TokenTrim

TokenTrim intelligently trims OpenAI messages to fit within a model's token limit (shortening a message by removing characters from the middle), making it easy to avoid exceeding the maximum token count.

It's best suited for use directly in OpenAI API calls:

import tokentrim as tt

model = "gpt-4"

response = openai.ChatCompletion.create(
  model=model,
  messages=tt.trim(messages, model) # Trims old messages to fit under model's max token count
)

TokenTrim's behavior is based on OpenAI's own best practices.

Installation

Use the package manager pip to install TokenTrim:

pip install tokentrim

Usage

The primary function in the TokenTrim library is trim(). This function receives a list of messages and a model name, and it returns a trimmed list of messages that should be within the model's token limit.

import tokentrim as tt

# Force a system_message to be prepended to your messages list. This will not be trimmed.
system_message = "You are a helpful assistant."

response = openai.ChatCompletion.create(
  model=model,
  messages=tt.trim(messages, model, system_message=system_message)
)

Parameters

• messages : A list of message objects to be trimmed. Each message is a dictionary with 'role' and 'content'.
• model : The OpenAI model being used (e.g., 'gpt-4', 'gpt-4-32k'). This determines the token limit.
• system_message (optional): A system message to preserve at the start of the conversation.
• trim_ratio (optional): Target ratio of tokens to use after trimming. Default is 0.75, meaning it will trim messages so they use about 75% of the model's token limit.
• return_response_tokens (optional): If set to True, the function also returns the number of tokens left available for the response after trimming.

Return Value

By default, trim() returns the trimmed list of messages. If return_response_tokens is set to True, it returns a tuple where the first element is the trimmed list of messages, and the second element is the number of tokens left available for the response.

License

This project is licensed under the terms of the MIT license.