detect_then_analyze.py

"""Routing a mixed inbox.

The pattern is detect (1 credit per 10 calls, amortised) → analyze with the chosen preset
(normal price). Worth it when the *type* is the decision: you can refuse to spend 40
credits on a 40-page PDF that turns out to be a scanned fax cover sheet.

If you just want the data and don't care which preset produced it, skip this and pass
preset="auto" — classification bundled into an extraction is free.

    python examples/detect_then_analyze.py ./inbox
"""

import json
import sys
from pathlib import Path

from visionapi import VisionAPI, unwrap

inbox = Path(sys.argv[1] if len(sys.argv) > 1 else "./inbox")
vision = VisionAPI()

# What we are willing to spend extraction credits on, and what to do otherwise.
HANDLED = {"invoice", "receipt", "purchase_order", "packing_slip"}
SUFFIXES = {".pdf", ".png", ".jpg", ".jpeg", ".webp", ".tif", ".tiff"}

for path in sorted(p for p in inbox.iterdir() if p.suffix.lower() in SUFFIXES):
    guess = vision.detect(file=path)

    # fallback=True means nothing matched confidently and the generic preset was
    # substituted. That is "shape unknown", not a match — check what was rejected.
    if guess["fallback"]:
        closest = guess["detections"][0] if guess["detections"] else None
        extra = f" (closest: {closest['preset']} — {closest['reason']})" if closest else ""
        print(f"{path.name}: unrecognised{extra}")
        continue

    if guess["recommended"] not in HANDLED:
        print(f"{path.name}: {guess['recommended']} — not a type we process, skipping")
        continue

    # `recommended` is exactly what preset="auto" would have run, so it is safe to commit.
    res = vision.analyze(file=path, preset=guess["recommended"])
    summary = json.dumps(unwrap(res["result"], drop_null=True), ensure_ascii=False)[:160]
    print(f"{path.name}: {guess['recommended']} → {res['credits_used']} credit(s)  {summary}")

print(f"\n{vision.credits()['balance']} credits remaining")