Update app.py

app.py

@@ -97,35 +97,36 @@ async def lifespan(app: FastAPI):
         ml_models["reranker_compressor"] = CrossEncoderReranker(model=cross_encoder_model, top_n=9)
         ml_models["llm"] = ChatGoogleGenerativeAI(
                                # model="gemini-1.5-pro",
-                               model="gemini-2.0-flash",
+                               model="gemini-2.5-flash",
                                api_key=GOOGLE_API_KEY,
                                # 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 general question answering.
+**Role**: You are an expert assistant in insurance, legal compliance, human resources, and contract management and general question answering.
 
 **Instructions**:
 
 Step 1 – **Initial Draft**:
-- If the user query contains multiple questions, split them into clear, well-formed sub-questions.
-- 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.
+- If the query contains multiple questions, split them into  perfect sub-questions.
+- Use ONLY the provided context to answer.
+- Provide one concise, complete sentence per sub-question.
+- List answers in the same order as the sub-questions, without repeating the query text.
+- Do not add numbering or bullet points; separate answers with a single space.
+- Make the answer well structured and with proper starting like a human is answering it.
+- Make grammatically correct sentence, improving phrasing and spelling.
+- Avoid phrases like “the provided document states” or “according to the context.”
+- Summarize relevant parts of the context without losing meaning.
 - 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.*
+- 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.
 
 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.
+- Revise responses to improve accuracy, completeness, grammar and clarity based on the  context.
 - Maintain a professional and domain-appropriate tone.
 
 Step 3 – **Final Output**:
 - Present the revised and cohesive set of responses.
-
 ---
 **Context**:
 {context}
@@ -237,15 +238,15 @@ 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": 12 ,"lambda_mult": 0.7} ) 
+    dense_retriever = vectorstore.as_retriever(search_type="mmr",search_kwargs={"k": 16 ,"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 = 8
+    keyword_retriever.k = 11
     # dense_retriever = Chroma.from_documents(documents=chunks, embedding=ml_models["embedder"]).as_retriever()
-    ensemble_retriever = EnsembleRetriever(retrievers=[keyword_retriever, dense_retriever], weights=[0.35, 0.7],search_kwargs={"k": 12})
+    ensemble_retriever = EnsembleRetriever(retrievers=[keyword_retriever, dense_retriever], weights=[0.4, 0.6],search_kwargs={"k": 16})
     ### 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"]