GhostCoT

A decorator to enable Chain-of-Thought (CoT) reasoning for non-thinking LLMs (e.g. gpt-4o-mini and deepseek-chat).

Installation

pip install ghostcot

Enable Non-Thinking LLMs to Think

Question: 2 + π ≈ ?

Using ghostcot, you can enable non-thinking LLMs (e.g. gpt-4o-mini) to reason and get:

Thinking: To estimate 2 + π, we need to know the approximate value of π. The value of π is approximately 3.14. Therefore: 2 + 3.14 = 5.14

✨ Final Answer: 5.14

You can run a demo with the following code:

import ghostcot
api_key = "your-api-key" # OpenAI/DeepSeek/Other
base_url = None # None for OpenAI, otherwise for other LLMs
ghostcot.run_demo_cot_stream("What is 2 + π?", model_name="gpt-4o-mini", api_key=api_key, base_url=base_url)

How to Use

The only change you need to make is to decorate your chat function of OpenAI API with @ghostcot.enable_cot(), and all the remaining code is the same as before.

# Apply GhostCoT decorator
@ghostcot.enable_cot()
def chat(messages, **kwargs):
    return client.chat.completions.create(
        model=model_name,
        messages=messages,
        **kwargs
    )

To distinguish between reasoning content and final answer:

• Non-streaming: response.choices[0].message.reasoning_content and response.choices[0].message.content
• Streaming: chunk.choices[0].delta.reasoning_content and chunk.choices[0].delta.content

Demos:

• Non-streaming demo
• Streaming demo