Update src/app.py

src/app.py

@@ -21,37 +21,48 @@ _CACHED_RETRIEVER = None
 def get_retriever():
     global _CACHED_RETRIEVER
     if _CACHED_RETRIEVER is not None: return _CACHED_RETRIEVER
-
+    
+    # 1. Setup Keys
     pinecone_key = os.environ.get("PINECONE_API_KEY") or st.secrets.get("PINECONE_API_KEY")
     google_key = os.environ.get("GOOGLE_API_KEY") or st.secrets.get("GOOGLE_API_KEY")
-
-    if not pinecone_key or not google_key:
-        raise ValueError("Missing API Keys.")
-
+    
+    if not pinecone_key or not google_key: raise ValueError("Missing API Keys.")
     os.environ["PINECONE_API_KEY"] = pinecone_key
     os.environ["GOOGLE_API_KEY"] = google_key
 
+    # 2. Setup Vector Store (Pinecone) with SCORE THRESHOLD
     embeddings = GoogleGenerativeAIEmbeddings(model="models/text-embedding-004")
     vector_store = PineconeVectorStore(index_name=INDEX_NAME, embedding=embeddings)
-    vector_retriever = vector_store.as_retriever(search_kwargs={"k": 30})
+    
+    # CRITICAL PROFESSIONAL FIX:
+    # We set a "score_threshold" of 0.5.
+    # This means: "If the AI is less than 50% sure, DO NOT show the result."
+    # This kills the "Prayer Card B" noise immediately.
+    vector_retriever = vector_store.as_retriever(
+        search_type="similarity_score_threshold",
+        search_kwargs={"k": 20, "score_threshold": 0.5} 
+    )
 
+    # 3. Setup Keyword Store (BM25)
     keyword_retriever = None
     if os.path.exists(CHUNKS_FILE):
         try:
             with open(CHUNKS_FILE, "rb") as f:
                 chunks = pickle.load(f)
             keyword_retriever = BM25Retriever.from_documents(chunks)
-            keyword_retriever.k = 30
-        except: pass
+            keyword_retriever.k = 20
+        except Exception as e:
+            print(f"BM25 Error: {e}")
 
+    # 4. Create Hybrid Ensemble
     if keyword_retriever:
         final_retriever = EnsembleRetriever(
             retrievers=[vector_retriever, keyword_retriever],
-            weights=[0.5, 0.5]
+            weights=[0.4, 0.6] # 40% Vector (Concepts), 60% Keyword (Precision)
         )
     else:
         final_retriever = vector_retriever
-    
+
     _CACHED_RETRIEVER = final_retriever
     return final_retriever