RAG for Real-World AI Applications
Connect AI to your data for smarter applications
30 lessons4 hr 20 minPlus
Lesson 1 is free now. The rest comes with Plus.
What you'll gain
Build production-grade Retrieval-Augmented Generation systems. Learn to connect AI models to your own data sources, build vector search pipelines, and create AI features that deliver accurate, contextual answers.
Syllabus
30 lessons · 4 hr 20 minChapter 1 - What is Retrieval Augmented Generation (RAG)?
4 lessons1
Intro to Real World RAG
Hello everyone, and welcome to the course RAG for Real-World AI Applications. In this first lesson, let's answer the fundamental question, what is RAG? Well, if you've ever asked a question to ChatGPT or any other LLM and gotten a confident but totally wrong answer, then you've been the victim of a hallucination.
3 minFree
2Real World RAG Use-Cases from Top Companies
When talking about RAG, I've been asked on multiple occasions whether or not RAG is something that's being used in the real world. The answer is a resounding yes. In fact, some of the world's top companies are already implementing RAG inside of end user product solutions, as well as internally for operations.
4 min
3How RAG Works – A High Level Overview
In order to understand how RAC works, let's walk through a simple use case together. So let's say you wanted to build a company policy chatbot to interact with your company documentation. but unfortunately there were no LLMs trained on your company policy docs.
6 min
4Project Setup and Vibe Coding the Chat UI
Now that you know why RAG is useful and have a nice high-level overview of how it works, let's build a practical application that will use a RAG pipeline in the back end to answer questions specifically about the MDN docs. Now, we will build this application with React, specifically with Next. js.
14 min
Chapter 2 - A Basic (Naive) RAG Pipeline
16 lessons5
Gather and Clean Documents with Llama Parse
Now that we have a web application with an interface scaffolded to actually make use of our rag pipeline for something practical, let's start working on the actual pipeline itself. And the first step in any rag pipeline is gathering and cleaning the data that you want to use to supplement the LLM. Now, data could come from absolutely anywhere.
11 min
6Open Source Alternatives to Llama Parse
In this lesson, let's take a look at a couple of open source alternatives for cleaning up our documents. Pandoc is something that I ran across that, despite the look of their website, actually worked pretty well. Notice right down here, you can see what all document formats it can handle. It can convert to and from Markdown.
6 min
7All About Chunking for RAG
Before embedding documents, we should break them up into smaller chunks. There are three main reasons for this. Number one, chunking helps us overcome model limitations. Many embedding models and language models have maximum input size constraints. So splitting allows us to process these documents that would otherwise exceed those limits.
5 min
8Chunk the MDN Docs for our Project
In this lesson, let's chunk up the MDN docs that exist in our project. In order to do that, I'm going to solicit help from our AI agent, specifically from Claude for Sonnet. Here is the prompt that I'm going to give it based on what I think we need. So I say implement a chunking script.
16 min
9Compare Embedding Models for RAG
Now that we have our documents cleaned and chunked, it's time to start creating vector embeddings for those chunks and storing those vector embeddings inside of a database.
5 min
10Compare Vector Databases for RAG
In the last lesson, we picked out an embedding model. The next step in the process is to choose a vector database where we can store those embeddings. So why do we need a vector database in the first place? As we already mentioned, embeddings are just a big list of numbers, AKA vectors.
3 min
11Use AI Agent to Setup a Postgres Database and Drizzle
In this lesson, let's get Postgres set up on local with that PG vector extension. In order to make this work, you will need to install Docker on your machine. If you don't have it already, I've left a link to the website. It's really easy to download and install. Once you have Docker installed inside of our project, create a docker-compose.
12 min
12Update the Database Schema for Documents and Chunks
In this video, I'm gonna start our agent chat fresh so that we're not working with any previous context. Then we're going to start planning out the database schema, that database architecture together with the agent. To do that most effectively and quickly and naturally for me, I'm going to enable the dictation feature on my machine.
14 min
13Use AI Agent to Create Database Seed to Store Documents
Now that the database is set up, I'm going to ask the same agent that was working on setting up that database to seed the database with the data from the MDN docs that we already downloaded, as well as those chunks that we've already created.
9 min
14Create Embeddings and Store in a Vector Database
Now, we're ready to start creating embeddings for our chunks. We've already decided to use Voyage AI to create those embeddings, but we're also going to use the AISDK as a JavaScript library, really a TypeScript library for interacting with Voyage. Why? Well, because AISDK supports tons of different LLM providers out of the box.
15 min
15Retrieve Semantically Related Content
Now that we have embeddings in our database for all of our different chunks, we can generate an embedding for a question and then search for semantically similar chunks inside of our database. Ultimately, we want to accept that question through the UI, the web app UI that we created a little earlier.
6 min
16What is Cosine Similarity and Top-K in RAG?
Now that we have the code generated to retrieve document chunks relevant to a user query, let's talk a little bit more about how it works. First inside of the semantic search script, we have this generate question embedding function. This does the exact same thing that we did for each of the chunks to the user's question.
6 min
17Augment Generative AI Input with Related Content
In this lesson, let's augment an LLM generation with our relevant docs. Of course, generating anything with AI means we have to have a provider to do that for us. However, Voyage AI doesn't really generate tags.
16 min
18Documenting RAG App Scripts and Service Functionality
I know, I know, I told you in this lesson we would hook up our RAG pipeline to our web app interface. But hey, we're not vibe coding here. I don't want to get ahead of myself. I do want to take just a few minutes to go over this readme file that was created by the agent in the last lesson. I think its existence is actually pretty nice.
9 min
19Hook Up the RAG Pipeline to the Web App
Off camera, I went ahead and prompted an AI agent with CloudForce Sonnet chosen with this prompt. I said I have a RAG pipeline created, see at readme. markdown, which points to that scripts readme that we just looked at in the last lesson, and an API UI pointing it to the source directory.
13 min
20Streaming The RAG Response With the AI-SDK
🚧 **This lesson is coming soon.**
Stay tuned — new content is added regularly. You'll be notified when this lesson is available.
12 min
Chapter 3 - Evaluating RAG Results
5 lessons21
Criteria for Evaluating RAG Results
Now that we have a working RAG pipeline, let's talk about evaluating its effectiveness so that we can systematically improve its output over time.
5 min
22Get Started with PromptFoo for RAG Evals
In order to evaluate our RAG pipeline, we're going to use a tool called PromptFu. You can get started with PromptFu at promptfu. dev. That's the page that you see in front of you right now. It has lots of different features available, including red teaming, guardrails, model security, MCP, as well as evaluations.
11 min
23Evaluate RAG Context Recall: Retrieval Effectiveness
In this lesson, let's run a context-based eval. Specifically, let's check the context recall to see if the expected documents are recalled or returned from the database based on a question that we ask. I've already written a little code to demonstrate this to you inside of the evaluation, and then to-retrieval-eval-deterministic folder.
7 min
24Evaluate RAG Context Adherence-Faithfulness
In this lesson, let's talk about the context-faithfulness assertion in PromptVue and write a test using this assertion type in our code. So first off, note that context-faithfulness is useful for checking if the LLM's response only makes claims that are supported by the provided context.
10 min
25An Eval to Check for Linked Sources and Some Refactoring
In this lesson, I want to introduce you to a new eval, as well as some refactors I made to clean up our application. Okay, so let me show you how it started. It started with me adding this new eval, which is more of a custom eval. It doesn't fit into the category of context faithfulness or factuality.
6 min
Chapter 4 - Strategies for Better Accuracy
4 lessons26
An Overview of Strategies to Improve RAG Results
In this lesson, let's explore the key strategies that can dramatically improve your retrieval quality and overall RAG performance, including BM25 with Reciprocal Rank Fusion, Contextual Embeddings, and then finally Reranking. Let's start with a classic strategy, BM25.
6 min
27How to Implement BM25 for RAG
Now that you know some of the different strategies available for improving your rag pipeline results, let's implement something very close to bm25 in our pipeline. Setting up bm25 with a Postgres database actually requires an external indexing service.
17 min
28How to Implement Contextual Embeddings for RAG
In this lesson, let me show you how I've implemented contextual embeddings for our project. I've done all the work off-camera, and we'll go through the results together. So the first thing I did was create this contextualizer helper file. The only reason it exists is to house this generateContextForChunk helper function.
6 min
29How to Implement Reranking for RAG
In this video, let me show you how we could implement re-ranking inside of our app. Remember, re-ranking is about passing our matched chunks, along with the users and query, to a specialized LLM so that it can intelligently make a decision about which chunks are actually relevant.
5 min
Certificate
Previewunlearn
Certificate of completion
Your name
RAG for Real-World AI Applications
Shareable, with a public verification link.
Finish all 30 lessons to claim it.
Access
It's part of Plus. Lesson 1 is free to watch for everyone, no account needed.
Every lesson of every course, the full notes, transcripts and prompts, and the Unlearn-built tools. Blog posts and the tools that are public on GitHub stay free for everyone.
Yes. Finish all 30 lessons to earn the RAG for Real-World AI Applications certificate.
Yes. Team plans can assign this course, or a whole role, and follow progress in Team.
Next step
More in Agents and context
The Future Proof Dev: An Intro to AI for Web Developers
Full course · 11 lessons · 1 hr 7 min