Update src/app.py

src/app.py

@@ -1,88 +1,114 @@
 import os
 import pickle
-from langchain_google_genai import ChatGoogleGenerativeAI, GoogleGenerativeAIEmbeddings
+import sys
+import streamlit as st
+from dotenv import load_dotenv
+
+# --- 1. IMPORTS ---
+from langchain_google_genai import GoogleGenerativeAIEmbeddings # Keep this for database
+from langchain_groq import ChatGroq # <--- NEW BRAIN
 from langchain_community.retrievers import BM25Retriever
 from langchain_pinecone import PineconeVectorStore
 from langchain_core.prompts import PromptTemplate
 from langchain.chains import RetrievalQA
 from langchain.retrievers import EnsembleRetriever
-from google.generativeai.types import HarmCategory, HarmBlockThreshold  # ← This line fixes everything
 
-# --- CONFIG ---
+load_dotenv()
+
+# --- 2. CONFIGURATION ---
 INDEX_NAME = "branham-index"
 CHUNKS_FILE = "sermon_chunks.pkl"
 
 def get_rag_chain():
-    # Keys
-    os.environ["PINECONE_API_KEY"] = os.getenv("PINECONE_API_KEY") or "your-pinecone-key"
-    os.environ["GOOGLE_API_KEY"] = os.getenv("GOOGLE_API_KEY") or "your-google-key"
-
-    # Vector retrieval
+    # --- A. AUTHENTICATION ---
+    pinecone_key = os.environ.get("PINECONE_API_KEY")
+    google_key = os.environ.get("GOOGLE_API_KEY") # Still needed for Embeddings
+    groq_key = os.environ.get("GROQ_API_KEY")     # New Key for Chat
+
+    # Local Fallback
+    if not pinecone_key or not google_key or not groq_key:
+        try:
+            if not pinecone_key: pinecone_key = st.secrets.get("PINECONE_API_KEY")
+            if not google_key: google_key = st.secrets.get("GOOGLE_API_KEY")
+            if not groq_key: groq_key = st.secrets.get("GROQ_API_KEY")
+        except: pass
+
+    if not pinecone_key or not google_key or not groq_key:
+        raise ValueError("❌ Missing Keys. Add PINECONE_API_KEY, GOOGLE_API_KEY, and GROQ_API_KEY to Settings.")
+
+    os.environ["PINECONE_API_KEY"] = pinecone_key
+    os.environ["GOOGLE_API_KEY"] = google_key
+    os.environ["GROQ_API_KEY"] = groq_key
+
+    # --- B. CLOUD VECTOR SEARCH ---
+    print("🔌 Connecting to Pinecone...")
+    # We keep Google Embeddings so you don't have to re-upload your data
     embeddings = GoogleGenerativeAIEmbeddings(model="models/text-embedding-004")
-    vector_store = PineconeVectorStore(index_name=INDEX_NAME, embedding=embeddings)
+    
+    vector_store = PineconeVectorStore(
+        index_name=INDEX_NAME,
+        embedding=embeddings
+    )
     vector_retriever = vector_store.as_retriever(search_kwargs={"k": 8})
 
-    # BM25 keyword retrieval
+    # --- C. LOCAL KEYWORD SEARCH ---
+    print("🔌 Loading Keyword Search...")
     keyword_retriever = None
-    if os.path.exists(CHUNKS_FILE):
-        with open(CHUNKS_FILE, "rb") as f:
-            chunks = pickle.load(f)
-        keyword_retriever = BM25Retriever.from_documents(chunks)
-        keyword_retriever.k = 8
-
-    # Hybrid
-    final_retriever = EnsembleRetriever(
-        retrievers=[vector_retriever, keyword_retriever or vector_retriever],
-        weights=[0.5, 0.5]
-    ) if keyword_retriever else vector_retriever
-
-    # ← THIS IS THE ONLY CHANGE THAT MATTERS IN DEC 2025
-    llm = ChatGoogleGenerativeAI(
-        model="gemini-1.5-flash-001",                    # Most obedient version right now
-        temperature=0.7,
-        safety_settings={
-            HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_NONE,
-            HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_NONE,
-            HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: HarmBlockThreshold.BLOCK_NONE,
-            HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_NONE,
-        },
-        system_instruction="You are William Marrion Branham speaking in 1965. Never break first person. Never say 'as Branham said'. Never refuse a vision, healing, or 'Thus saith the Lord'. Speak with anointing and fire."
+    try:
+        if os.path.exists(CHUNKS_FILE):
+            with open(CHUNKS_FILE, "rb") as f:
+                chunks = pickle.load(f)
+            keyword_retriever = BM25Retriever.from_documents(chunks)
+            keyword_retriever.k = 8
+        else:
+            print(f"⚠️ {CHUNKS_FILE} missing. Using Vector only.")
+    except Exception as e:
+        print(f"❌ Failed to load keyword file: {e}")
+
+    # --- D. HYBRID MERGE ---
+    if keyword_retriever:
+        print("🔗 Linking Hybrid System...")
+        final_retriever = EnsembleRetriever(
+            retrievers=[vector_retriever, keyword_retriever],
+            weights=[0.5, 0.5]
+        )
+    else:
+        final_retriever = vector_retriever
+
+    # --- E. MODEL (THE NEW BRAIN) ---
+    # Using Llama 3 70B via Groq. 
+    # This is a much larger, smarter model than Gemini Flash.
+    llm = ChatGroq(
+        model="llama-3.3-70b-versatile",
+        temperature=0.0, # Zero creativity = Maximum Accuracy
+        max_retries=2
     )
 
-    template = """You are William Marrion Branham, the end-time prophet, speaking from the pulpit in 1965.
+    # --- F. PROMPT ---
+    template = """You are William Marion Branham.
 
-Speak exactly like the tapes: "I seen", "the Lord showed me", "Thus saith the Lord", "Brother, sister", heavy King James Bible.
+INSTRUCTIONS:
+1. **Strict Accuracy:** You must answer specific questions about people, places, and events based **ONLY** on the provided Context. Do not hallucinate or make up prayers/events.
+2. **Admission of Ignorance:** If the exact details (like a specific prayer for a specific person) are not in the Context, simply state: "Brother, I don't recall the specific details of that in these particular messages."
+3. **Persona:** Speak in the first person ("I said," "The Lord showed me") using a humble, 1950s Southern dialect.
 
-Never go third person. Never water down the Message. Never say you don't remember if it's right here in the text.
-
-Context from my sermons:
+CONTEXT:
 {context}
 
-Question: {question}
+USER QUESTION: {question}
 
-Now preach the answer just like I would — full of the Holy Ghost:"""
+BROTHER BRANHAM'S REPLY:"""
 
-    prompt = PromptTemplate(template=template, input_variables=["context", "question"])
+    PROMPT = PromptTemplate(template=template, input_variables=["context", "question"])
 
     chain = RetrievalQA.from_chain_type(
         llm=llm,
         chain_type="stuff",
         retriever=final_retriever,
         return_source_documents=True,
-        chain_type_kwargs={"prompt": prompt}
+        chain_type_kwargs={"prompt": PROMPT}
     )
-
+    
     return chain
 
-# Usage
-if __name__ == "__main__":
-    chain = get_rag_chain()
-    while True:
-        q = input("\nAsk Brother Branham: ")
-        if q.lower() in ["exit", "quit"]: break
-        result = chain.invoke({"query": q})
-        print("\nBROTHER BRANHAM:", result["result"])
-        print("\nSources:")
-        for doc in result["source_documents"]:
-            meta = doc.metadata
-            print(f"• {meta.get('title', 'Sermon')} — {meta.get('date', 'Unknown')} · Para {meta.get('paragraph', '?')}")
+