called Bloomberg GPT and they have they might have trained a custom model it's a 50 billion parameters llm model that they have built from scratch you know on their Finance data they have abundance of Finance data that they can build these uh models with larger name models from scratch and you know get you know comparatively similar accuracy like we see in charge GPT or GPT X power models and uh similarly we have something called doctor GPT you know and you know Bloomberg GPT says they have trained a custom model you know uh they have trained a model on their data set but multiple tools are there in the community right now you'll see in IT industry has gone crazy about you know with these tools like GPT power tools we see Dr GPT we see Bloomberg GPT we also have something called kisan GPT you know kisan is an you know basically a Hindi word okay and it's basically relates to the farm or the agriculture so kisan is for Farmers so we have Farmers back over here in India then equip-based agri-focused country okay and you see it says kisan GPT is an AI voice assistant for agriculture related queries based on open AI chat GPT and with some models okay so they have combined both the models that can be used for open source model by open Ai and also they have an API now and they are using chat GPT API for example to build this kisan GPT where you can ask all your queries related to agriculture the farming okay this somebody has built and the credit goes to them you can see over here kisan GPT okay by the way they have multiple languages they have multi lingual capabilities okay and we have something called Gita GPT as well it's so fascinating to even talk about guys right we have multiple tools so every day you know every hour of the day we are seeing some GPT powered application or genetic power generative AI powered application now if you see Gita GPT for example so it's the holy book you know uh in India and across the world people follow this bhagavad-gita which is the holy book and you can see it says ask bhagavad-gita question so you can ask all your queries to you know uh this tool these are basically for a specific domain you know an industry now we have agree based you can ask the same query to charge GPT I am pretty sure that chair GPT will also be able to answer any agree related question that you have any you know holy book related like Quran or Gita or Bible you know even chat GPT will be able to answer it okay of whatever questions you asked Till You have some kind of custom data which data which is very confidential and you know you want to train using some llm models then you can use alpaca or llama or any other generative AI models which are open which are open so that you can you know use it but we have seen this tool a few of them are trained okay on their data set of the customer data set using open source models most of them are again backed by GPT 3 3.5 turbo or gpt4 if they have got the access or chat GPT apis okay most of them are built and this is what we are going to do as well guys so we are here using the same open AI apis to build this BJP GPT but on BJP document so what I have done I'll show you we are going to build this tool in this video guys okay so let me just show you some demo okay so I'll just ask some of the questions that I you know I was just playing with this tool I'll just say you know can you tell me or let me come down I'll say who is who is Narendra Modi this is my first query that I'm asking I'm interacting with this uh bot now okay let me first uh ask this question and see what it responds so you can see it over here my question is who is Narendra Modi and I got a beautiful response it says Narendra Modi is the current prime minister of India he is the leader of the vatidanta party which is again BJP and has been in office since 2014. this is my first question I can ask some of the questions you know that I also was playing earlier I just said right I say can you you know can you list uh can I list a few of the achievements of prime minister Prime Minister Modi and this is my question let me just you know quickly fix this so I'm asking some follow-up questions now okay and we'll also test the context and memory part now this is a very very basic bot extremely powerful you see the capability of this kind of you know uh tool now it's up to you that how you want to extend this tool further you see I asked this question can you list a few of the achievements of Prime Minister Modi and it's a yes some of the achievements of Prime Minister Modi include adopting a whole of government approach to ensure coordination across various sectors for pandemic management launching the Iceman father we have Pradhan mantri yojana the world's largest health insurance program you know we have multiple programs he is coming up with his aims Iams iits and all these uh Academia and you know trying to uh create a better ecosystem for Education the healthcare you know different you see the highway is being and the infrastructure in India has been booming like anything so few of the achievements now if I ask a follow-up question can I ask this question Pradhan mantri tell me more about this let's see if it can answer of your follow-up question tell me more about uh this is this is my question now I'm asking a follow-up question okay from the uh the previous answer and let me just hit enter and let's see so how we build this tool you know I'll show you that what kind of document that I have written I can see fantastic right Pradhan mantri yojana you know uh is a flagship health insurance scheme launched by the government of India GUI in 2018 it is the world's largest health insurance program and I I believe so can provide Financial Protection to around 18 crore households in India now it's a huge program you know by Prime Minister Modi and of course the government and provides cashless and paperless access to services for the beneficiary at the point of service you know it's fantastic uh I'm really loving this question now let me just ask some few of them other questions as well so I'll ask can you list down some of the priorities of BJP mainly BJP can you list down some of the priorities of BJP maybe from their uh election Manifesto now this is my question now I'm asking this question can you list down some of the priorities of BJP from the election Manifesto so uh what kind of document did I have used as so even all these documents are you know available in public domain you know just to tell you very clearly guys okay this is not related and I'm a fan of you know BJP when I'm a fan of Prime Minister Modi why I have taken this specific document I wanted to build something you know uh which which might Inspire some of my audience you know that how they can uh take their Knowledge Management or knowledge basis or the database or any kind of Text data that they have how they can derive some kind of intelligence to enhance customer experience basically right what we are doing we are enhancing a customer experience nobody wants to go and read a document and try to find out what are the you know priorities of BJP or any other political party in their election Manifesto nobody wants to read those Manifesto right so you need some kind of Bot where the citizen can interact with this board to understand what are the priorities in that Manifesto what what are the schemes that they can you know get benefited from so those sort of queries that they have and they if somebody has queries about Pradhan mantri or what more what Prime Minister Modi is currently you know doing some great stuff in in healthcare you know in manufacturing infrastructure that can be anything right so you need some kind of application that will help the citizens to interact Hazel free okay and that's why you know I wanted to build something which you know which makes some sense you can use the same underlying concept to build like IPC you know if you are Indian we have Indian Penal Code we have IPC no but you know very few people understand ipca rules and regulations in all those code of conduct you know so if you build something you know take those data you go and take this PDF which are available on Internet and try to create this kind of Bot we can call it IPC GPT for example or chat IPC you you might have seen in Industry we have chat PDF your product PDF file on that application you can interact with it but they have some limitation they have made it paid and this is what is going to happen guys everybody is taking the advantage of global AI literacy you know this organization this individuals but anyway that's a different story but what I'm what I'm trying to say here is that you try to solve a problem you take this data abundance of data which is available in public domain you can build IPC GPT you know you can build football tactics GPT you have so much of data available around football tactics what what are the tactics that Guardiola follows what are the tactics that future is following or jidan follow the Carlos absolutely is following or ten hag is following what are the go-to tactics you can create all this for aspiring coaches or aspiring managers in football this is one of the examples you know you can you can build it something in Telecom manufacturing education you you build something like testing GPT where you have these testing documents you create some kind of automated test cases you can do whatever you want to do you know that make some sense you are solving a problem with this uh generative AI models you know this is just an example BJP GPT that I wanted to show you now let's ask one or two more questions that I'd like to ask now I have also or let me just also ask can you list down some of uh hit schemes or let's say some of uh uh movie schemes now this is my questions is my next query and I'll just say okay let's see and you know you can also add voice capability here you can use some of the basic python libraries you can like if you see kisan GPT they are using whisper here they say whisper models where you're interacting with it they will transcribe this voice the audio they have a speech to text capability into it you go on you take this it's open source completely free very easy documentation few lines of code and you will be able to do it right so if I ask this question it says I ask you can you list down some of this schemes it says some prime minister Narendra Modi scheme schemes include Pradhan mantri yojana you know different kind of yojana that you know scheme that here and let me just ask the last question over here I'll say hey okay uh tell me more about this now this is what I'm asking now and let's see and this also helps you know from an organization standpoint that you you want to utilize the data that you have you know you try to generate some kind of intelligence or insights or even if you're not able to do that part because it might be very complex to understand what kind of influence that you have to generate this will help you enhance the customer experience you know if you want to attract more consumers or the more citizens toward your scheme or how you create a citizenship engagement platform this can also help you you know creating a citizenship engagement platform where citizens would like to engage with your schemes or your plan through this kind of interaction or interactive bot so ask this question tell me more about this Pradhan mantri jandan yojana and you say it says pmgdy the financial inclusion program launched by the government of India in 2014 when you know uh Narendra Modi became the prime minister of India and the program aims to expand and make affordable access to financial services such as bank accounts remittances Credit Insurance and pencils and whatnot right so this is the example the this is a demo guys BJP GPT you know GPT genetic pre-trained Transformer if the if you're not aware about this model you can go and read on internet there's a plethora of documents available for GPT now so this is the tool that we are going to build okay I'll show you that how we can build this tool okay so now let's start building this tool guys so if you see guys I am currently on this uh vs code IDE and I am inside the folder today I'm not going to write the code from scratch guys I will explain the code line by line and this code base will be available on the GitHub repository okay because most of the code components are already available that I have covered in my previous videos okay you can find that video uh I'll also share that video link in the description and if you remember in that video you we used a graduate application so we used gradu application which we have built a QA bot with help of Lang chain and GPT index also llama index and we use open AI to build that QA bot but in that q a box that we built the Grady application there were no history so we have not had memory or context we could not ask a follow-up questions you know and also see it right on the UI aspect I'm talking about so in this that we also have done it so you know to build this application or this sort of application what you need you need uh few of the dependencies and the app that you see the back end is again powered by fast API so fast API is a wave framework okay uh of in Python again it's extremely capable of building a scalable solution okay it has an upper hand when you compare it with the flask or even Django it provides asynchronous function capabilities it has a lot of other properties like pedantic Starlet you know the a better data validation so it provides features you can read about fast API on their documentation okay so fast API is also uh nowadays being used to build web applications powered uh you know again by Python and then we have we're using some of the libraries like dot EnV okay to handle the API key uh that we we have stored in that this dot EnV file and then we have some something called multi-part okay because we we are uploading documents right we are we are using document we need this kind of uh multi-part or AIO file that is required for by fast API because then we have uvcon the server okay that that will be used to run this fast API application okay then we have Ginger 2 which kind of you know uh helps in the templating because we are using an HTML as well okay so we are using a Ninja 2 template if you are familiar with flash you would have already worked with ginger or Ginger too then we have open AI you know the Revolutionary open AI okay we are using the libraries to interact with uh we're talking about cosine similarity and those things that we'll use uh especially there then again you know we have open eyes twice by the way we have a GPT index and I said I'll also explain you know why you're using GPT index here uh and then we have Pi pdf2 because we are uploading PDF file whatever operation that we you know perform on that top of PDF file we use pi pdf2 to perform those operations let me go back to app.pi guys or before that I will show you how the folder structure looks like so let me go back to my folder I have a virtual environment that you can see I have created over here that's called bjpgpt I've activated my venv and I have installed all my dependencies into this V and B this is how you create a virtual environment if you're on a Windows okay so python hyphen am VNV dot VNV and then you activate it with the CD vnvcd script dot slash activate and then you do pip install iPhone R requirements.txt to install all your dependencies I I have already installed it so I'm not going to do that so if you come back this is how the folder structure looks like we have something called docs now docs is very important because this is a folder where you are going to keep all your data you have one files or you have n number of files you know you keep all your data or the document here inside this docs folder now if I go inside this docs folder you will see I have multiple documents related to bharti rental party or BJP the political party I was talking about you can do it for congress as well guys if you are you know watching it from a congress fan you can also look at different uh political parties if you are trying to build this sort of solutions for politics and stuff you can also build it for sports and other things so these are the documents so we have BJP we have you know this booklet where we're talking about you know eight years of their uh governance or their whatever achievement that they have achieved their schemes everything then we have major achievements this is kind of listing it down you see what see it over here they have all their achievements you know gsts and multiple others right here and then we have their manifestos as well uh for 2019 election you know this Manifesto that you see now that depends you want to create a large Corpus Optics data go ahead and do that right so you can upload n number of documents you know related to you for solving a business problem you know then we have something called Static now static and templates these are basically a standard folder structure of the format when you are working with fast API or flask or you know similar you know web framework so in static what we do we keep our all our static things like in assets we are keeping our CSS files or JavaScript files or all the images like logos header Footers logos you are using multiple images you can put it over here in CSS we have style CSS okay so in static this is why we are using all this static components then we have templates where we keep our HTML files you know they are the templates so we have only a single HTML file now we have multiple HTML files you can also do that so we have index.html the UI that you saw and that we have a EnV file where we are storing our API key you know then we have app.pi that I will discuss that you know in history you know in history we are saving all our segments I will talk about it in bit and then we have index Json which is very important as well these are the way the embeddings are happening we are having word embeddings you know we are using DaVinci 3 to create this embeddings of the token we are getting this tokenization and all these tokens and this is what this is why we are using index to store those things right and then we have requirements file now go back to uh excuse me sorry I'll go back to this vs code and I will head back to app.pi now in app.pi you see I have imported all my dependencies or libraries like Fast API we have we are importing Lang chain for composability you know we have chunks of prompts even lend your prompts and provides multiple other features I have also covered Lankan in the same video that I was talking about the gradu QA bar you know then we are using Spacey numpy and all those standard libraries we are ignoring some warnings and then we are loading this you know core Web MD you know this is a side we have SM we have MD we have large and you can install it from the Spacey website I will talk about it where are we using this and then we are initiating this app using fast API this is how you initiate an app okay so app equals fast API this is a method if you are using flask you replace this with flask and then we are mounting this static we are you know trying to assign the directory okay this is the directory that we are reading the static through this method called Static files then we're also defining the same for templates we are using a ninja two templates and that's called a templates folder this is what we are doing here then we are using open AI API key you know the OS dot environment using the pi you're using this load dot EnV with help of this OS dot get EnV open Ai and API key now I will discuss these functions in bit now the first thing if you come down I'll first show you this app dot get the thing that we are doing over here okay see we are using an async function you know that's a fast API inbuilt async Define index and we are using this index HTML now what this does this kind of whatever UI that you are using you use this function and it will be able to you know present the UI for you okay yeah using fast API this is what we are doing now if you come over you will see we have written multiple functions python function the first thing is Define save underscore history to file there you see a history.txt file over here right we are saving it over here in this uh working directory now I save history to final history is very important guys okay now if you see chat GPT a revolutionary charity I'm using this word revolutionary okay revolutionary chat GPT in the left hand side you see a list of your history okay the history shows all your previous interactions and they are kind of when they started initially there was no history you know in the first few weeks but later they realize and that's that's how it also became very powerful people started using it and they started liking that feature of the history this is what we can also do guys okay we can create a history.txt by the way if you are building a very escalable or large solution you might have to create a database there okay a txt file will not work in that case Okay uh and you create a keys like user IDs and then you map that uh history with the rest effective user IDs okay then we are creating this history in the load history file very much self-explanatory if you are if you know programming we are loading this file now so we have defined a python function function load underscore history underscore from underscore file and we are you know importing this history txt and now what we are doing we are looking at a NLP technical natural language processing techniques that's called cosine similarity how similar to strings are okay now I eat mango and you know I like mango okay how similar these two sentences are for this kind of you know sentence you know basically this helps in recommendation engine uh when there are text Data involved or movie recommender you can also use this kind of technique or basically for search for search cosine similarity were used like anything you know in years ago now we have better algorithms and techniques like word embeddings now we have better indexing we are using fast you know Facebook AI Source uh system then we have a lot of tools as well which kind of uses different now we are using this Transformer based models you know for embeddings and uh performing this search but cosine similarity also uh it does the job guys okay so what we are doing here we are using this NLP now we have used Spacey uh dot load to load this model uh English core web uh medium size model OKAY NLP and we are you we are passing our both of the segments so A and B you see what it says you know in this uh used to compare the similarity between the user input what user input is and the segment which has already been stored in this history.txt now if I come back on this hd.txt you will see all my questions over here and I asked this question if you come down can you list down some of Modi schemes you can see it has been saved in this history file who is Narendra Modi tell me more about this okay or whatever it is now if you go back to app what it does is first doing some pre-processing like stop words we are handling those okay and then it tries to find out the similarity between both the uh components like the similarity between the user input and the segment already stored in the history okay this is what we are doing here now we are sorting those history guys it's very important now if you say what it says sort the history of interactions based on cosine similarity between the user input and the segment in the history okay history listing of segment that already uploaded and we are using this separator by the way let me come down I will explain that we are using this separator you know so if you come down to come down to this history.txt you see we have a separator for each questions and answers okay you see this uh hash that we are using right this is a separator and I'll say what are this uh U and AI so we already have created an initial prompt okay for this conversation so we have this default U and a I so U Is the End user and AI that we already have you know using this gpt3 power DaVinci 3 Model okay that we have used an embeddings by the way so we already have this initial prompt so in prompt to what we are doing we are taking this input from the end user you see this U and then we have this curly braces where we are storing the input and then we are combining this initial prompt one plus prompt two to follow to follow the conversation okay from here to the uh user input and we are using a separator to save this in history.txt so if you come up you see here we are sorting the history between user input and segment then append in the list the list that you're present empty list similarities and then we are sorting the similarities using numpy.org okay we are using this Arc sort method of simulating we're sorting it and then we are saving history to file sorted history okay if you come down you're getting latest in from history get the latest n segment from the history is again the same thing right and if you come down now here is the interesting thing guys construct index okay History part I hope you understood okay you look at you run this code you will understand you can also customize this history in a way you want to customize this is just the logic that I have learned and I have you know created uh using this techniques and methods you want to customize this you can customize it completely you can also use some other techniques now if you come down uses construct index function is the backbone of this entire app okay what we are doing you know we can also increase this number output by the way let me improve it to one zero nine six okay now we have Max input size okay the input size that's where line chain also helps okay so we have four zero nine six tokens we as this is going to be the max input size then we have number of outputs the 1096 we are getting this outputs from the BJP GPT that you see right number outputs 1096 which is basically nothing but the number of tokens and these are the maximum token that one we one can get in the output when they are asking a query okay then we have chunk overlap and chunk size limit please watch the previous video to understand this more this parameters in details where line chain helps it's you know chunking out your in your in input prompt it chunks The Prompt because you might ask you know landlier questions no that's where line chain also helps and then we have the prompt helper again passing this parameter in this prompt helper method they're using this llm predictor by line chain we are using this open Ai and we are having having temperature 0.7 you can again tweak this temperature you don't want creativity or Randomness please make it zero if you don't want that but you can play with this okay and then we have loading this uh using simple directory reader method you know from this docs folder that we have stored and then we are passing this to GPT index GPT index kind of uh takes this documents in a data structure you know it's and you are using llm predictor and prompt Helper and then it index it okay and then we are saving this indexed save to disk index.json file over here and then we're returning this index once we do once we do this guy once we've completed this step now we have index.json we have the history functions written over here now we are creating this QA board we have input text okay we are using all these functions we are taking this prompt you know and we are we are using all these functions to return the output if it's over if you see it over here index same thing we're using this function in we are loading this index.json and then we are passing uh the responses through the query that has been asked you see index.query full prompt response more compact again this is very interesting response mode you want a compact response you want a linear response you want a very short response you can also read this in the documentation okay a GPD index and then we have some time thing okay that how much time it took to generate responses and then we are having this UI you know just enable encoder to uh for the chat Reserve here so we have to flow it on the UI okay so this is this is the way it works guys okay I haven't you know written the code in this video from scratch before I already have developed this I just wanted to explain line by line or the function now you can play around it the code will be available on GitHub now let me just run it and see so what it does guys once you run this okay it kind of when we are running it for the first time by the way just for your information you have multiple documents now suppose you have 1000 document PDF files okay or doc file or text file when you have those number of documents it might take from hour to two right hours to two hours or three hours okay because you have number of documents and you will see the number of tokens kind of goes into the lags and you know millions or billions you know on tokens that it says then I'll show you what I'm talking about you see there's a total embedding token images for this sort of documents okay we have this document like five documents and it says total embedding token uses it's like you know a hundred thousand or something right now if you have multiple document it might take up to minutes two hours to load okay now we're running with the first time okay now let me go back to application and we're gonna ask the same question okay I so who is again in the Modi kind of a stuff okay and I'll just run this no once I run this it'll come over here and you'll see let me show you in the terminal guys okay so we are printing it all this it's a total llm token uses 234 tokens Emery token is 109 tokens so Narendra Modi is the current prime minister of India and the same output time taken like two seconds or something okay if you come back it says the same thing like remember the current prime minister of India he is leader of the party and I've asked him tell me about uh you know BJP for example okay tell me about bdp and I'll just uh hit enter here and you can completely change your design you can add more features as I was discussing earlier as well in this video you use whisper you use text to speech you can also have this uh text read itself okay you can use some text to speech like etas or Pi httx3 or anything you want to use you can use it right you have total https a lot of other models are available the way you explore it says the right-wing political party in India it is the largest political party and something something right and okay uh tell me about what can you tell me about what can you tell me on GST scheme or something okay now this is going to be the last question guys okay I hope you got the uh underlying message uh that how generative AI can be used to you know enhance customer experience you know and that too in a specific domain okay like I'm talking about a chatbot genetic bi chatbot for a political party okay how this can be utilized in an election campaign okay or you are uh just for your citizenship engagement okay or your social listening platform that can be used for any uh those purposes now it depends on you that what kind of problem statement you are trying to solve okay now it says the goods and services tax is a comprehensive indirect tax on a supply of goods and services and you can tell me on GST scheme so it's working you know perfectly fine right a very simple design I have used bootstrap okay I will give you the code bootstrap a very simple if you go back to you know uh this thing here in the template index HTML receiver using booster five you know sort of things arranging some of the we are using a container and we have this classes so I'll give this I'm not a web technology expert guys but I can code a simple UI okay and that too with help of cold pain and bootstrap and all those things which available so I hope you understood okay you go on you find out some documents create a docs folder put it over there use LinkedIn and GPT index trying to create the embeddings and indexing and then you create a simple UI and then integrate it with the help of a backend with framework like Fast API or flask and you know you can build something like bjpgpt or kisan GPT or Bloomberg GPT or you know Dr GPT or Gita GPT you go build on for example un hdg is GPT why not you are building up something for a basketball or something for you know awareness child abuse that can be anything right so this is what I wanted to show you this video guys I hope you liked the video okay and if you have any doubt you are facing any errors if you think that okay this can be scaled up or is you can scale it for some other uh you know big problems you can reach out to me you know if you have some feedback and thought please let me know in the comment box or you can also reach out to me through uh social uh media credentials or crates that's available and please reach out to me uh if you have any kind of queries and if you like the content please uh hit a like button and if you haven't subscribed the channel yet do subscribe please share the Channel with your friends and to your peer that's all for today's video guys thank you so much for uh watching see you in the next video