Level 3 covers the two patterns behind most real GenAI products: retrieval-augmented generation (RAG), which grounds answers in your own documents, and agents, which call tools to take actions. You learn embeddings, vector databases such as Pinecone, Weaviate and Qdrant, tool calling, handoffs between agents and how to trace what went wrong. It takes about 25 hours over 3 to 4 weeks.
By the Build Fast with AI teamPublished Updated
Outcomes
Syllabus
Practice
Finish these 6 tasks before moving on. Open a task to see its steps.
Time: 2 hours
Time: 1.5 hours
Time: 2 hours
Time: 2 hours
Time: 3 hours
Time: 3 hours
Build
5 runnable notebooks, all free to open in Google Colab.
Build triage, handoff, and tool-calling workflows with the OpenAI Agents SDK.
Compare frontier models with cost, latency, and red-team checks using Promptfoo.
Route simple tasks to Flash-Lite and harder ones to Pro with a practical hybrid router.
Track traces, metrics, and evaluation runs for LLM and agent workflows.
Add replay analytics, benchmarking, and monitoring to multi-agent systems.
Build a chatbot that retrieves and references documents in its responses.
Create an autonomous agent that can search the web and synthesize information.
Build a system to analyze and extract patterns across multiple documents.
Full-stack application with AI code completion and intelligent suggestions.
Build a collaborative app with WebSockets and AI features supporting multiple users.
Create a mobile app (iOS/Android) powered by GenAI with offline capabilities.
Engineer a large-scale RAG system supporting millions of documents and high QPS.
Design and implement distributed agents orchestrating complex workflows.
Research and implement innovative RAG approaches with published benchmarks.
Research optimal agent architectures for different problem classes.
User and market research for positioning an AI product.
Analyze competitive AI products and create detailed feature benchmarks.
FAQ
A RAG roadmap is the order in which to learn retrieval-augmented generation: embeddings, chunking, vector databases, retrieval, prompting with context, then evaluation and reranking. On this roadmap that sequence is Level 3, after you can already build a basic LLM app.
Use RAG when the model needs facts it does not know or that change often, such as company documents. Use fine-tuning to change style, format or behaviour. Most teams start with RAG because it is cheaper to update and easier to debug.
Any one of Pinecone, Weaviate, Qdrant or Chroma teaches the core ideas: embeddings, similarity search and metadata filters. Pick a managed service if you want zero setup, or Qdrant or Chroma if you want to run it locally.
A RAG system retrieves context and then answers once. An agent runs a loop: it picks a tool, reads the result and decides what to do next. Many production agents use RAG as one of their tools.
Build a fixed set of 20 to 50 questions with known answers, then measure retrieval hit rate and answer correctness every time you change something. Tracing tools such as Opik and AgentOps show which step failed.