Add project files

pipeline/verify.py

@@ -0,0 +1,76 @@
+"""
+Optional third-party AI verification, layered on top of the core
+feature-matching pipeline: sends the winning template's original photo and
+the user's original upload to an external vision-LLM endpoint that runs a
+detailed QC-style comparison (flowers, vase, ribbon, composition) and
+returns a MATCH/DISCREPANCIES/CONFIDENCE verdict plus free-text description.
+
+This never affects the core match result -- if the endpoint is slow, down,
+or unreachable (it's a Cloudflare tunnel, which can go stale), the caller
+is expected to treat any exception here as a soft failure.
+"""
+
+import logging
+import os
+import re
+
+import requests
+
+import config
+
+logger = logging.getLogger(__name__)
+
+_VERDICT_RE = re.compile(
+    # DISCREP\w* rather than a literal "DISCREPANCIES" -- the LLM behind
+    # the endpoint doesn't reliably spell it the same way every time
+    # ("DISCREPANCIES" vs "DISCREPENCIES" have both been observed), and a
+    # missed match here silently dumps the whole raw block into the UI.
+    r"MATCH:\s*\[?\s*(?P<match>YES|NO|PARTIAL)\s*\]?\s*"
+    r"DISCREP\w*:\s*\[?\s*(?P<discrepancies>.*?)\s*\]?\s*"
+    r"CONFIDENCE:\s*\[?\s*(?P<confidence>High|Medium|Low)\s*\]?",
+    re.IGNORECASE | re.DOTALL,
+)
+
+
+def _parse_result_text(text: str) -> dict:
+    """The endpoint's `result` field is free text with an embedded
+    MATCH/DISCREPANCIES/CONFIDENCE block, sometimes followed by a prose
+    description, sometimes not. Pull out the structured bits; whatever's
+    left over (if anything) is the description."""
+    text = text or ""
+    m = _VERDICT_RE.search(text)
+
+    if not m:
+        return {"match": None, "confidence": None, "discrepancies": None,
+                "description": text.strip()}
+
+    return {
+        "match": m.group("match").upper(),
+        "confidence": m.group("confidence").capitalize(),
+        "discrepancies": m.group("discrepancies").strip() or "None",
+        "description": text[m.end():].strip(),
+    }
+
+
+def verify_images(reference_path: str, actual_path: str) -> dict:
+    """reference_path = the matched template's original photo (IMAGE 1),
+    actual_path = the user's original upload (IMAGE 2)."""
+    with open(reference_path, "rb") as f1, open(actual_path, "rb") as f2:
+        files = {
+            "image1": (os.path.basename(reference_path), f1, "image/png"),
+            "image2": (os.path.basename(actual_path), f2, "image/jpeg"),
+        }
+        response = requests.post(config.VERIFY_ENDPOINT_URL, files=files,
+                                  timeout=config.VERIFY_TIMEOUT_SECONDS)
+    response.raise_for_status()
+    data = response.json()
+
+    parsed = _parse_result_text(data.get("result", ""))
+    return {
+        "match": parsed["match"],
+        "confidence": parsed["confidence"],
+        "discrepancies": parsed["discrepancies"],
+        "description": parsed["description"],
+        "raw_result": data.get("result"),
+        "pixel_precheck": data.get("pixel_precheck"),
+    }