Update app.py

app.py

@@ -56,10 +56,27 @@ def _get_capacity_snapshot() -> dict:
 # there's no position bias and it's faster than even a 2-token LLM call —
 # just one forward pass per script+candidate pair, batched.
 # ══════════════════════════════════════════════════════════════════════════
+# Pixabay boilerplate/metadata tags that describe the *asset* rather than
+# its visual content, and can falsely spike similarity (e.g. "ai generated"
+# just means the clip was made with generative AI tools — it says nothing
+# about the subject matter — but it collides hard with any script mentioning
+# "AI"). Stripped before embedding.
+_NOISE_TAGS = {
+    "ai generated", "live wallpaper", "background", "wallpaper",
+    "beautiful wallpaper", "backdrop",
+}
+
+
+def _clean_tags(tags: str) -> str:
+    parts = [t.strip() for t in tags.split(",")]
+    kept = [t for t in parts if t.lower() not in _NOISE_TAGS]
+    return ", ".join(kept) if kept else tags  # never return empty
+
+
 def _select_best_position(script: str, videos: list) -> int:
     if not videos:
         return None
-    tag_strings = [v.get("tags", "") for v in videos]
+    tag_strings = [_clean_tags(v.get("tags", "")) for v in videos]
 
     script_emb = embedder.encode(script, convert_to_tensor=True)
     tag_embs = embedder.encode(tag_strings, convert_to_tensor=True)