Update app.py

app.py

@@ -55,32 +55,36 @@ class EmbeddingBackend:
         return np.asarray(emb)
 
 # --------------------------------------------------
-# Hybrid (exact ➜ semantic) index
+# Hybrid (exact → semantic) index
 # --------------------------------------------------
 class HybridIndex:
     def __init__(self, df: pd.DataFrame, text_col: str, backend: EmbeddingBackend):
-        self.df        = df
-        self.text_col  = text_col
-        self.backend   = backend
-        self.ac_tree   = self._build_ac()
-        self.embeddings = self._build_emb()
+        self.df          = df
+        self.text_col    = text_col
+        self.backend     = backend
+        self.text_to_rows = defaultdict(list)  # passage → [row ids]
+        self.ac_tree     = self._build_ac()
+        self.embeddings  = self._build_emb()
 
     # ---------- exact match ----------
     def _build_ac(self):
         tree = KeywordTree(case_insensitive=True)
         for i, passage in self.df[self.text_col].astype(str).items():
-            tree.add(passage, i)
+            tree.add(passage)
+            self.text_to_rows[passage].append(i)
         tree.finalize()
         return tree
 
     def exact_hits(self, query: str) -> List[int]:
-        return list({m[1] for m in self.ac_tree.search_all(query)})
+        rows = set()
+        for keyword, _ in self.ac_tree.search_all(query):
+            rows.update(self.text_to_rows[keyword])
+        return list(rows)
 
     # ---------- semantic ----------
     def _build_emb(self):
         docs = self.df[self.text_col].astype(str).tolist()
         emb  = self.backend.encode_corpus(docs)
-        # Normalise for cosine similarity via dot‑product
         emb_norm = emb / np.linalg.norm(emb, axis=1, keepdims=True)
         return emb_norm.astype(np.float32)