diff --git a/code/scripts/build_similarity_cache_parquet.py b/code/scripts/build_similarity_cache_parquet.py
index 99d784d..cc39f6d 100644
--- a/code/scripts/build_similarity_cache_parquet.py
+++ b/code/scripts/build_similarity_cache_parquet.py
@@ -202,7 +202,8 @@ def build_cache(
     df = similarity.cards_df
     df["is_land"] = df["type"].str.contains("Land", case=False, na=False)
     df["is_multifaced"] = df["layout"].str.lower().isin(["modal_dfc", "transform", "reversible_card", "double_faced_token"])
-    df["tag_count"] = df["themeTags"].apply(lambda x: len(x.split("|")) if pd.notna(x) and x else 0)
+    # M4: themeTags is now a list (Parquet format), not a pipe-delimited string
+    df["tag_count"] = df["themeTags"].apply(lambda x: len(x) if isinstance(x, list) else 0)
 
     # Keep cards that are either:
     # 1. Not lands, OR