Update src/ranker.py

src/ranker.py

@@ -1,7 +1,6 @@
 """
 ranker.py
 ---------
-
 This module implements functionality for ranking candidate sentences by
 their relevance to a given claim.  The ranking is performed by
 embedding both the claim and the candidate sentences into a semantic
@@ -17,12 +16,10 @@ approach still yields reasonable relevance orderings without requiring
 deep learning dependencies.
 
 Example:
-
 >>> from ranker import rank_sentences
 >>> ranked = rank_sentences("Cats are adorable pets", ["Cats purr when happy", "Airplanes fly"], top_k=1)
 >>> print(ranked[0][0])
 ...  # prints the sentence most similar to the claim
-
 """
 
 from __future__ import annotations
@@ -40,7 +37,6 @@ _use_transformers = False
 
 def _load_sentence_transformer(model_name: str = "all-MiniLM-L6-v2"):
     """Load the sentence transformer model lazily.
-
     Attempts to import and instantiate the specified sentence
     transformer model.  If the import fails, sets a flag to indicate
     fallback use of scikit-learn.
@@ -71,7 +67,6 @@ def _embed_with_st(texts: Iterable[str]) -> np.ndarray:
 
 def _rank_with_tfidf(claim: str, candidates: List[str], top_k: int) -> List[Tuple[str, float]]:
     """Rank candidates using TF-IDF cosine similarity.
-
     This fallback method uses scikit-learn's TfidfVectorizer to
     construct vectors for the claim and candidates and then computes
     pairwise cosine similarity.  It does not require any heavy
@@ -97,10 +92,8 @@ def rank_sentences(claim: str, sentences: Iterable[str], top_k: int = 10) -> Lis
     ----------
     claim:
         The short textual claim against which candidates are compared.
-
     sentences:
         An iterable of candidate sentences to score.
-
     top_k:
         The maximum number of top-ranked sentences to return.  If the
         number of candidates is less than ``top_k``, all candidates are
@@ -116,6 +109,9 @@ def rank_sentences(claim: str, sentences: Iterable[str], top_k: int = 10) -> Lis
         ranking, the scores may be lower but are still comparable within
         the same run.
     """
+    # IMPORTANT: declare globals before any usage in this function
+    global _use_transformers, _st_model
+
     # Convert the iterable to a list so we can index and iterate
     candidates = list(sentences)
     if not candidates:
@@ -144,9 +140,8 @@ def rank_sentences(claim: str, sentences: Iterable[str], top_k: int = 10) -> Lis
                 exc,
             )
             # Mark the transformer as unusable for subsequent calls
-            global _use_transformers
             _use_transformers = False
             _st_model = None
 
     # Fallback to TF-IDF ranking
-    return _rank_with_tfidf(claim, candidates, top_k)
+    return _rank_with_tfidf(claim, candidates, top_k)