Local Vision Bridge (OpenWebUI Function)

Give your text-only LLMs the ability to "see" using a secondary local Vision model.

This OpenWebUI Function intercepts image uploads, sends them to a local Vision-Language Model (like Qwen2.5-VL running on llama.cpp), and seamlessly injects detailed text descriptions into the chat context.

This allows you to use massive, high-intelligence text-only models (like large MOE) while still enjoying multi-modal capabilities via a smaller, faster dedicated vision model.

Features

• Zero-Latency Caching: Hashes images so you only pay the "GPU tax" once. Subsequent turns in the chat are instant.
• History Aware: Scans the full conversation context to ensure the model doesn't "forget" images in multi-turn chats.
• Model Agnostic: Works with any text-only model in OpenWebUI.
• Universal Compatibility: Handles both modern OpenAI-format image uploads and legacy/Ollama formats.
• System Framing: Injects descriptions as "System Tool Output" so the model knows it is seeing the image, rather than thinking the user typed the description.