Skip to content

Vision·

take_photo — a frame plus a question for the multimodal model, shared by every personatools/vision.py, 1 tool.

From the module docstring

Vision tool — capture a TINY camera frame and inject it into the bidi voice agent's multimodal context (BidiImageInputEvent), so the realtime model SEES the image natively. Adapted from neon/tools/vision.py.

tool does
take_photo LOOK.

take_photo·

async take_photo(question: str = '', device: int = 0) -> dict

LOOK. Capture a frame from TINY's head camera and put it in front of the voice model right now — the model sees the image and answers in audio.

Call this FIRST whenever someone says "look at me", "what do you see", "who's there", "what is this", "can you see …" — never answer about what you see without calling it, and never say "I'll take a look" instead of calling it.

argument meaning
question what to answer about the image (default: describe what you see).
device dev fallback camera index (macOS). Ignored on the robot.

source: tools/vision.py:152