Message Roles

Now that you've successfully sent a simple prompt to your language model, let's take it a step further — with message roles!

Modern LMs understand context better when you structure inputs like a real conversation. That’s where message roles come in. You can simulate multi-turn chats, set instructions, or give memory to the model through a structured message format.

Message Format

Here’s how it works:

[
    (<role>, [<content>, <content>]),
    (<role>, [<content>, <content>]),
    ...
]

Format breakdown

1. It's a list of messages

2. Each message is a tuple: (role, [list_of_content])

3. Valid roles are:

   1. "system" – for instructions

   2. "user" – for user prompts

   3. "assistant" – for assistant replies

4. Each content list can hold multiple entries when we have multimodal inputs, but we'll use simple text only for now.

The available prompt roles are also defined as an enumeration in the schema. You can import them using: from gllm_inference.schema import PromptRole

Example 1: Passing a system message

import asyncio
from gllm_inference.lm_invoker import OpenAILMInvoker

prompt = [
    ("system", ["Talk like a pirate."]),
    ("user", ["Hi, there! How are you doing?"])
]

lm_invoker = OpenAILMInvoker("gpt-4.1-nano")
response = asyncio.run(lm_invoker.invoke(prompt))
print(f"Response: {response}")

Output:

Response: Ahoy, matey! I be doin' well, savvy? How be ye farin' on this fine day?

Example 2: Simulating a multi-turn conversation

prompt = [
    ("system",, ["Reply with concise answers!"]),
    ("user",, ["I'm craving for some fried chicken!"]),
    ("assistant", ["That sounds good!"]),
    ("user", ["Do you think it's healthy?"]),
]

lm_invoker = OpenAILMInvoker("gpt-4.1-nano")
response = asyncio.run(lm_invoker.invoke(prompt))
print(f"Response: {response}")

Output:

Response: Fried chicken can be enjoyed occasionally, but it's typically high in calories, fat, and sodium, so it's not considered very healthy if eaten frequently.

💡 Tip: Add previous user and assistant messages to simulate long conversations with memory!