Please share Gemini tokenize information

[hyunjunseo]

hello.

Thank you so much for the recently announced Gemini-Pro API availability.

We use a lot of APIs, and in the case of OpenAI, we expose cl100k_base so that users can pre-calculate the number of tokens and avoid API errors.

But Gemini-Pro doesn't know anything about the token, so it has to rely on the character count. 😅

Is it possible to share token information like OpenAI's tiktoken?

Thank you for creating good model. 😃

[angel_perez · Accepted solution]

With the Vertex AI SDK (python)  -- we compute the number of tokens (and characters)  like:

Token Count Docs 
It looks like this:

from vertexai.preview.generative_models import GenerativeModel

gemini_pro_model = GenerativeModel("gemini-pro")

print(gemini_pro_model.count_tokens("why is sky blue?"))
I do miss having a local implementation that we could use like tiktoken, and it will be greater if it exists (I am not aware of it) ---
I hope it helps. 

[angel_perez · Accepted solution]

With the Vertex AI SDK (python)  -- we compute the number of tokens (and characters)  like:

Token Count Docs 
It looks like this:

from vertexai.preview.generative_models import GenerativeModel

gemini_pro_model = GenerativeModel("gemini-pro")

print(gemini_pro_model.count_tokens("why is sky blue?"))
I do miss having a local implementation that we could use like tiktoken, and it will be greater if it exists (I am not aware of it) ---
I hope it helps. 

[PicardParis · Accepted solution]

Hey!

• You can now count tokens locally with the Vertex AI SDK for Python (starting with version 1.57.0).
• Check out this Medium article for details: Counting Gemini text tokens locally.

[hyunjunseo]

Thank you so much

Now i can request just once not twice anymore ^^

[myudak]

how to do this with nodejs?