Generate responses with the ChatCompletions API

Completed

The OpenAI ChatCompletions API is commonly used across generative AI models and platforms. Although the Responses API is recommended for new project development, it's likely that you'll encounter scenarios where the ChatCompletions API is useful for code maintenance of cross-platform compatibility.

Submitting a prompt

The ChatCompletions API uses collections of message objects in JSON format to encapsulate prompts:

completion = openai_client.chat.completions.create(
    model="gpt-4o",  # Your model deployment name
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "When was Microsoft founded?"}
    ]
)

print(completion.choices[0].message.content)

Retaining conversational context

Unlike the Responses API, the ChatCompletins API doesn't provide a stateful response tracking feature. To retain conversational context, you must write code to manually track previous prompts and responses.

# Initial messages
conversation_messages=[
    {
        "role": "system",
        "content": "You are a helpful AI assistant that answers questions and provides information."
    }
]

# Add the first user message
conversation_messages.append(
    {"role": "user",
    "content": "When was Microsoft founded?"}
)

# Get a completion
completion = openai_client.chat.completions.create(
    model="gpt-4o",
    messages=conversation_messages
)
assistant_message = completion.choices[0].message.content
print("Assistant:", assistant_text)

# Append the response to the conversation
conversation_messages.append(
    {"role": "assistant", "content": assistant_text}
)

# Add the next user message
conversation_messages.append(
    {"role": "user",
    "content": "Who founded it?"}
)

# Get a completion
completion = openai_client.chat.completions.create(
    model="gpt-4o",
    messages=conversation_messages
)
assistant_message = completion.choices[0].message.content
print("Assistant:", assistant_text)

# and so on...

In a real application, the conversation is likely to be implemented in a loop; like this:

# Initial messages
conversation_messages=[
    {
        "role": "system",
        "content": "You are a helpful AI assistant that answers questions and provides information."
    }
]

# Loop until the user wants to quit
print("Assistant: Enter a prompt (or type 'quit' to exit)")
while True:
    input_text = input('\nYou: ')
    if input_text.lower() == "quit":
        print("Assistant: Goodbye!")
        break

    # Add the user message
    conversation_messages.append(
        {"role": "user",
        "content": input_text}
    )

    # Get a completion
    completion = openai_client.chat.completions.create(
        model="gpt-4o",
        messages=conversation_messages
    )
    assistant_message = completion.choices[0].message.content
    print("\nAssistant:", assistant_message)
    
    # Append the response to the conversation
    conversation_messages.append(
        {"role": "assistant", "content": assistant_message}
    )

The output from this example looks similar to this:

Assistant: Enter a prompt (or type 'quit' to exit)

You: When was Microsoft founded?

Assistant: Microsoft was founded on April 4, 1975 in Albuquerque, New Mexico, USA.

You: Who founded it?

Assistant: Microsoft was founded by Bill Gates and Paul Allen.

You: quit

Assistant: Goodbye!

Each new user prompt and completion is added to the conversation, and the entire conversation history is submitted in each turn.

While not as fully featured as the Responses API, the ChatCompletions API is well established in the generative AI model ecosystem, so it's useful to be familiar with it.