Recap

Tools, structured data, & chatbots

🔧 Tool calling

• Agent = LLM + tools + loop

• A tool can be a typed, documented function — the model decides when to call it.

• register_tool() turns a Python function into something the model can call.

• Letting an LLM execute arbitrary code is powerful but risky — sandbox it, run it remotely, or restrict its scope.

• Tools can be extended beyond your own process using MCP (over HTTP or STDIO).

📦 Structured data extraction

• Extract structured data from text/images/PDFs

• Transform LLM responses into data your code can branch on, not just prose.
  • Useful for a lot of stuff beyond chat apps!

• .chat_structured() + a Pydantic model extracts typed data from text, images, or PDFs.

• str | None fields handle data that’s sometimes missing.

💬 shinychat

• Turn any chatlas instance into a web app in just a few lines of code.

• client= gives you history, tool display, streaming, cancel, attachments, and error handling for free.

• Return a ToolResultDisplay to show the user a polished view while the model sees raw data.

• Customize with tool displays, slash commands, and more

Quick check 🙋🏻‍♀️

• What two things does a Python function need to become a tool?

• Name one way to keep LLM-driven code execution safe.

• What does client= give you for free in shinychat?

Up next

📊 querychat — the capstone: self-service analytics

✔️ Measuring correctness — evals

🧪 Production tooling — observability