Large Language Models Bootcamp- Information Session

Key Takeaways

some of you are at least there on the zoom call uh and I think the webinar has also started okay so I will go ahead and start sharing my screen here we'll go ahead and get started here okay I will just get situated uh my screens and my monitors just align them and then we'll get started okay so so welcome to the information session everyone my name is rajal I'm one of the lead instructors and uh a chief data scientist at data science dojo and what we will go over today is uh the general curriculum the intent uh who should attend the boot camp um what do we teach um at the boot camp and what you will be able to do um after the boot camp and um and then also answer any questions um and uh I I will try my best to actually answer questions around um you know some open-ended questions that are relevant to the boot camp but sometimes uh uh you the questions can be around well um how do I build a career in data science while I would love to actually talk about that uh you know sometimes the questions can be so open-ended uh that I will not be able to answer um in the interest of time and in some cases well uh you know it might be difficult completely difficult altogether to really uh for me to give a very clear answer to things but we'll we'll do our best we'll do uh the best we can uh to answer any questions so uh um so a bit about us we have we are one of the oldest companies in this space in Ai and machine learning and analytics uh um space um and uh and I have to actually say that one of the most respected companies in terms of customer trust right so more than 11,000 graduates this is 11,000 people who have actually attended one of our in-person physically uh collocated or virtual programs postco um and then U more than 3,000 companies globally and you know some of you who are watching or many of you um I'm pretty confident that some of someone from your company uh has at some point attended one of our program so we have a a fairly big Global footprint uh in terms of the the boot camp uh why this boot camp we happen to be the only uh the boot camp in the world at the moment right so there's no other boot Camp as far as I know I mean so well I mean definitely some other trainings can be called boot camps but boot camp to my uh my definition of a boot camp is that it has to be complete uh it has to um cover the entire stack of whatever we are teaching whether it is data science or large language models um and in addition to that I mean there is this uh theory and practice component and you know the comprehensiveness theory and practice and really in the condensed amount in a condensed manner uh yes you can actually uh take one course here one course there and then you know combine all of that that's possible but what we do here is uh within five days you come in uh you enter the program and leave with the ability to actually build um llm application so that's that's where we are coming from um so the use cases of large language models that they are emerging uh as we as we you know see I mean right now companies they are trying to figure out you know how to build these applications and uh and so on um building these application building interesting Megan you only see a blank screen that is interesting so let me let me reshare my screen thank you for pointing this out Megan uh I will stop stop sharing and reshare and let me do it maybe I can move my screen to a different screen okay here sorry about that uh I did not realize because my own Zoom uh widget it is showing I can see it myself but I don't know if it was only for you uh Megan or other people are experiencing the same problem uh okay I am going to now uh reshare my screen here and um and some of you who are on the zoom call if you can confirm that you can see it okay and uh is it is the screen visible okay thank you so much for confirming okay so when you start building these applications um uh you know we there are some real challenges there are some challenges that happen um some challenges uh like uh you know how do you manage the token cost uh what are the limits uh on in terms of context Windows right um you know uh whether you should use a bigger contact window model or smaller contact model there are regulatory challenges um uh your company can actually get in trouble um a few weeks back Air Canada they um ran into this uh or uh ended up with a legal in a legal situation where one of their Bots uh offered um uh bement fair to uh to uh to a customer the customer came back and uh tried to claim that uh uh bement Fair discount only to be told hey I mean we did not tell it but your your Bot did uh and then Court ruled in favor of the customer so how do you handle all of those regulatory issues um when when we talk about Enterprise application when you talk about uh Enterprise llm applications we no longer are talking about that a single page uh you know PDF that you are um doing this as a hobby project we are actually talking about some serious application um um okay uh Megan you cannot see the screen uh again so can others see my screen Megan Megan it could be an issue on your side or it could be can others see my screen right now okay okay thank you so much so Megan may you may want to maybe disconnect and rejoin uh possibly it is an issue on your side thanks for confirming Ahmed and Arman okay okay I will I I will get back to this so uh so there are actually challenges around regulatory um requirements um you know you're you're a healthcare bot um and you cannot give medical advice you are a legal bot and you cannot get give legal advice and to what are the extents so uh you know if I if I keep going uh you know maybe um what kind of Bot are you deploying does it have to respond very quickly or it it can wait when it responds right so are there any data in AI governance issues uh will you be using open source models or close Source model open source being uh most notably Lama uh series of models and close- sourc uh uh open Ai and uh Google uh some of the Google models being uh um one of the notable ones um then there is this lack of reproducibility even if you ask the same question because of nature of these uh Technologies uh your answers may or may not be the same um hallucination whatever that means and we you know that is a problem knowledge is moving how do you evaluate a model uh there are so many issues and when you start building the models you realize well using chat GPT is easy but using uh uh using or building an llm application that your customers are going to be happy with uh any application that is going to keep the cost low any application that is going to you know deal with uh regulatory challenges and uh you know keep your proprietary data safe and keep the customers happy about how quickly it do respond responding um keep your costs under control it won't hallucinate or maybe if it hallucinates I mean there's a way to detect hallucination uh you will be able to evaluate whether the model is doing well or not it won't be very brittle to any changes in the prompt it is tough all of that is actually quite tough so uh the boot camp actually uh the large language models boot camp by data science Dojo it is going to teach you how to build llm applications on Enterprise data and once again sometimes you know even during interviews right so when you're interviewing data scientist yeah what's a big deal I mean I built that in an hour yes you can build an llm application in an hour uh cloning a g repository from somewhere but what if I asked you would you be comfortable uh it's a medical uh bot would you be comfortable actually giving it to your loved one to make decisions uh uh about their health or would you be comfortable deploying it as a a legal adviser for your own company uh would you be comfortable actually as a customer using this and suddenly the answers change so what we do is uh we have a curriculum that has been designed by practitioners uh I spend I'm one of the Le instructors uh for the boot camp I will show you other instructors as well and uh I have been working in machine learning for last you know close to 20 years now um and then um we uh and as a company we have a product that we have built and all uh and we have Fortune 500 companies big companies who are actually consuming that product so you will actually hear from us and other uh uh other instructors who are not just instructors and presenters of slides you will actually hear from people who are actually building these systems okay we cover uh most of the mainstream LM tools and libraries and packages I will explain what that means um and uh Hands-On exercises um maybe an hour or two hours of presentation followed by hour or two hours of practical exercises and on the last day of the training uh the fifth day it's a 5-day training eight hours a day fairly intense uh uh there is a comprehensive project with mentoring and support so you will leave with a deployed URL of your own application so that is the plan and right now we are offering the class in both formats some people they show up uh at data science dojo in Seattle and others they attend this class remotely so far we have had people uh from all over the world um you know as far as Australia know people from Middle East some people attend remotely many people they decide to travel uh we have had people from India uh uh France UK South America um and and different so um so we have the audiences geographically distributed many people actually decide to travel to Seattle to in the boot camp um so what does uh what does the uh boot camp include we want it to be a very complete experience so we do not want you to come in hey my laptop uh cannot install this package hey how do I actually sign up and my credit card is not working on uh to get this GPU cluster right so when you register um any API Keys any llm tokens that are going to be consumed any GPU clusters um any infrastructure that you need during the boot camp we provide that in some cases we take the cost in some cases uh you know our partners they provide the cost vv8 is our partner for uh Vector databases so you will when we do Vector database exercises you will be getting a VV server where you do all the exercises uh then um you know symol AI um Union is there arise is there you know Run part and uh blank chain Sano Academy all of them uh runpod actually provides us GPU clusters everyone gets uh you know a certain amount of credit uh when we go and find tune Al llama 2 model uh on custom data set so for the Hands-On exercise so you know we take care of all the costs associated with uh with any compute any Services any software so that is included and um we have partnered with some of the top companies in this space uh for the upcoming C boot camp so far we have these Partners I mean this list may change and we have other partners who have been involved with us they have been very kind to actually provide uh you know either instructors or infrastructure support to us um so the curriculum as I as I pointed out it is much more than just just Theory uh it is we also talk about the business aspect we also talk about the um you know the regulatory or um you know IP or pii type issues um we have guest speakers who bring in uh um will it would be uh they will be bringing some perspective from industry uh about adoption about so the talks can be real case studies it could be from ranging from you know purely business and cultural challenges that you make run into when you're building a large language model application or it could be purely technical right so different kind of issues we we balance them um and then uh there is a question that was posted here uh from YouTube will there be Hands-On product in uh projects in this boot camp yes I mean uh I'm I'm about to show you uh but there are going to be um this uh the what differentiate this uh this uh training is the completeness uh of the training um let me actually show you uh one picture here and then I will uh go back but yes the short answer is yes there will be a final project to build an llm application uh a complete endtoend application and deploy it and then there will also be uh you know exercises on Vector databases then exercises on Lang chain exercises on embeddings and exercises on observability and guard rails and exercises on fine-tuning and so on and I will show you those exercises in just a moment um so if you look at this somewhat of outdated version of you know the what the llm system might look like I need to work on these these slides so at the core of it um when we go and log into chat GPT most of us are only exposed to well uh you know what um an a chat GPT like application but when you build um uh when you build an uh an Enterprise application the ecosystem is going to be much more involved so at the core of these applications is your uh uh it is um uh um some kind of Open Source or close Source large language model whether you have deployed it locally you're self-hosting or you're calling the API um you can actually self host um some open source model like llama 2 llama 3 or you can use an API like uh Google gini or you can use an API like open AI gbd4 or gbd4 Omni then in addition to that uh you need a vector database and uh Vector database why do you need it uh you know just like a SQL database you need a vector database and what is a vector database think of that as you know where you store the embeddings of your vectors or embeddings of your documents so you need a vector database um and then and before you put anything in the vector database you have to convert them to embeddings so we'll talk about uh at length we will talk about embeddings we'll talk about Vector databases and at this point uh you know I can I can see maybe some level of uh discomfort hey I do not have the math background I do not have the programming background background I'm not a CS major uh will I be able to understand it absolutely and I will show you how um and I think the question actually popped up right here what is the level of codes uh scripting SQL knowledge is required for the boot camp if you can read python code uh so let me actually phrase uh or maybe scope the question um if you're not a coder before you came to the training um can you finish the boot camp successfully um yes uh but my assumption is if you're not a coder before the training you don't expect the boot camp to turn you into a uh a deaf professional because that's not the intent you will remain whoever you are if you're Pro if you're a product person when you come in you will be an llm products person if you are um a Dev uh a decent Dev who comes in you will become um uh you will you will know how to Implement these things so uh uh we'll give you quote samples I will show you the entire ecosystem in a moment um after this uh infographic is done um so if you know some basic programming you will be fine we can take care of rest of the things and um you know maybe it is hard to believe but when you see how we have structured it you will see you will understand why uh I'm making this claim uh um can I I work on my own project as a Capstone uh it is very hard I mean so if you look at this uh you know um uh yes you can but if you expect uh expect us to support on that project so uh this is your own project and if you bring into the bring into the boot camp and you know last days is allocated last half a day for 4ish hours they are dedicated to the Capstone uh in four hours uh it depends upon how ambitious your project is right so because uh we are just uh here to actually enable you show you how to build things but your own Project without knowing if you want to build your startup at the boot camp that's not going to happen right so I mean you know that's uh yeah um yeah I mean you can you can I mean you can bring in and um your own data and build it but we don't know what data it is and what idea it is right so I mean I'm a straight shooter I mean so I'm going to actually so I will not actually be around the bush to just to uh get you to sign up right um but the project the project that we do I'm very very confident when you go back you will be able to build things around it I mean I can I can vouch for that but I cannot guarantee that if you bring in your own data and you have your own idea we will be able to help you actually get up and running during the boot camp because I don't know the data I don't know the problem how how properly uh structured the problem is but um uh can we help yeah sure we can help can we guarantee no we cannot so I hope that answers the question okay uh um so you will learn um about Vector databases and um and in general what we call the rag retrieval augmented generation pipeline uh then embeddings of models um we have a very very detailed session on Lang chain uh Lang chain about four to six hours you will be a uh you will be a I would say a fairly um knowledgeable user of L Chain by the end of these six to eight hours because we have uh a very comprehensive moduel there uh then uh it's not only about building applications it is also about building performant and uh safe applications right uh so from a performance standpoint uh we'll talk about semantic caching we'll also talk about safety right so and logging and monitoring and operation um and deployment of uh models So to that and guard rails right so how do you stop your model from doing things that it should not be doing uh and so on uh how do you validate the model um so I can keep going but you will you will um you will you should get a fairly good idea I will move to the Hands-On component I will show you how the learning platform is structured but um the idea is to give you a little bit of you know not a little bit actually fairly detailed practice of the theory and practice of vector database theory and practice of embeddings and Theory practice of L chains and uh theory and practice of fine tuning and monitoring and Rag and all of that and to on the last day combine all of this put it together and you build an application and deploy it on streamlit okay and why should why streamlit because at the end of the day uh we want the is to be approachable to everyone if you're not a web developer by training this course is not about web development we cannot teach you web development streamlit is a very uh easy to use Tool uh devs of course can use it very easily uh those who do not are not devs even they can deploy uh application streamlet so it helps us focus on the The Core Concepts as opposed to the the finer details and the nuances of coding and web development I mean because that's not the intent of the course the course is um well if you're a developer you have Cod samples you have uh you know will'll give you the same thing if you're have a product guy you will be looking at uh you know if you're a product person you'll be looking at it from a product standpoint if you're are a hardcore Dev you will be looking at it as as a Dev and uh will look at it um let me actually go and show it to you this is our learning platform and let me actually see there is one more question here um is this boot camp intended for recent graduates this boot camp is intended for anyone who wants to learn how to build Enterprise llm applications you could be a recent graduate we have had people who recently graduated um we have people we have had people who were product managers uh did not know coding they attended it we have had um amazing awesome coders they attended the boot camp they loved it so it is it depends upon what your intent is the goal is to enable you Empower you give you the right uh resources and the right foundation and then you go back and build whatever you want to build um and Charles uh is this boot camp onsite in Seattle or is it virtual it is both actually so uh the way it works is that we have in our training room uh we have attendees um many attendees uh are iners uh sitting you know around you know uh and then there we have some people who are attending remotely so it's it's a mix of remote and inperson attendees and it has actually worked out uh far better than we had expected because uh you know um our conferencing equipment is that you know anytime someone focuses uh someone is speaking uh you know the focus shifts on them so you know you can see them who's speaking even the classroom right so you know we have this Advanced conferencing equipment so it's fairly interactive uh um fairly interactive uh you know online attendees in many cases in one of the boot camps actually the online attendees were actually very very active uh to the point that they were some of them were far more active than some of the inperson attendees so it's a very interactive environment is it better to attend in person or uh so you will not have any setup issues um and I will I will tell you and better subjective if you're talking about in terms of setup issues um uh I don't think uh I don't foresee any setup issues um it really depends upon your personal choice and your budget because when you're traveling of course you have to allocate some budget for actually staying in Seattle and you know traveling to Seattle so that would be there but otherwise in terms of the quality the content The Experience U pretty much seems to be uh the same except of course you know those hallway conversations and lunch and coffee conversations that you would have during breaks with each other of course uh you know online uh sometimes online attendees also I mean since the cameras are on people walking by talking to each other so they are part of it but there is that human component uh that maybe uh a small uh difference but but other than that uh you know in terms of learning in terms of uh the content the the experience the friction uh it is about the same uh so we have about it has varied but for the large language models boot camp the average size of the CL class has been around in person uh when we were purely in person we had 20 20 22 uh people I think that's U average fall around 20 and then online um you know anywhere it ranges from 10ish 10 to 15 [Music] so um that's that's roughly um that's roughly uh where we are right so maybe around think of it around 15ish on either side 15 16 on online and maybe 10 to 15 um uh so 10 15 to 15ish uh in person and then about the same 10 to 15 online okay uh so let me uh I will hold a pause button on the questions because I do not want to uh so please feel free to keep those questions coming uh but uh what I am going to do here is uh I'm going to uh actually um go to the um uh to the learning platform uh the way um it works is uh all we expect you to know is basic working understanding of python uh so if you can uh if you can write uh uh python code wonderful that's great but if you can understand python code by reading it um that's okay too but keep in mind we are not going to be able to actually if you have only reading coding ability and then you after the boot camp you will not able to build llm applications end to end from scratch I'm assuming that if you are you're a product manager you are some kind of manager you're you're a program manager who wants to learn and understand the ecosystem you'll understand it very well but then you will have to actually have a team of devs who will implement this but if you're a Dev already I'm very confident that you will be able to build application so you will get what you want out of it but you we are not going to magically you know there's not going to be some kind of uh you know uh transformation in a week that you know you come out of this as a rockstar coder that's not the intent of the boot camp very clearly for uh we have a two-hour basic Python tutorial um covering some Concepts that uh we will encourage you to watch before you come in other than that everything else will take care of it now um so um we will start with embeddings embeddings are uh the basic unit the most important thing in this whole understanding of you know this semantic understanding of uh Speech or language so we will first start with embeddings and uh you know attention mechanism and Transformer architecture and then we are going to have these labs around uh attention mechanism let me see for instance right so uh and the way these Works work is you're going to go and click on these labs and these labs they will open up like this and you know you have these Labs uh uh and you run these Labs then once we have gone through uh the idea of embeddings and let me show you let me drill um you know go inside of one of the modules so let's say once you understood embeddings now what do you do after embeddings you are going to actually store these embeddings in a vector database well what is a vector database so we are going to uh have uh you know we'll go we going to actually tell you what is a vector database in at length you know how is it different from a traditional database uh you know what does a vector database query look like uh this is real material that we actually I'm showing you from uh from the from the boot camp so we explain to you how does a vector database do search how does it index things uh uh how does a search work I mean what is hierarchical uh you know this indexing and organization of uh the material uh so we talk about this in quite a lot of detail you you will see that this is a very very detailed material but once we have done this we are not just going to leave you with the theory part of it we are going to actually ask you now that you know Vector databases how do you perform a vector search and and then uh you will click on this lab and you notice that I clicked here this lab popped up why are we doing it this way uh well if you are very Savvy and if you have your own local python environment set up and you you don't expect that you will be running into issues like you know resolving any package dependencies and your it does allows you to install all the packages and there is so many things that can go wrong and what we have done is we have created this dedicated Compu and storage everyone gets their own small you know uh space in our uh our uh learning platform so all of you you have your own uh versions of these and you can modify if you like and now if you look at this we are going to give you uh these uh you know setup Labs will give you this a you know URLs there is a vector database that is already deployed there uh and then we will walk you through the exercises here right so you know explaining hey what do you think this is happening and this is where I uh this is what I meant you should have the ability to actually understand uh the code even if you cannot write code from scratch that's okay uh so um if you look at this we are showing you near Vector example and near object example now that you're done then we actually go to the next Lab that is a similarity search this whole module is about four to five hours including Theory again you connect to vv8 you create a collection you import the data uh near text example and uh if you look at this uh you know if you look at this so how do you search for uh this query and then you know it is returning the related objects and it you also see see the embeddings and uh you know all of that um and then this is done then hybrid uh so now you're combining keyword and Vector search so um and keyword and your uh sematic search right so you're combining both of them then you have generative search and then you have multi-tenancy and then you have Vector compression right how do you compress these vectors uh you can expect them like to have hundreds of millions of vectors potentially for non-trivial scenario then how do you actually do product uh or vector quantization so um if you look at this uh we are not dwelling on Theory only and we are not also just going in mindlessly running code samples we are first setting the context explaining to you all the all the issues that surround um uh that's around building uh um or storing a headings in a vector database and then we actually walk you through all the exercises so you know it's completely you know ingrained in your mind I mean how how is this whole thing done and now after that you know uh how do you do semantic search and how do you build a rag Pipeline and then now now you know embeddings and now you know Vector databases so what is Rag and what is retrieval augmented generation then we uh talk about uh you know prompt engineering and uh you know F tuning of models um when you when we do fine tuning of models we explain to you what fine tuning is and how is it different from rag uh what is transfer learning uh I'm trying to remember very quickly here uh what is uh low rank adaptation what is quantization um some of these are very mathematical topics very deeply mathematical topics but you know uh we will explain to you in a manner so you understand um you know what it is and then we show you how to fine tuned llama 2 7 billion parameter model and compare it it against uh an unfin tuned model uh for the lack of better word right so you will show you how a fine tune model is performing and how a model without fine tuning is uh performing so you first went through fine tuning transfer learning quantization low rank adaptation all of that and then you apply all of these principles uh then we um we um talk about Lang chain Lang chain is a fairly big module such a big module that we have a separate complete separate course for it here let me show it to you so um let's say we are going to Lang chain here and when I go to Lang chain um let me see is it opening yes um so when I go to Lang chain um in Lang chain we are going to talk about you know all the major components of L chain startle from model IO um you know retrieval chains memory agents how do you build multi-agent systems and Lang graph which is something more recent um I will give a very high level idea what each of these are so let's say when I go into model iio how do you connect uh how do you connect um how do you create templates for different prompts because you you don't expect your users to be giving uh fixed prompts every time in the UI so we start with this you can see that there is this uh you know uh prompting exercise here so we'll uh go for uh prompting and how do you give few short examples uh how do you uh give it different types of models different models uh how a retrieval is how do you connect it to different data sources you can see that you know there are different kind of data sources different kind of parsers um is your data in CSV format is it in Json format is it cell is it sitting in a uh in a URL is it sitting in a SharePoint folder so we have exercises for different kind of uh uh different kind of parsers and retrievers and all of these exercises are going to be done in class so this is not something that uh you know that will be hey go and do it yourself you can see that we are taking um some piece of text here and we are chunking it if you can see you know basically and we are CH changing the chunk sizes chunk overlaps and you can see uh by changing these parameters you can see how these are changing so fairly Hands-On once again if you even if you don't know coding you uh you know you know what Chun size would mean uh What uh you know what overlap would mean uh and you're a product uh person right so you're product manager and now you when you go back you will be able to understand what your depths are talking about and what your direct reports you're maybe a Dev manager you're not writing code anymore uh or you know whatever your scenario is maybe you're a Founder who's working on a jni or llm startup you can still actually uh carry out the discussion uh you can still guide the product knowing what are the limitations uh without even knowing how to write code and these L these all of this content is accessible to you for one year uh from the date of the boot camp so you have plenty of time to actually play around and learn uh you can download code samples as well right so we don't stop you I mean if you want to run it locally you can download code samples run it locally if you prefer um then uh after that I mean we'll go through chains uh if you look at chains here right so what are chains chains is this idea that uh you know uh you are um you know basically uh the uh you are um ask asking a question that in connection to the uh uh or maybe you want you want your uh question to be broken into multiple um multiple pieces maybe I should show an example how these labs are run actually in the UI let me actually go here and I will copy my API key so all of you will be getting your own um all of you will be getting API keys so you are a ble to actually uh so you're able to run this code sample so I just copied and pasted an API key and uh how is it running now it is saying uh you know uh I am there is um you know your job is to come up with a classic dish from the area that the user suggest and there is this prompt template I run it and then there is the second part of the chain is now based on this uh dish that was recommended recommend a recipe and then after that given the meal so just a dessert that goes with it and then if you look at this now we are running these uh one by one and uh you know I think this is from my last training I don't know where everyone is based uh so let's say Seattle right so I will just put in Seattle here so when I run it for Seattle it is overall chain is not defined did I not run it I'm sorry real time real world demos okay it worked I think I did not run that cell so if if you look at this um and I will I know some questions are coming up I will answer the questions so uh to avoid any distraction here so one classic dish from Seattle area is the clam chowder right so and now clam chowder uh and the next thing is what is a recipe for clam chowder and out of that what is a good dessert that goes with it right so you you can see that uh it is giving me all of that so uh very practical very relevant uh if you look at this uh it is giving us all the options uh and I think during uh during I think my last one I think it was Mars I think that's what I did and these models also happen to have some a good sense of humor so it says Martian Red Red Rocks 2 and all of that right so you can see that uh and then we talk about the router chain and all of that so so you get the idea once again um do you need to be um uh a coder for this I don't think so I mean so you can actually because the code is already structured and we are you're being guided to actually go through the code um but will you be able to write code from scratch after this well it depends upon your pre-existing coding ability um so in addition to all of this that I've mentioned um we will also emphas heavily emphasize on evaluation so we talk about your traditional uh evaluation uh metrics like uh um more machine translation metrics like root score and blue score and uh then we slowly go toward more semantic evaluation techniques um like Bird score we slow one thing one approach that we take is we don't directly show you the shiniest thing you know it's almost I mean I sometimes um uh sometimes uh what happens is that uh um if you directly focus on the shiniest things um you don't have that appreciation and understanding uh of how this evolved right so sometimes going through uh you know um poty and then reaching that affluence it it makes you a wiser person I mean then there is this analogy that would like to use here right so we show you some of the less than perfect way of evaluation and in most cases we'll show you first we'll start with the less than perfect way of doing things and then move all the way to uh a better way of doing things so we'll start with um non-semantic evaluation approaches and uh once again anything that I'm telling you assume that it has a practical component right if you look at this evaluate uh rag with llm evals right so Lang email extraction so if you look at this so we will start with uh um we have a lab we'll start with the very basic techniques and then move on to the more some of the more um more sophisticated techniques that uh to the point that one uh you know there is something called ragas which is an evaluation framework also will dive into ragas as well uh and then we'll also one model evaluating the output for the model and all of that right something around uh GPD well we we'll talk about deployment uh we have um uh sessions uh on uh guard rails and uh monitoring your model and on the last day we'll give you instructions for uh building your own uh project and then we'll help you you know deploy an llm application so let me go here this is at a very high level the technology stack uh that uh roughly I mean in in this uh um realm I mean different areas we have this technology stack that we use uh give and take um we use all of this I've gone through all of this curriculum already let me see and so these are the expected instructors uh because uh you know it it's very hard to get on the schedules of all of these people I mean organizing this boot camp is actually quite a bit of challenge so but well uh here we are I mean this this was our uh set of instructors in the previous boo camp and uh I mean our partners they love us uh you know they have been actually very kind generous uh with uh you know providing all the support that they can and our customers also love us and uh believe me or not already I mean people are traveling all the way from like um as far as as I mentioned earlier as well Australia right so from India people have traveled from India to attend uh you don't have to travel but I'm mentioning right so so these are only some of the companies who have attended because you know we are unable to actually keep this slide updated uh the next one uh next boot camp is happening uh in Seattle October 21st to 25th uh and you know we would love to uh have you at the boot camp let me actually go back and start taking questions uh and going backward so Manish your question let me see Manish your question was what training assets can I take away with me so Manish all of these are yours at as long as uh our intellectual property is um is respected uh many of these are resources that are open source if you if there's anything that is open source feel free to uh you know use it but if you plan to monetize on it that is again going to be against deros service I mean if you just say I mean you know uh so starting tomorrow you're starting a company called data science Mojo right so that's not cool but anything else I mean it is yours we are pretty cool with actually you know do whatever you want to do with the material uh that's okay it is there we are pretty happy with you know providing resources um keeping letting people retain access we don't have a problem with that I mean uh and we want I mean this uh spirit to be reciprocated by everyone okay um I hope this an the question Manish so it is yours right so you can you can keep it as long as there's no commercial commercial repurposing of our content uh do we match the techniques to the specific elements of a task meaning decomposing a task and matching to the technique need it yeah I mean we try to do it so when we talk about uh rag right so rag has different components and we have a lot of I mean uh so you know once again I mean so when when uh a course is done by practitioners right so then uh we will actually be involved in like very very nuanced discussions and I I love teaching it because uh because people bring in hey Raja how will you do this and then suddenly you're scratching your head yeah I mean that's a very Val question right so so uh you know let me let me actually give you an idea that uh we are we are going to actually look at all of these minor points for sure I mean so uh in the in the greatest detail possible so I'm showing you one of the modules uh that I don't think I mentioned right so uh on around end of day four early day five uh we will have this uh this uh this session how do you build high performance rack pipelines they're not uh single page or 10 page PDF they are not just 10 files or 20 files you're talking about half a million files with complex formats uh you know native image native PDFs or image type PDFs or doc uh rag is for retrieval augmented generation so and you will know very well uh by the end of first day you will know very well what rag is so and so you know we'll say this is the r Pipeline and then we'll talk about well where are some of the problems that can happen uh you know our problems can happen data ingestion retrieval response and query uh pre- retrieval so and then we talk about all sorts of issues right so you know data pre-processing and chunking and embeddings and Vector database there uh uh the can be problems and it may sound something less approachable to you but by this time uh after that four days of learning people are actually they have very strong opinions and very you know this this session I've never been able to finish this I plan around uh two hours for this I'm unable to finish this in two hours because there's so many things uh and so many curve balls that are thrown at me uh you know what what about this and what about that right so because there is uh the field is still evolving there is no textbook uh things are changing very quickly um so we bring in a lot of different ideas right so what is the right strategy for chuning which embedding model should you find tune the embedding model or not um how do you scale your application what do you do with querying um do you uh and we talk about query optimization uh we talk about you know multi-step ret reble and query rewriting and query alteration and fine-tuning the qu fine-tuning the model that does query um that does query aeration so we talk about a lot of uh uh we talk about a lot of practical issues uh in here and this is not just this slide and this slide and this slide sometimes a slide may take you know 10 15 minutes just on a slide and you're now discussing you know all the you you're almost visualizing what may be happening in the rag pipeline so let me come back I mean because um I I think I started from the bottom so I'm I'm stuck in the questions that were asked later I will come back to the later questions if I can't join the complete meeting will the boot camp be recorded and viewed later yes Ling uh it is entirely possible uh to do it however um uh because sessions are long uh and when they're recorded as some kind of postprocessing they may not be available immediately within the one hour the next one hour or two hours right because sessions are long I mean it's an eight hour long uh day so if you miss some session it might we it usually becomes available the next morning not because of anything that on our side there's some post-processing that needs to be done uh and we have no control over that laws of pH physics um your question is is is there any pre-boot Camp material that we can review to prepare for sessions python coding Etc yeah I mean definitely we have this uh pre boot camp uh material that is there but if you are comfortable with very basic python we will teach you everything else uh there and there is a question by Sayad can someone with Fox Pro programming experience can learn it um it's a difficult question I've never done foxb programming I I but I mean I see I mean if it is programming if you told me I mean you have coded in VBA uh or maybe Pascal or Fortran programming so yeah I mean yes I mean it can be done it really depends upon your problem solving skills uh and uh and set up a time with us and we would never tell you uh and I hope that has been the case so far we would never tell you if we think that you you will not be able to do it and we think that whatever you're trying to achieve the boot camp does not align with it will let you know right away I mean so I think life has more to offer than deceiving people into attending a boot camp and make a few uh uh some money from them right so we'll we'll give you honest advice right so because you know we have enough uh people who are already attending would love to help you right so just let us know okay um will the will the bot be uh will the bot be develop Works similar to chat GPT meaning it will respond to national language yeah yeah absolutely right so it will respond to but it is going to be much more than chat GPT because what we are showing you is it is your own bot you will upload your document and you can ask questions about that document you will give it a URL uh you can customize the prompt maybe you are you added some chains so think of this as something that is your own um um something that is your um that you can um you can customize uh so you're programmatically customizing we also show you uh other things uh in your Bot um I think I did not mention it when I was talking about Lang chain so for instance agents right what if you want to give your uh your own chatbot uh some agentic functionality um so um how do you get the agentic functionality how do you how do you if you want to while you answer the question before your llm answers the question you want it to go and fetch data from an API or you want it to do perform a web search or you want to it to perform a weather search right for instance if I go and ask um uh an llm not chat gbt I think chat GPT has already buil it built it but llms uh their knowledge and their view of the world is Frozen at the time when they were built so how do we go and uh how do we go and uh um make our uh uh make our applications more aware of what is happening in the present right so you may want to you may want to have some kind of uh tool as we call it tools and agents so you may want to have a Search tool or some kind of programmatic API tool embedded and integrated integrated into it so there is uh a bunch of different things that are possible uh and of course you know in this one hour it is very hard to Encompass but the point here is to actually give you uh to give you this sense that it is um we have uh this has been buil built by uh the curriculum has been built by um you know the people um you know I'm one of the contributors but I mean bul of work has been done by by the amazing data science team uh behind the scenes who are actually building these products and they daily they would run into these issues I mean maybe I can give you an idea right so uh we have a product someone is using uh not someone some companies are using and now we thought we have built something very scalable and then um one F morning morning come a customer comes in here and says yeah I uploaded documents uh and some of the documents are not being indexed right so well how many documents did you upload uh 80,000 documents in one go well okay uh and then we went back and tried to fix it right so these kind of issues the kind um this kind of practical insights um yes I mean you can learn from from you know conference papers or you know a few presentations here and there but when when the people who are talking to you they have built things themselves um they uh they have uh either run into many of those issues and if they have not they will be able to appreciate the question uh and um give you a response and if they don't know they will let you know because they don't have any uh anything to worry about they don't have anything to prove because you know well I mean we we don't know everything right so uh I think that is there uh I think the last question is what is the cost of the boot camp uh the cost of the boot camp uh uh you can uh you can please take a look on the website because I don't want to miscot the the cost I mean um if you look at maybe someone from my team can actually post link to the web page uh on different uh you know in zoom and also on ac across different social media um but I mean check it out uh the full price of the boot camp uh is $5,000 but uh depending upon how early you register whether there's a group discount or not um the the cost varies if you are a data science Dojo alumni please do not register on the website we have a discount for anyone who's a returning customer and we have a substantial discount for those people so please do not register on the website if you are already a data science Dojo customer which actually interestingly 30 to 40% of the attendees are returning customers who attended our data science boot camp while back and they know um the quality that we deliver to our customers and uh in if you would like to know uh anything about the boot camp uh just you know set up a call I see that uh we have uh uh we have posted links to um how to set up an advisor call and we would love to help you uh and in our calls we call them more of uh you know so usually uh we don't use any heavy-handed you know sales Tech tactics right so we really uh go with the intent that we are going to help you and in many cases we'll tell you maybe maybe you should just go and do this if you're trying to do this maybe this food Camp is not for you and in some case will uh uh you know tell you that yes this boot camp is for you uh so just set up a call would love to help you out and if there are any other questions please uh feel free to otherwise uh we'll call it a day sounds good uh thank you so much everyone I'm looking forward to seeing at least some of you at the camp thanks everyone and thank you and uh we need to connect