dsh-vision-bridge

In conversational AI applications, pure text models such as DeepSeek V4/4-Flash are inherently limited in processing image inputs due to their architecture. When users paste images, systems often throw a MODEL_DOES_NOT_SUPPORT_IMAGES error, disrupting multimodal interactions. The dsh-vision-bridge plugin elegantly solves this pain point by acting as a "vision bridge": it automatically invokes a user-configured OpenAI-compatible vision model (e.g., Qwen-VL-Max, GLM-4V) to convert images into textual descriptions, which are then fed to DeepSeek for further reasoning. This allows users to seamlessly integrate image understanding while leveraging the strong text capabilities of pure text models.

The plugin's core strength lies in its transparent workflow. By patching the model info resolution and streaming request processes, it tricks the system into believing the text model supports images, permitting image attachments into the session. Before sending requests to the model, the plugin scans all image blocks in messages and replaces them with descriptions generated by the vision model. For models that natively support images, it passes through the original images without conversion to avoid overhead. Additionally, it features caching based on image IDs to prevent redundant vision model calls and associated costs, and gracefully degrades with error messages if vision calls fail, ensuring uninterrupted dialogue.

This plugin is ideal for developers, researchers, or content creators who wish to leverage pure text models like the DeepSeek series but need to handle image content. For instance, in customer service chatbots, educational tools, or content analysis scenarios, users might send screenshots or photos, and the model must respond based on this visual information. With dsh-vision-bridge, you can extend the capabilities of text models without switching to expensive multimodal alternatives, achieving a balance between cost and functionality.