ocr/gemini_face_extractor.py

from ocr.base import GeminiExtractor

_PROMPT = """
Find the photograph of the person on this card. Get the exact position of the entire photo.

Return JSON:

{
 "label": "face_photo",
 "box_2d": [ymin, xmin, ymax, xmax]
}

Coordinates must be normalized between 0 and 1000.
Return ONLY JSON.
"""

# Single model; resize=False so returned bbox coordinates match the
# original image dimensions used later in crop_face().
_extractor = GeminiExtractor(_PROMPT, models=["gemini-2.5-flash"], resize=False)


def detect_face(image_bytes: bytes) -> dict:
    result = _extractor.extract(image_bytes)
    return result["data"]