ds4f-vision

name	ds4f-vision
description	Local vision for DeepSeek-V4-Flash-0731 via the on-prem Qwen3-VL sidecar. Use when the user attaches or mentions an image path/URL and needs visual facts, OCR, or image comparison — stay on 0731; do not switch to a VL model. Call from IPython: await ds4f_vision.describe_image(...), ocr_image(...), compare_images(...).

DSpark local vision (Prime Agent)

Prime Agent only supports HTTP MCP integrations today, so this skill talks to the Qwen3-VL sidecar (http://127.0.0.1:8889) directly from the kernel.

When to use

• User gives a local path or image URL and asks what is in it, colors, text, etc.
• Prefer these helpers over guessing from filenames.

IPython usage

import ds4f_vision

print(await ds4f_vision.describe_image("/abs/path.jpg", question="What color is the sweater?"))
print(await ds4f_vision.ocr_image("/abs/path.jpg"))
print(await ds4f_vision.compare_images(["/a.jpg", "/b.jpg"], question="What changed?"))

Notes

• Sidecar must be running (./start-deepseek-v4-flash-dspark.sh, ENABLE_VL_SIDECAR=1).
• Override base with env DSPARK_VL_BASE_URL (default http://127.0.0.1:8889).
• Max 4 images for compare. Huge images are auto-downscaled.