3 The OpenAI Python library

This chapter motivates moving from the web interface to programmatic access so you can scale beyond single, manual prompts, embed GPT calls inside larger workflows, and combine language models with other tools. It introduces OpenAI’s Python library, notes that similar libraries exist for other providers, and walks through getting your environment ready: having a recent Python (3.9+), ensuring pip works, installing a pinned library version, and obtaining an API key. It emphasizes safe key handling—prefer environment variables and avoid hard-coding—and reassures that cloud notebooks can be used if local setup is troublesome.

Once set up, you install the specified library version and create a client to make requests. You can list available models and select by model ID (with a brief note that some IDs may be private fine-tuned models). The core capability used throughout is chat completion: you supply a list of messages, each with a role (user, assistant, or optional system) and text content, and the model returns a reply. The response contains choices (completions), termination info, and usage statistics such as prompt, completion, and total tokens. Because costs scale with tokens and vary by model, the chapter urges choosing models and prompt sizes judiciously.

The chapter then shows how to steer outputs with configuration parameters. You can bound or end generations with max_tokens and stop sequences; shape repetitiveness with presence_penalty and frequency_penalty; bias or forbid specific tokens with logit_bias (token IDs obtainable via a tokenizer tool); and control randomness via temperature or top_p (tuning only one at a time is recommended). When variability is useful, you can request multiple candidates with n and pick the best downstream. Practical guidance includes setting sensible token limits to cap cost, using stop conditions when appropriate, encouraging or discouraging repetition depending on the task, and selecting an appropriate model (the examples use a multimodal model while focusing on text).

The Google Colab platform can be used to run the following examples.

Managing secret keys for accessing the OpenAI API.

• You can use OpenAI’s language models via a Python API.
• Other providers offer similar libraries for accessing their models.
• To use OpenAI’s library, create a client object.
• You can use OpenAI’s models to complete chats.
• Chats to complete are specified as a list of messages.
• Each chat message is characterized by content and a role.
• Roles can be one of user, assistant, and system.
• Obtain chat completions via the chat.completions.create function.
• You can configure models using various parameters.
• The max_tokens parameter limits the number of tokens generated.
• Alternatively, stop allows defining phrases that stop text generation.
• You can penalize or encourage specific tokens via logit_bias.
• The presence_penalty parameter penalizes repetitive output.
• The frequency_penalty parameter penalizes repetitive output.
• The temperature parameter chooses the degree of randomization.
• Parameter top_p determines the number of output tokens considered.
• Parameter n chooses the number of generated completions.

FAQ

import openai
client = openai.OpenAI()
models = client.models.list()
for m in models.data:
    print(m)

[{
  "role": "user",
  "content": "Tell me a story!"
}]

import openai
client = openai.OpenAI()

result = client.chat.completions.create(
    model="GPT-4o",
    messages=[{"role":"user", "content":"Tell me a story!"}]
)

# First reply text:
print(result.choices[0].message.content)