1
The Knowledge Problem
5:00 – 5:15 PM 15 min
--:--
Key Talking Points
  • Ask the class: What's your company's refund policy? Now imagine asking Claude — it invents one.
  • This is THE #1 problem in enterprise AI. Claude is smart but doesn't know YOUR business.
  • RAG = search your docs first, then answer. Simple concept, profound impact.
  • Every Fortune 500 company doing AI needs this. You're learning the most in-demand pattern.
Demo Cues
OPEN: lecture-guide.html
Ask the Class
What's your company's refund policy? Now imagine asking Claude...
Transition
So we know the problem. Now let's look under the hood — HOW does RAG actually work?
2
RAG Architecture Deep Dive
5:15 – 5:40 PM 25 min
--:--
Key Talking Points
  • Three levels: No data (hallucination), stuffed context (simple but limited), RAG (scalable).
  • We're using Claude's 200K context window — no vector DB needed for our scale.
  • Walk through the trace demo: question → search → rank → verify → generate.
  • Key insight: the search tool IS the agent's knowledge. No tool = no knowledge.
Demo Cues
OPEN: rag-trace-demo.html
Ask the Class
Why can't we just paste all documents into the prompt every time?
Transition
Enough theory. Time to build. Open your Colab notebooks.
3
Hands-On — Build a Knowledge Agent
5:40 – 6:25 PM 45 min
--:--
Key Talking Points
  • Everyone open Colab. I'll walk through Phase 1 together, then you fly solo.
  • Walk the room. Common issues: API key not set, wrong model name, tool results format.
  • Celebrate first working queries loudly — momentum matters.
  • If stuck >3 min, ask a neighbor. If neighbor stuck too, raise hand.
Demo Cues
SWITCH TO: Google Colab OPEN: workshop-timer.html
Ask the Class
What's the most surprising answer your agent gave?
Transition
Your agent now knows your business. Grab a coffee — next we break it.
4
Break
6:25 – 6:35 PM 10 min
--:--
Key Talking Points
  • Stretch break. Remind: "Your agent now knows YOUR business. Next: what happens when the data is wrong?"
  • Check: who got all 3 phases done? Partial credit is fine.
Transition
Welcome back. Your agent works. Now let's see what happens when the data is WRONG.
5
RAG Failure Modes & Trust
6:35 – 7:05 PM 30 min
--:--
Key Talking Points
  • Air Canada story is REAL. Court ruled them liable. $812. This is not theoretical.
  • Demo each failure mode live — outdated doc, contradicting docs, unanswerable question.
  • Key message: "I don't know" is a FEATURE, not a bug. Best agents know their limits.
  • Link to failure-modes-demo.html for interactive exploration.
Demo Cues
OPEN: failure-modes-demo.html
Ask the Class
Should a company be liable for what its AI chatbot says? Why?
Transition
So how do real companies handle all this? Let's look at the real world.
6
RAG in the Real World
7:05 – 7:20 PM 15 min
--:--
Key Talking Points
  • Most companies buy (Glean, Notion AI). You learn to build so you can evaluate.
  • Ask: "Who has used a company knowledge base at work? Was it good?" Get stories.
  • Build vs buy framework: <100 docs = long context, <10K = simple RAG, >10K = vector DB.
Ask the Class
Who has used a company knowledge base at work? Was it actually good?
Transition
You now know what companies build AND buy. Let's wrap up.
7
Wrap-Up + Homework
7:20 – 7:30 PM 10 min
--:--
Key Talking Points
  • Stack recap: Tools → APIs → Memory → ReAct → Knowledge. Next: Multi-Agent.
  • Homework preview: Build YOUR OWN knowledge agent. Domain you actually care about.
  • Tease Session 8: "One agent is useful. A TEAM of agents is powerful."
Demo Cues
OPEN: homework-guide.html
Ask the Class
What domain would YOU build a knowledge agent for?