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_key")
+        GOOGLE_API_KEY = os.getenv("gemini_api_key3")
         print("🔑 gemini_api_key:", "FOUND" if GOOGLE_API_KEY else "NOT FOUND")
 
         if not GOOGLE_API_KEY:
@@ -99,30 +99,32 @@ async def lifespan(app: FastAPI):
                                # model="gemini-1.5-pro",
                                model="gemini-2.0-flash",
                                api_key=GOOGLE_API_KEY,
-                               temperature=0.15,
-                               max_output_tokens=300
+                               # temperature=0.15,
+                               # max_output_tokens=300
                            )
         ml_models["prompt_template"] = ChatPromptTemplate.from_template("""
-**Role**: You are an expert assistant in insurance, legal compliance, human resources, contract management, and Question Answering.
+**Role**: You are an expert assistant in insurance, legal compliance, human resources, contract management, and general question answering.
 
 **Instructions**:
-Step 1 - **Initial Draft**:
+
+Step 1 – **Initial Draft**:
 - If the user query contains multiple questions, split them into clear, well-formed sub-questions.
-- Use ONLY the context provided below to generate answers.
-- For each sub-question, provide a concise, complete, single-sentence response.
-- Maintain the original order of the sub-questions in your responses.
-- Do NOT repeat the query text, do NOT number the answers, and separate them with a single space.
+- If the query or subqueris is a universal or general knowledge question (e.g., scientific facts, laws of motion), answer using your own knowledge if you think proper information is not there in the context given.
+- For all other domain-specific queries (e.g., insurance, legal, HR), rely strictly on the provided context to generate responses.
+- Answer each sub-question in  concise and grammatically correct sentence, improving phrasing and spelling.
+- Preserve the original order of sub-questions in your responses.
+- Do not repeat the query text, do not number the answers, and separate the answer of the subquery with a new paragraph.
 - Do NOT use line breakers ("/n" or "\") in between the answers.
-- Avoid boilerplate phrases like “the document states” or “as per the context.”
-- If the answer to a sub-question is not in the context, say exactly: *I do not know the answer of "subquery",Please ask query related to the Document only.*
+- Avoid boilerplate phrases like “the document states” or “according to the context.”
+- If a sub-question cannot be answered from the context (and is not general knowledge), respond with: *I do not know the answer of "subquery", Please ask query related to the Document only.*
 
-Step 2 - **Critique & Revise**:
-- Review the initial answers to identify any sub-questions where relevant information from the context might have been missed.
-- Revise those answers to better incorporate relevant text and improve completeness and accuracy.
-- Ensure language is professional, clear, and aligned with domain tone.
+Step 2 – **Critique & Revise**:
+- Review the initial answers for any missing or underused context.
+- Revise responses to improve accuracy, completeness, grammar and clarity based on the full context.
+- Maintain a professional and domain-appropriate tone.
 
-Step 3 - **Final Output**:
-- Output the revised, cohesive set of answers (one sentence per sub-question).
+Step 3 – **Final Output**:
+- Present the revised and cohesive set of responses.
 
 ---
 **Context**:
@@ -235,15 +237,15 @@ 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": 13 ,"lambda_mult": 0.7} ) 
+    dense_retriever = vectorstore.as_retriever(search_type="mmr",search_kwargs={"k": 17 ,"lambda_mult": 0.65} ) 
     # dense_retriever = vectorstore.as_retriever(search_type="similarity" ,search_kwargs={"k": 11} )
     
 
     # Create retrievers using the pre-loaded models from our ml_models dictionary
     keyword_retriever = BM25Retriever.from_documents(chunks)
-    keyword_retriever.k = 9
+    keyword_retriever.k = 12
     # dense_retriever = Chroma.from_documents(documents=chunks, embedding=ml_models["embedder"]).as_retriever()
-    ensemble_retriever = EnsembleRetriever(retrievers=[keyword_retriever, dense_retriever], weights=[0.35, 0.65],search_kwargs={"k": 12})
+    ensemble_retriever = EnsembleRetriever(retrievers=[keyword_retriever, dense_retriever], weights=[0.35, 0.7],search_kwargs={"k": 18})
     ### 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"]