1. Your first model call

Time: 20 minutesEdit: code/01-first-conversation/starter/chat.py

Every AI app, agent and framework sits on top of one thing: an HTTP request to a model. In this module you send that request yourself.

The chat completions API

Most model servers speak the same protocol, first made popular by OpenAI and now supported by Ollama, LM Studio, vLLM, llama.cpp and nearly every cloud provider. You POST a list of messages to /chat/completions and get the next message back.

Request

{
  "model": "qwen3:4b-instruct",
  "messages": [
    {"role": "system", "content": "You are a friendly conference assistant."},
    {"role": "user", "content": "What's a good first AI talk to attend?"}
  ],
  "temperature": 0.7
}

Response (trimmed)

{
  "choices": [
    {"message": {"role": "assistant", "content": "Try a beginner keynote..."}}
  ]
}

There are three roles:

• system: instructions from you, the developer. The user doesn’t see them.
• user: what the person typed.
• assistant: what the model replied.

Exercise: make the call

Open code/01-first-conversation/starter/chat.py. The chat loop is written for you. Your job is the ask() function and two lines in the loop.

TODO 1: send the request with httpx and return the reply text.

01-first-conversation/starter/chat.py

def ask(messages):
    response = httpx.post(
        f"{config.BASE_URL}/chat/completions",
        headers={"Authorization": f"Bearer {config.API_KEY}"},
        json={
            "model": config.MODEL,
            "messages": messages,
            "temperature": 0.7,
        },
        timeout=config.TIMEOUT,
    )
    response.raise_for_status()
    return response.json()["choices"][0]["message"]["content"]

config reads your .env file, so the same code works with any provider.

Run it:

python 01-first-conversation/starter/chat.py

Ask it something. Then ask a follow-up like “Can you say that more briefly?” What happens?

Models have no memory

The model forgot what it just said. That’s not a bug. Models are stateless. Every request is a fresh start, and the only memory is the list of messages you send. Chat apps feel like they remember because they resend the whole conversation every time.

TODO 2 and 3: append the user’s message before calling ask(), and append the model’s reply after.

        messages.append({"role": "user", "content": user_input})
        reply = ask(messages)
        messages.append({"role": "assistant", "content": reply})

Run it again and try the follow-up. Now it remembers.

Why this matters later

Because the conversation is just a list you send, anything you put in it becomes part of what the model reads, including text from tools, documents and web pages. Keep that in mind for Module 5.

Try this

• Change the system prompt to make the assistant answer only in pirate speak. Then ask it to stop. Who wins, the system prompt or the user?
• Set temperature to 0 and ask the same question three times. Then try 1.5. What changes?
• Ask it which sessions are on Day 2 of the Lumen AI Summit. It will probably make some up. Hold that thought.

Checkpoint. You can call a model with plain HTTP and keep a conversation going. You also saw its biggest weakness: it confidently answers questions it has no data for.

Stuck? Run the finished version: python 01-first-conversation/solution/chat.py