Open WebUI filter to prevent thinking re-injection

filter.py

"""
title: Strip Thinking Tags
author: Sam Morris <sam@robots.org.uk>
author_url: https://gist.github.com/yrro/b0f2765ea55ae3414e06b319dd07ae8e
version: 0.1
"""

# Open WebUI re-injects thinking into the assistent messages in the conversation
# history which is contrary to Gemma 4's requirements. 
# It's a workaround for <https://github.com/open-webui/open-webui/issues/23339>.
#
# There is a setting Admin Panel → Settings → Connection → [your llama.cpp connection] → Provider;
# Set this to "llama.cpp" and thinking is sent in a separate 'reasoning_content' field instead
# of being rendered into the message stream. Maybe that makes this filter obsolete?

from pydantic import BaseModel, Field
from typing import Optional


class Filter:

    def inlet(self, body: dict, __user__: Optional[dict] = None) -> dict:
        new_messages = list(
            self.strip_thinking(message) for message in body["messages"]
        )
        return body | {"messages": new_messages}

    def strip_thinking(self, message: dict):
        new_message = {}
        if message["role"] == "assistant" and message["content"].startswith("<think>"):
            thinking, sep, new_content = message["content"].partition("</think>")
            new_message["content"] = new_content.lstrip()
        #print(f"new_message: {new_message!r}")
        return message | new_message

Thank you very very much for this! It fixed my only gripe with Gemma 4.