Video ID: u5Vcrwpzoz8
YouTube URL: https://www.youtube.com/watch?v=u5Vcrwpzoz8
Added At: 13-06-25 21:19:14
Processed: No
Sentiment: Positive
Categories: Education, Tech
Tags: AI, Knowledge Management, Chatbots, Large Language Models, Fine-Tuning, In-Context Learning
Summary
The video discusses the importance of AI in knowledge management. The speaker shares their experience building AI chatbots for different business use cases and highlights the challenges they faced. They introduce two common ways to give large language models private knowledge: fine-tuning or training one's own model, and putting knowledge into prompts some people call 'in-context learning'.
Transcript
if you ask me what is one use case that clearly AI can provide value it's going to be the Knowledge Management no matter which organization you work in there are huge amount of weeky documentation and meeting notes that is everywhere and organized no better than a library like this it will take forever for any human being to read and digest all those information and be on top of everything but with the power of large language model this problem finally is having a solution because we can just get language model to read all sorts of different data and retrieve answer for us this why end of last year there was big discussion about where search engine like Google going to be disrupted by lar langage model cuz when you have lar langage model that has a word knowledge and can provide hyper personalized answer to you why do you still want to do the Google search and we already start seeing that happen there huge amount of people now go to platform like chat gbt or plexity to answer some of their day-to-day questions and there also platform like link focusing on Knowledge Management for corporate data and as many of you already tried it is actually very easy to spin up a AI chat bot that can chat with your PDF PowerPoint or spreadsheets but if you ever try to build something like that yourself you will quickly realize even though a lot of people thinks the AI is going to take over the world the reality is somewhat different many of time the AI chat B build probably even strugg to answer most basic questions so here's a huge gap between what does the world think AI is capable of today versus what is actually capable and for the past few months I've been trying to build different sorts of AI bar for different business use cases to figure out what is working what is not so today I want to share some of learning with you how can you build a rock application that is actually reliable and accurate so for ones who don't know there are two common ways that you can give large language model your private knowledge one method is fine-tuning or training your own model it basically Bak knowledge into the model weights itself and so this method can give large Lage model precise knowledge with fast inference because all knowledge already baked into the weights but downside is that it is not a common knowledge about how to find tune in model effectively because there are so many different PRS and you also need to prepare the training data properly that's why the other method is a lot more common and widely used which is you don't really change the model but put knowledge into the part of prompt some people call it in context learning but you might also just refer it as rack which represent for retrieval augmented generation it basically means instead of getting the large Lage model answer users question directly we'll try to retrieve Reven knowledge and documents from our private database and insert those knowledge as part of prom so that large L model can have additional context and if we want to dive into a bit more details to set up a proper R pipeline it normally start from data preparation where you extract information from real data source convert them into a vector database which is special type of database that can understand the sematic relationship between different data points so that when a user have a new question it will retrieve relevant information and send to larg Range model and if you want to learn how does Vector database and embedding Works in depth I actually made another video a couple months ago that is specifically talking about that so you can check out if you want to learn more the challenge of rag is that even though it is really simple and easy to start and build a pro concept building a production ready rack application for business is actually really complex because there are many challenges and problems with just simple rag implementation firstly the real world data is really messy many of them are not just simple text paragraph it can be combination of different image diagram charts and table format so if you just use normal data passer or data loader of this PDF file quite often it will just extract incomplete or messy data that large Lage model cannot easily process so many of the rack use case failed at the very beginning since it couldn't extract knowledge properly on the other side even though you create database from the company knowledge to Accurate retrieve relevant information based on question is also really complicated because different type of data and documentation normally involve different retrieve methods for example if your data is actually spreadsheets or SQL database Vector search might not be the best answer while keyword search or SQL search will yield better and more accurate for you for some of the complex questions it might involve knowledge across unstructured data like par paragraph text as well as structured data like table content and on the other hand sometimes you might just return a sentence within a paragraph that is most relevant to the question people are asking but adjacent content could be critical for answering the question properly and also some of the question people ask might seem simple but actually quite complicated in the rack context so if someone ask how is the sales training from 2022 to 2024 to answer this question properly CH model need to have contest from multiple different data source it might even need to do some pre calculation to answer this question popular so in short a lot of Real World Knowledge Management use case cannot be easily achieved with a simple KN rack but good news is there are many different tactics that you can use to mitigate those risks Jerry from llama index actually made a really good chart and summary about all the different Advanced rack tactics from some of table State methodes like better procor chunk size to some of really Advanced a gentic behavior today I want to pick up a few that I found works really well but before I dive into it I know many of you are either Founders or part of AI startup teams I'm always curious how does AI native startups operate and how do they embed AI into every part of business and H spot did they research recently where survey more than 1,000 top startups with heavily adopting AI to scale their go-to Market process and figure out what worked what didn't and what's the best practice for example they dive into how does AI in starup sales actually works what type of use case deliver the most impact on goto Market strategy across thousands of startups from how companies use AI to customer targeting and segmentation to developing intelligent pricing model and even looks into how logistic and supply chain startups utilizing AI to predict problems before it actually happens to significantly improve productivity and as a AI Builder I also found really interesting in what kind of AI tools goto Market teams currently are using this give me pretty good insights of the current goto Market AI Tax St to if you want to learn how AI native stups should operate and scale goto Market I definitely recommend go check out this free research doc you can click the link below to download this report for free now back to how can we create a reliable and accurate rack firstly better data password this is probably one of the most important but also the easiest one to improve the quality immediately so challenge as we mentioned before that the real world data is really messy if you're just dealing with website data a little better but once you get into format like PDF or PowerPoint the data start became really really messy and difficult for lar model to interpret because there image charts diagrams and all sorts of different things and even though there's huge amount of different data pass on platform like llama Hub or lanching already many of them if you try is not that great for example if you're using pypdf which is one of the most popular and common PDF passer when you try to read Apple's financial report it can often extract numbers and data incorrectly and most of the time data extracted is in a quite a massive format that hard to understand the relationship between different numbers and with a round number to starways of course your AI app is going to fail to answer the question accurately but luckily for the past few weeks there are a few really awesome new poer that large langage model native can help you prepare data a lot more effective one is the Llama part so this a passor that implemented by llama index which is team that probably has the most knowledge in the world about RX so they introduced Lama Parts a few weeks ago which is parts that specific on converting PDF file into a large Lang model friendly markdown format it has a lot of higher accuracy in terms of extracting table data compared with other type of par we normally use and it is a really smart passor where you can actually pass on prompts to tell the passor what the document type is and how do you expand them to extract information so you can even pass on a comic book PDF and provide some information that this document is a comic book most page do not have title it try to reconstruct the dialogue happening in a cohesive way result you can see here it focus on extracting only the dialogue and Main content and on the other hand you can even use that to extract Mass formulas accurately by giving a special prompt like output any Mass equation in latex markdown format then it can extract formul in markdown format that can be rendered as formul Pop so this llama password extremely powerful and totally change the game for rack on your local files it is already live on llama Cloud so you can use it for free I definitely recommend check out llama Parts if you need to handle a big amount of complex local documents but part from those documents like PDF file we also need to deal with a huge amount of website data that's where I want to introduce the second parer called fire craw so fire craw is introduced by MBO where they provide scer that folks turning website data into clear markdown format that large langage model can very well for example if I have this URL from for news where they have an article about AI agents if I paste this into fire CW it turn this website into clean markdown format for me with title and image as well and everything is clean public structur so this will reduce the amount of noise that the large L model actually receive a lot and also get all the metad data ready as well so if you want to do some additional filtering you can also do that and they allow to script single URL cross the whole domain or even search across web and the best part is because now I'm using llama parts for the local file and the file C for the website that data most of the data I need to handle is UniFi into markdown format so I just need to optimize my rack pipeline for the markdown format so this is the first part of better script the next one is chunk size so assuming you extract all the information from the website of local file populate to create a vector database we actually need to break down the whole documents into small chunks and toize each Chunk we can map all the chunks into a vector space where we can understand which two senten are more semantically similar to each other so that next time when user have a question we basically just toonize this question they got into the same Vector space as well and retrieve the most relevant chunks adding them as part of prompt so that lar Range model has a contact to answer question and one of the key factors that going to impact the performance here is the chunk size which is how big each text Chunk should be so one question you might have is why do we even break down documents into small chunks why don't we just keep size as big as possible so the large L model has full contacts so there are a couple reasons one obvious reason is of course large L model has limited contact window so you can't just feed every possible things into the prompt and even though you can feed everything in the performance often not that great because of the loss in the middle prop that's basically a phenomenon that we observe when you feed a big prompt to motel the large damage Motel will pay a lot more attention to the beginning end of The Prompt but the things in the middle can often get lost and there are a lot of different type of tasks people did already that showcase even for big model like Turbo with 1208 model once the contact window passed around 70k lar Dage model will fail to extract some contact and content from a large PR but on the other hand if you keep the chunk size too small then it also has a lot of problem because the information it retrieve probably don't have full contest for large Lang model to understand so there's a lot of trade-off and balance you need to find between the chunk size different type of documents can have its own optimal chunk size and the most scientific way to find the optimal chunk size is experiments so you can play with different chunk size and maybe even predefine list of evaluation criterial like the response time the faceful and the relevance then do evaluation again your testing data sets with different chunk size to find what is the most optimal chunk size for your document type and one my colleague from R side called Satia actually did some quite an interesting implementation so because different type of documents can have different optimal chunk size what he did is he tried to figure out a optimal chunk size and whole rack pipeline for different type of documents and then when we receiveing new documents we'll just try to classify these documents and give the most optimal rack configuration if the file I upload is resume.pdf then it will router to the best practice of r documents where it will choose the right passer and pass prompt as well as the optimal chunk size and retrial methods so this is second technique try to play and find the optimal chunk size for specific documents the third one I talk about is reent so this is a common tactic that we use to improve the retrieval accuracy and part I try to optimize is the relevancy of document when we do Vector search against a user question if we Define topk to be 25 which means we want the vector search to return the top 25 most rant chunks the chunks it return has a mixed level of relevance and they are not sorting a way that most relevant document is at top in reality the most relevant chunks are spread across return chunks so if you just simply passing all 25 chunks to the large model it will have a few problems one it will consume a lot more tokens and second it have a lot more noise so the answer quality going to be a bit lower common method here is called reim instead of sending those 25 chunks directly to the large energ model we can use another Transformer model specifically to find the relevance between documents so we can pass on this list of chunks use ranker to pick up the top most rant chunks out of initial search results so that the answer generation will be faster and more accurate another comment is called hybrid search as we mentioned before Vector search is not necessarily the best search methods for many use case for example think about e-commerce side where user search for a product you actually want to sure the product name is exactly matched from actual product name in your database and to making sure the result is super relevant we want to do a keyword search this where hybrid search can offer much better results and the way it works is in instead of just doing Vector search we can do both Vector search and keyword search and then mix both result together pick up top most randant ones so there's a few quite common and practical ways you can improve your rag pipeline but the part I want to really talk bit more is the agentic rag so till now you probably realize there a lot of different rack techniques you can use and the real challenge here is that there's no real best practice across all sorts of different documents and the beauty of agent rack is that we can utilize agents Dynamic and reasoning ability to decide what is optimal rack line even do things like selfcheck or chain of s to improve the answer and one very simple but powerful method is query translation of planning so the idea here is basically instead of doing the vector search with the question user ask brly which in many of time is not opal for Vector search we can get agent to modify the question a little bit so that it is more retrieval friendly for example if the user ask a question the went to which school between August 1954 and November 1954 directly do Vector search against this query might not yeld the best result instead we can abstract this question to be something like what was L's education history then to aor search against this modified and more abstract question to return the four results and this is Method called step back which is introduced originally from Google Deep Mind and same thing you can even utilize similar concept to get agent or L model to modify the question a little bit before it doing the retrieval so if user had question like how's the sales trending from 2022 to 2024 the agent can actually break down this complex question into three sub queries and each query search the sales data for that specific year then we can merge things together so that the lar model has four context and on the other hand you can get lar lar model to do some metadata filtering and routing as well so each documents we got from customers you can give some metadata like title year Country summary and this will be extremely useful because we can combine with some agentic Behavior so instead of just do the vector search across all the possible database you have that probably returns some Reven data from some of database that that is not relevant at all you can get agent to generate metadata first so if the user ask best burger in Australia you can generate metadata to the country in Australia first then futter down the list of data that it will do Vector search against to only C and brisin so the result it return will be much more relevant and you can imagine all those techniques that we mentioned here can be a tool for the agent and when we receive a new question we just get agent to decide whether you should utilize certain tactics to improve the result on the other hand you can even introduce some kind of self-reflection process into the rack pipeline to improve accuracy one of the most popular concept is called corrective rack agent it is a pipeline that really aim to deliver high quality results so if the user have question and after we're doing some retrieval we will get L model to do some evaluation and decide if the retrieve documents are correct or relevant to the question that we were asking if the classification is correct then we'll go through some process to do knowledge refinement to clean up and knowledge but if it's ambiguous or incorrect then agent go on internet and search for some internet results instead and repeat this process few times until it feel like got a correct answer then it can generate the results from there so by adding those self-reflection you can see the quality of this Rec P will be much higher even though there are tradeoffs in terms of speed but the quality of the answer is going to be much more rant and accurate and today I want to show you a quick example of how can you build a corat rack agent with llama stre on local machine as well as file crawl for the website script going to use GL graph to build this creative rack agent lens from L chain actually give it very detailed a tutorial about how can you build such agent with land graph but today I just want to introduce a simplified version with some script that we just introduced before so the way to work is that when the user asks a question we will try to retrieve most relevant document but after that we will get lar model to grade whether the documents retrieve is relevant to the question you ask if yes then go generate answers but if not we'll do web search and we'll use TBL which is a web search engine designed specific for agents and after answer is generated we also do another round of check where the answer is actually hallucinating if yes generate again if no check whether the answers actually answer the original question if not then go web search find random information and repeat this per again until the question can be answered and as I mentioned before we're going to use land graph so it basically allow you to define the high level workflow and Logics but still getting agent or larange model to complete task at every single stage it still give control about what the flow look like but utilize Lear model capability at every single step to complete tasks and we're going to use llas three as decision-making model here so first you're going to download o Lama which allow you to run llama streight on your local machine D and once you download that you can open your terminal and do old Lama pull llama 3 this is where download the Llama stre model on your local machine and after that let's just do a quick test we can do all llama around llama stre so I can type hi who made Facebook so I'm running this model on my MacBook and you can see the speed is actually still pretty good so once we confirm that you can run this llama 3 Model on your local machine we can close the terminal and open Visual Studio code we can create a dup notebook by called rack agent l street. iynb so this will create notebook and I will run you through this example so first let's install some libraries going to use including lanching langra T GPD for all which open source embedding model they can run on your local machine as well as file core and after that I will set my lens Miss API key which should automatically log all the interactions so that we can keep track and I'm going to set up a variable called local LM equal to Lama 3 and first thing is I want to use file code to create a vector database from a few blog post that I have on my website so I will import a few different libraries Define the list of URLs then I will run file CR loader so file CR does have a l chain integration already where I just need to pass API key and this a mode as well so script means your script just individ URL you can also change to craw which will C through the whole domain and then I'm going to split up the documents into small chunks each chunk 250 and also future out some metadata because at default file Crow return some metadata as array which is supported so we're going to clean up those ones and in the end create a vector database using gbd4 all embedding as well as the future documents and in the end create retriever so that we can use this retriever to retrieve relevant documents in this Vector database anytime so now we get retriever ready next is we want to create where the document is relevant to the question so we create a retrieval Creator Define large L model with the chat of llama that point to the Llama stre we created before then we will create a prompt template so LL stre has very special prompt style that you need to follow to making sure the performance is good you can click click on the link below to get more details about their prompt style but it normally looks something like this you have beginning of Tex St header ID which is a rooll and then the message itself and here we can quickly test if I give a question how to save lar Modo cost it will give me a score of yes and yes basically means it is relevant no means it is not relevant but if I change the question to be something like where to buy iPhone 5 the score will be no so this will be the first checkpoint to decide if the documents is relevant if yes then refine documents but if the answer is actually relevant next is we want to generate the answer using L stream model and create large model chain called rack chain and for the same question how to save lar Modo cost you can see it actually retrieve information from my blog post pretty accurately but if the retrieve document is not relevant then we want to do a web search and as I mentioned before we're going to use Tav so tavet is like a web search service for large langage model where you can just give a natural language it will return search results very similar service to EXA so here we're just going to put in your TBL API key and then create a web search tool so now we have create document grader a l engine model step to generate answer as well as web search the last part is that we want to create some function to check if the answer is hallucinating and whether the answer actually answer the question so create a hallucination grader with the special prompt here and again the result will be yes or no if yes it is not hallucinating if it's no that means it didn't pass the criterial as well as a answer grader with the same yes and no message so that so that we can keep the result pretty consistent that's pretty much it now we have all the key components ready and next we just need to turn them into different function and set up a lang graph State and notes so we first say set up a langra state so state is like what kind of value that you want to share across all the different steps in our case it will be the question us ask the answer is that lar mode generated the research result from web search as well as a retrieved documents then will create different notes one is retrieve node which is responsible to retrieve documents it basically to just call the retriever that we created earlier and return documents and question this will basically override Global State then will create this function to create documents to see if it is relevant if it's not relevant it will stop the full loop and just to set web search to be yes if it's relevant it will keep checking every single documents and then generate note which will call the lar model to generate answer as well as a web search then we create a few conditional age so you can think about all lines between nodes to be age it can be simple age that connect two different nodes together or it can be conditional age that going to run some function and based on result can route to different notes so here we're going to create two conditional age function one is going to base on result whether document is relevant or not to decide whether to do web search or just generate answer and the other is going to check whether the answer is hallucinating if it's not hallucinating then it will check if the answer actually anwers the user's original question and that's pretty much all we need and then next thing is we're going to add all the four nodes that we're going to use in the end we're going to connect everything together so all Lang gra will start from entry point and here I will set entry point to be retrieve document first and then I will add age so AG as I mentioned before is a link between different notes and here I'm going to connect retrieve notes to grade documents and I can also add conditional EDS which means after grade document is notes I want to run this function to decide whether it should do web search or should just generate answer from retrieve document right away if it's web search then after web search results I wanted to connect generate notes to generate answer and after we generate answer I want to run this function to decide if there's any hallucination and if answer answer the question if it is hallucinating then go back generate again and if the answer didn't answer question I'll do web search if it's actually good and the workfl and in the end I can just do workflow. compile and test out this question how to save large L mod cost and in this log you can see that so it first say retrieve documents and then start check every single document to see if it is Rel to the question in end addition is that all the documents revant now go generate answer and after the answer is generated check hallucination and decide the answer is actually grounded with the information will retrieve from documents and then it check whether the generate the answer answer original question and it decide it does so you finish all the checks and then output the final answer so here's example of how can you create this fairly complex agentic rack and as you can see that those agentic rack obviously have very clear tradeoff that it is a lot slower to generate a quality answer but upside is that you can actually making sure the quality is really good and documents relevant I'm really ke to see what kind of interesting rack agent that you're going to create please comment below for any tactics that has been really effective to you that I didn't mention here uh I will continue to post interesting AI project that be building so if you enjoy this video please come consider give me a subscribe thank you and I'll see you next time