Update app.py

app.py

@@ -99,21 +99,31 @@ async def lifespan(app: FastAPI):
                                # model="gemini-1.5-pro",
                                model="gemini-2.0-flash",
                                api_key=GOOGLE_API_KEY,
-                               temperature=0.1,
-                               max_output_tokens=350
+                               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 Question Answering.
+
 **Instructions**:
-- 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.
-- 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.
+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.
+- 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.*
+
+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 3 - **Final Output**:
+- Output the revised, cohesive set of answers (one sentence per sub-question).
+
 ---
 **Context**:
 {context}
@@ -225,7 +235,7 @@ 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": 12 ,"lambda_mult": 0.7} ) 
+    dense_retriever = vectorstore.as_retriever(search_type="mmr",search_kwargs={"k": 13 ,"lambda_mult": 0.7} ) 
     # dense_retriever = vectorstore.as_retriever(search_type="similarity" ,search_kwargs={"k": 11} )
     
 
@@ -233,7 +243,7 @@ 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=[0.35, 0.65],search_kwargs={"k": 10})
+    ensemble_retriever = EnsembleRetriever(retrievers=[keyword_retriever, dense_retriever], weights=[0.35, 0.65],search_kwargs={"k": 12})
     ### 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"]