token_counter.py

# This is a work in progress. There are still bugs. Once it is production-ready this will become a full repo.
import os

def count_tokens(text, model_name="gpt-3.5-turbo", debug=False):
    """
    Count the number of tokens in a given text string without using the OpenAI API.
    
    This function tries three methods in the following order:
    1. tiktoken (preferred): Accurate token counting similar to the OpenAI API.
    2. nltk: Token counting using the Natural Language Toolkit library.
    3. split: Simple whitespace-based token counting as a fallback.
    
    Usage:
    ------
    text = "Your text here"
    result = count_tokens(text, model_name="gpt-3.5-turbo", debug=True)
    print(result)

    Required libraries:
    -------------------
    - tiktoken: Install with 'pip install tiktoken'
    - nltk: Install with 'pip install nltk'

    Parameters:
    -----------
    text : str
        The text string for which you want to count tokens.
    model_name : str, optional
        The OpenAI model for which you want to count tokens (default: "gpt-3.5-turbo").
    debug : bool, optional
        Set to True to print error messages (default: False).

    Returns:
    --------
    result : dict
        A dictionary containing the number of tokens and the method used for counting.
    """

    # Try using tiktoken
    try:
        import tiktoken
        encoding = tiktoken.encoding_for_model(model_name)
        num_tokens = len(encoding.encode(text))
        result = {"n_tokens": num_tokens, "method": "tiktoken"}
        return result
    except Exception as e:
        if debug:
            print(f"Error using tiktoken: {e}")
        pass

    # Try using nltk
    try:
        import nltk
        nltk.download("punkt")
        tokens = nltk.word_tokenize(text)
        result = {"n_tokens": len(tokens), "method": "nltk"}
        return result
    except Exception as e:
        if debug:
            print(f"Error using nltk: {e}")
        pass

    # If nltk and tiktoken fail, use a simple split-based method
    tokens = text.split()
    result = {"n_tokens": len(tokens), "method": "split"}
    return result

class TokenBuffer:
    def __init__(self, max_tokens=2048):
        self.max_tokens = max_tokens
        self.buffer = ""
        self.token_lengths = []
        self.token_count = 0

    def update(self, text, model_name="gpt-3.5-turbo", debug=False):
        new_tokens = count_tokens(text, model_name=model_name, debug=debug)["n_tokens"]
        self.token_count += new_tokens
        self.buffer += text
        self.token_lengths.append(new_tokens)

        while self.token_count > self.max_tokens:
            removed_tokens = self.token_lengths.pop(0)
            self.token_count -= removed_tokens
            self.buffer = self.buffer.split(" ", removed_tokens)[-1]

    def get_buffer(self):
        return self.buffer

To use the TokenBuffer class, you can create an instance with an optional max_tokens argument and then call the update() method to add text. The get_buffer() method returns the current buffer.

tb = TokenBuffer(max_tokens=2048)
tb.update("Some text to add to the buffer.")
print(tb.get_buffer())

The update() method ensures that the buffer always contains at most max_tokens by removing characters from the beginning of the buffer when the limit is exceeded.

Using a token buffer like the one implemented in the TokenBuffer class is useful for working with OpenAI's API for several reasons:

Token Limit: OpenAI's models have a maximum token limit per API call (e.g., 4096 tokens for gpt-3.5-turbo). By using a token buffer, you can manage and control the text input to ensure it stays within the allowed token limit, preventing errors when making API calls.

Cost Control: OpenAI's API pricing is based on the number of tokens processed. By maintaining a token buffer, you can keep track of the tokens used, helping you manage costs more effectively and avoid exceeding your budget.

Text Truncation: When dealing with long text inputs or a stream of text, using a token buffer can help you truncate or remove less relevant text while preserving the most recent or relevant information. This is particularly useful when working with conversational AI applications, where the latest information might be more important for generating appropriate responses.

Rate Limiting: OpenAI's API has rate limits based on tokens processed per minute. A token buffer helps you stay within these rate limits, ensuring that your application can operate smoothly without encountering rate limit errors.

Overall, using a token buffer like the TokenBuffer class is a practical way to manage tokens when working with OpenAI's API, helping you stay within token limits, control costs, and manage text inputs more effectively.

Example usage for TokenBuffer:

from token_counter import TokenBuffer

# Initialize a TokenBuffer with a maximum token count of 30
buffer = TokenBuffer(max_tokens=30)

# Add a sentence to the buffer
buffer.update("Hello, how are you doing?")
print(buffer.get_buffer())
print("Token count:", buffer.token_count)

# Add another sentence to the buffer
buffer.update("I'm doing well, thank you!")
print(buffer.get_buffer())
print("Token count:", buffer.token_count)

# Add a longer sentence to the buffer
buffer.update("I've been working on a project and making great progress.")
print(buffer.get_buffer())
print("Token count:", buffer.token_count)

# Add one more sentence to the buffer
buffer.update("That's great to hear, keep up the good work!")
print(buffer.get_buffer())
print("Token count:", buffer.token_count)

Output (YMMV):

Hello, how are you doing?
Token count: 6
Hello, how are you doing?I'm doing well, thank you!
Token count: 11
Hello, how are you doing?I'm doing well, thank you!I've been working on a project and making great progress.
Token count: 24
I'm doing well, thank you!I've been working on a project and making great progress.That's great to hear, keep up the good work!
Token count: 30

Tiktoken's github repo: https://github.com/openai/tiktoken

NLTK's github repo: https://github.com/nltk/nltk