dsh-siliconflow-vision

The main conversational model in DSH typically accepts text only, so when a user drops a screenshot, a scanned document, or a product photo into the session the model neither describes the image nor reasons about its content. dsh-siliconflow-vision bridges that gap by registering an analyze_image tool that forwards the image to a SiliconFlow vision model (default Qwen/Qwen3-VL-32B-Instruct) and feeds the structured result back into the conversation, giving the model genuine visual understanding without leaving the current session.

The plugin accepts three image sources: a local file path on the server, an http(s) URL, or a base64 data URL. Users simply ask the model to "look at this image" and can attach a custom prompt such as "extract all text" or "what animals appear in the picture." The underlying model is swappable, ranging from the speed-focused Qwen3-VL-8B for quick lookups to PaddleOCR-VL-1.5 dedicated to optical character recognition. Under the hood the plugin uses Node.js native fetch against the OpenAI-compatible API; local images are read and converted to base64 in memory, so no temporary files are written to disk and the runtime footprint stays minimal.

It is well suited for developers and operators who build DSH workflows around image understanding: interpreting customer screenshots of error messages, auto-extracting text from receipts or documents, or having a product manager paste a competitor screenshot and ask for a feature comparison. Steps that previously required opening a separate tool or switching contexts now collapse into a single natural-language request within the conversation.