Making LLM calls · Chapter 4

Call an LLM from Python in 15 lines with OpenRouter, and learn what messages, roles and temperature do.

The big idea

Calling an LLM is just sending a web request. You send your messages to a web address, and a reply comes back as JSON. That is all an "AI feature" is under the hood.

Your first call

Save this as ask.py:

python

import os
import requests

API_KEY = os.environ["OPENROUTER_API_KEY"]

response = requests.post(
    "https://openrouter.ai/api/v1/chat/completions",
    headers={"Authorization": f"Bearer {API_KEY}"},
    json={
        "model": "liquid/lfm-2.5-2.6b:free",
        "messages": [
            {"role": "system", "content": "You are a friendly teacher. Answer in two sentences."},
            {"role": "user", "content": "What is a knowledge graph?"},
        ],
        "temperature": 0.7,
    },
    timeout=60,
)
response.raise_for_status()
print(response.json()["choices"][0]["message"]["content"])

Run it:

bash

pip install requests
export OPENROUTER_API_KEY=your-key-here
python ask.py

What each part does

Part	What it means
The URL	OpenRouter's chat address. Every model uses the same one.
`Authorization` header	Your API key, so OpenRouter knows who is asking.
`model`	Which LLM answers. Change this one word to switch models.
`messages`	The conversation so far: rules (`system`) and question (`user`).
`temperature`	Creativity. 0 gives nearly the same answer every time; around 1 gives more variety.
`choices[0].message.content`	Where the reply text sits in the JSON that comes back.

When it fails: a 401 error means the key is wrong. A 429 error means too many requests; free models have limits, so wait a minute and try again. If the model is not found, it has probably been retired; pick a current free model at openrouter.ai/models (see the disclaimer in Chapter 3).

Where it lives in GraphRAG Chats

generate_answer in backend/src/backend/openrouter_client.py does exactly this. It uses httpx instead of requests so the server can do other work while it waits, but the URL, header, model, messages and reply path are the same.

Practice: make it your own

Run ask.py and read the answer.
Set temperature to 0 and run it three times. Then set it to 1.2 and run it three times. Notice the difference.
Change the system message to "Answer like a pirate." and run it again.

You are done when you can change the model's behaviour just by editing the messages.

Check yourself

Which part of the request picks the model?
You get a 429 error. What happened?
What does a low temperature do?