scripts/ab_fragment.py

"""A/B: does a typed fragment meaningfully steer the output?

Runs the same images twice — once with no fragment, once with a raw personal
line — and prints both interpretations + film titles so we can see whether the
words lead (task #16) rather than just tint.

    .venv/bin/python -m scripts.ab_fragment [model]
"""

import os
import sys

from dotenv import load_dotenv

load_dotenv()
if len(sys.argv) > 1:
    os.environ["CHALCHITRA_MODEL"] = sys.argv[1]

from backend import interpret  # noqa: E402
from scripts.smoke_test import moody_jpeg  # noqa: E402

# Ambiguous-ish frames: warm dusk + a quiet interior. On their own they read as
# soft nostalgia; the fragment should pull them somewhere specific.
IMAGES = [
    moody_jpeg((30, 28, 36), (120, 86, 60), blob=(255, 190, 110, 70)),
    moody_jpeg((18, 18, 22), (60, 55, 64)),
]

FRAGMENT = "the last evening in the flat before he moved out for good"


def show(label: str, frag: str) -> None:
    print(f"\n{'='*70}\n{label}\nfragment: {frag!r}\n{'-'*70}")
    r = interpret(IMAGES, frag)
    print(r["interpretation"])
    print("\nfilms: " + " · ".join(f"{f['title']} ({f['year']})" for f in r["films"]))


def main() -> None:
    print(f"model={os.environ.get('CHALCHITRA_MODEL', 'default')}")
    show("A — images only", "")
    show("B — images + personal fragment", FRAGMENT)


if __name__ == "__main__":
    main()