Interactive Lab · How LLMs stay honest

Why does an LLM make things up, and how does giving it a search box fix that?

An LLM only knows what was baked into its weights. Ask it something it never saw, your company's refund policy, a niche recent fact, and it will answer confidently and wrongly. RAG bolts a search step in front of it: retrieve the real documents, hand them over, and make it quote the truth. Flip retrieval on and off below and watch the same question flip from fiction to fact.

🎬 Watch the 45-second explainer, then try it yourself below ↓

▶ The playground, pick a question, flip retrieval, hit Ask

Pick a question

🔎 Use retrieval (RAG): ON ☠️ Poison the knowledge base: OFF

❓Your question

→

🔎Retrieversearches KB

→

🤖LLMgenerates

→

✅Answer

🧭 Where this fits

RAG is the architecture of virtually every enterprise AI assistant shipped since 2023, customer-support bots grounded in help docs, legal/medical copilots with citations, internal-knowledge chat. The poisoned-KB failure you just triggered is the field's core operational lesson: RAG moves the failure from the model's memory to YOUR retrieval pipeline, which is why retrieval evals (see the RAG-eval lab) are a job.

The code that's actually running

🔮 Predict, then test

Pick "BitBootcamp's refund policy" and turn retrieval OFF, will the LLM admit it doesn't know, or invent a plausible-sounding policy? Now flip it ON. Finally, keep RAG on but flip ☠️ Poison the KB: the answer is still grounded and confident, so why is it wrong? What does that tell you about where a RAG system actually fails?

The takeaway

Your new mental model of RAG

No retrieval → the model answers from memory alone. For anything private, recent, or niche it will hallucinate, fluent, confident, and wrong.
Retrieval ON → it first fetches the real chunks, then answers grounded in them and can cite its sources. Same model, but now anchored to truth.
Garbage in, garbage out → RAG is only as good as its sources. Poison the KB and the answer is grounded, cited… and still false. Retrieval quality is the product.
You didn't read a paper, you watched the same question flip from fiction to fact and back. That's the instinct that tells you when to add retrieval and where to guard it.

📖 Lecture notes Continue the journey → Try another lab