OpenWebUI + Weave (by Weights & Biases) Input/Completion Logging

OpenWebUI function to log your inputs and completions to Weave (by Weights & Biases) for LLMOps / observability.

Implementation notes:

• OpenWebUI (open-source LLM web UI).
• Weave (LLMOps, logging, and observability).
  • Uses manual call tracking to log inputs and completions to your Weave project.
  • Triggers the call tracking using OpenWebUI's filter functions before (inlet) and after (outlet) an LLM execution.

Installation

1. Setup your Weave project

• Go to https://wandb.ai/home (sign up if you don't have an account), create a new project
  • Remember your project's name, you'll need it in the next step
• Grab your API key from https://wandb.ai/settings

2. Run your OpenWebUI interface and install the function

• Run your OpenWebUI Docker container (read more), simplest way is to use the following command:

docker run -d -p 3000:8080 -e OPENAI_API_KEY=your_secret_key -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:main

• Visit the OpenWebUI Functions page at: http://localhost:3000/admin/functions (assuming you're local and used the default port)
• Open filter.py from this repo, copy content, paste the following code into the code field and click "Install".
• Select the gear icon next to the function and set the wandb_api_key and wandb_project_name you copied in the first step.
  • Your API key can be found at: https://wandb.ai/settings.
  • Make sure you use team/project_name format for wandb_project_name.
• Enable the function and set as "Global" to enable logging for all chat instances.

3. View your traces

• Visit https://wandb.ai/home and navigate to your project to see your logs of inputs and completions.

Note

• Set the priority of the function to the highest value to have it fire (log) after any other transformations you make to your context.
• Token counting is currently supported for OpenAI models using tiktoken. Non-OpenAI models will default to gpt-4os token count (which will probably be close to actual token usage).