Update app.py

app.py

@@ -68,7 +68,7 @@ async def lifespan(app: FastAPI):
     print("🚀 Initializing models and prompt template...")
 
     try:
-        GOOGLE_API_KEY = os.getenv("gemini_api_key3")
+        GOOGLE_API_KEY = os.getenv("gemini_api_key")
         print("🔑 gemini_api_key:", "FOUND" if GOOGLE_API_KEY else "NOT FOUND")
 
         if not GOOGLE_API_KEY:
@@ -100,10 +100,10 @@ async def lifespan(app: FastAPI):
                                model="gemini-2.0-flash",
                                api_key=GOOGLE_API_KEY,
                                temperature=0.1,
-                               max_output_tokens=300
+                               max_output_tokens=350
                            )
         ml_models["prompt_template"] = ChatPromptTemplate.from_template("""
-**Role**: You are an expert assistant in insurance, legal compliance, human resources, and contract management.
+**Role**: You are an expert assistant in insurance, legal compliance, human resources, contract management and Question Answering.
 **Instructions**:
 - If the query contains multiple questions, split them into  perfect sub-questions.
 - Use ONLY the provided context to answer.
@@ -113,10 +113,7 @@ async def lifespan(app: FastAPI):
 - Avoid phrases like “the provided document states” or “according to the context.”
 - Summarize relevant parts of the context without losing meaning.
 - If the answer is not in the context for some subqueries, respond exactly with: " I do not know the answer of "subquery",Please ask query related to the Document only." for that subquery.
-- Keep answers professional, clear, and direct, avoiding unnecessary jargon.
-**Tone & Style**:
-- Professional yet approachable.
-- Clear and direct.
+- Keep answers professional, clear, and direct.
 ---
 **Context**:
 {context}
@@ -228,7 +225,7 @@ async def run_hackrx(req: RunRequest):
     # end_time2 = time.time() - start_time2
     # print(f"vector done: {end_time2}")
     # dense_retriever = vectorstore.as_retriever(search_type="mmr",search_kwargs={"k": 8})
-    dense_retriever = vectorstore.as_retriever(search_type="mmr",search_kwargs={"k": 16 ,"lambda_mult": 0.8} ) 
+    dense_retriever = vectorstore.as_retriever(search_type="mmr",search_kwargs={"k": 12 ,"lambda_mult": 0.7} ) 
     # dense_retriever = vectorstore.as_retriever(search_type="similarity" ,search_kwargs={"k": 11} )
     
 
@@ -236,7 +233,7 @@ async def run_hackrx(req: RunRequest):
     keyword_retriever = BM25Retriever.from_documents(chunks)
     keyword_retriever.k = 9
     # dense_retriever = Chroma.from_documents(documents=chunks, embedding=ml_models["embedder"]).as_retriever()
-    ensemble_retriever = EnsembleRetriever(retrievers=[keyword_retriever, dense_retriever], weights=[0.3, 0.7],search_kwargs={"k": 15})
+    ensemble_retriever = EnsembleRetriever(retrievers=[keyword_retriever, dense_retriever], weights=[0.35, 0.65],search_kwargs={"k": 10})
     ### to make it faster we are now using our built reranker thats why commenting the code below
     # compression_retriever = ContextualCompressionRetriever(
     #     base_retriever=ensemble_retriever, base_compressor=ml_models["reranker_compressor"]