Knowledge Core ships with a live AI assistant. It’s not a generic chatbot — it has read your documentation and course content and can answer specific questions about them.
How it works: RAG in plain language
The assistant uses a technique called Retrieval-Augmented Generation (RAG). Instead of relying only on what a language model was trained on, it first retrieves relevant pieces of your content, then uses those as context for generating an answer.
Here’s the flow for every question:
Your question
│
▼
1. Embedding
Convert the question into a vector (a list of ~1000 numbers
that captures its semantic meaning).
│
▼
2. Similarity search (Cloudflare Vectorize)
Find the 5 content chunks whose vectors are most similar
to the question vector.
│
▼
3. Context injection
Prepend the retrieved chunks to the prompt as context.
│
▼
4. LLM generation (Llama 3 on Workers AI)
Generate an answer based on the retrieved context.
│
▼
Streamed answer in the chat widgetA language model’s training data has a cutoff date and knows nothing about your specific project. RAG bridges this gap: the model gets the facts from your docs, and only needs to handle reasoning and language.
What the AI knows
The assistant has indexed all content from:
- Documentation — every page under
apps/docs/src/content/docs/ - Course lessons — every lesson under
apps/courses/src/content/lessons/
Content is split into ~800-character chunks before embedding, so the AI can retrieve precise sub-sections rather than entire pages.
Good questions to try
The AI works best with specific, content-oriented questions:
| Good question | Why it works |
|---|---|
| ”How do I install Knowledge Core?” | Maps directly to the Installation guide |
| ”What UI components are available?” | Matches the Components overview page |
| ”How do I create a new course lesson?” | Targets the Creating Content guide |
| ”What is Cloudflare Vectorize?” | Covered in the AI Chat Integration guide |
| ”How does the quiz component work?” | Explained in this very course |
The tech stack behind it
| Component | Technology |
|---|---|
| Embedding model | @cf/baai/bge-large-en-v1.5 (1024 dimensions) |
| Vector database | Cloudflare Vectorize |
| Language model | @cf/meta/llama-3-8b-instruct |
| Runtime | Cloudflare Workers (edge, ~0ms cold start) |
| Transport | Server-Sent Events (streaming) |
Everything runs on Cloudflare’s edge network — no dedicated server, no GPU to manage, pay-per-request pricing.
Keeping the index fresh
The vector index is built once by running pnpm run ingest. It does not automatically update when you edit content. After adding new pages or making significant changes, re-run the ingest command to update the index.
If you ask the AI about something you just added and it says it doesn’t know — the index probably hasn’t been updated yet. Run pnpm run ingest and the new content will be searchable within seconds.
Quiz
1. What does RAG stand for?
2. How many content chunks does the chat worker retrieve per question?
3. What do you need to do after adding new documentation pages?