Recap

Foundations + chatlas basics

LLMs, Agents, & Context

• LLMs are jagged — great at some things, bad at others (e.g. counting).

• LLMs are really good at coding — so enable them to, and verify the result.

• Agent = LLM + tools + loop.

• Correctness and usefulness can (further) improve with the right context, constraints, and tooling.

chatlas, so far

• One interface (Chat{Provider}) to many LLMs — swap providers, keep your code.

• .chat() echoes responses; .stream() yields the chunks.

• Chat objects are stateful — they remember the conversation (get_turns(), set_turns()).

• A system prompt shapes every response, invisibly, without RAG or tool calling.

Quick check ✅

• What’s the difference between a system and user prompt?

• What method yields a chatbot in 1 line of code?

• Why might you want to use .stream() over .chat()?

Up next

🔧 Tool calling — give the LLM real capabilities

📦 Structured data extraction

💬 shinychat — turn a chat into an app