TinkerModule 1 · 7 chapters
Retrieval-Augmented Generation (RAG)
How AI assistants answer from your company’s documents, why they go wrong, and how to fix them.
- Context windows
- Embeddings
- Chunking
- Vector search
- Reranking
- Agentic RAG
- GraphRAG
Start RAG →
See how tech works by playing with it.
Hand-drawn stories you play: build the thing, watch it break, fix it.
Free · No sign-up · Nothing to install · Works on your phone
Why Tinker
Understanding comes from poking at things
Play, don’t just read
Every page has a live picture. Each step asks you to change one thing in it, and you watch what happens.
The real thing underneath
Nothing is faked for the animation. In RAG, real retrieval maths runs in your browser over real embeddings.
A story with stakes
You join a team with a problem to solve. Each chapter starts with something going wrong, and you fix it.
Modules
Pick a concept
Beginner
Module 6 · 7 chapters
Tokens & embeddings
How text becomes the numbers a model reads: tokens to count and pay for, and embeddings that carry meaning.
- Tokens & ids
- Byte pair encoding
- Counting & cost
- Tokenizers compared
- Embeddings
- Meaning by neighbours
- Token budgets
Start Tokens & embeddings →
Intermediate
Module 2 · 7 chapters
Tool calling & the Model Context Protocol (MCP)
How AI assistants go from answering to acting: tools, the Model Context Protocol, and keeping a human in charge.
- Tool definitions
- JSON Schema
- Multi-step calls
- MCP hosts & servers
- Tools, resources, prompts
- Approval
- Prompt injection
Start Tool calling & MCP →
Module 4 · 7 chapters
Evals & LLM testing
How a team knows the assistant got better or worse: test sets, graders and regressions.
- Test sets
- Regressions
- Code graders
- Model judges
- Judging the judge
- Grading the path
- pass@k & pass^k
- Release gates
Start Evals & LLM testing →
Module 5 · 7 chapters
Prompt injection & guardrails
Text that tries to take over the assistant, and the defences that hold it back: filters, output checks and the lethal trifecta.
- Direct & indirect injection
- Prompt rules
- Input filters
- Output checks
- The lethal trifecta
- Defence in depth
- Red teaming
Start Prompt injection & guardrails →
Module 8 · 7 chapters
Context engineering & memory
What goes into the model’s window each turn, in what order, and what it should remember between conversations.
- The context window
- Lost in the middle
- Just-in-time retrieval
- Prompt caching
- Clearing tool results
- Memory files
- Context policy
Start Context engineering & memory →
Advanced
Module 3 · 7 chapters
AI agents
How an assistant goes from one request to a whole case: the loop, plans, budgets, memory and helpers.
- Workflow or agent
- The agent loop
- Plans & approval
- Step & token budgets
- Stuck loops
- Memory & compaction
- Sub-agents
Start AI agents →
Module 7 · 7 chapters
Transformer architecture
What happens inside the model between a prompt and its next word: vectors, attention, heads and layers, and how the next token is picked.
- Next-token prediction
- Embeddings & position
- Attention
- The causal mask
- Heads & layers
- Temperature & top-p
- The KV cache
Start Transformer architecture →
Module 9 · 7 chapters
Fine-tuning vs prompting
When to train the model, when to prompt it, and when to give it documents.
- Prompt, documents or weights
- Training & loss
- LoRA adapters
- Facts vs style
- Forgetting & cost
- Evals before training
- Choosing an approach
Start Fine-tuning vs prompting →
Want a concept explained here? Propose a module.
How it works
Every module plays the same way
1 Meet the problem
A short story sets up what is going wrong and why it matters.
2 Change one thing
The step asks for one action in the picture: add a block, move a slider, pick an answer.
3 Watch what changes
The picture, the meters and the answer update at once, so cause and effect are visible.
4 Fix it yourself
Puzzles hand you a broken setup. Hints appear when you’re stuck, and “Show me” is always there.