extractor.py

import json
import re
from llm import generate_vision_pass

VISION_PROMPT = """Analyze this image and return ONLY a JSON object, no other text:
{
  "scene": "one sentence describing the setting",
  "mood": "one word emotional tone",
  "objects": ["key", "objects", "in", "scene"],
  "style": "visual style (cinematic, documentary, candid, etc)",
  "possible_interpretations": ["2-3 thematic meanings a debater could argue"]
}"""


def run_vision_pass(image_path: str) -> dict:
    """
    Extract structured scene context from image.
    Runs once per debate — result is passed to all persona turns as text.
    Sends image via local transformers ZeroGPU pipeline.
    """
    text = generate_vision_pass(image_path, VISION_PROMPT)

    match = re.search(r"\{.*\}", text, re.DOTALL)
    if not match:
        raise ValueError(f"Vision pass returned no JSON:\n{text}")

    return json.loads(match.group())


def build_context_block(scene: dict) -> str:
    """
    Format scene dict into a readable text block
    injected into every persona and judge prompt.
    """
    objects = ", ".join(scene["objects"]) if scene.get("objects") else "none"
    interpretations = (
        "; ".join(scene["possible_interpretations"])
        if scene.get("possible_interpretations")
        else "none"
    )

    return (
        f"IMAGE CONTEXT (factual, do not dispute this):\n"
        f"- Scene: {scene['scene']}\n"
        f"- Mood: {scene['mood']}\n"
        f"- Objects: {objects}\n"
        f"- Style: {scene['style']}\n"
        f"- Possible interpretations: {interpretations}"
    )