Update app.py

app.py

@@ -106,13 +106,13 @@ async def lifespan(app: FastAPI):
         ml_models["prompt_template"] = ChatPromptTemplate.from_template("""
 **Role**: You are an expert assistant in insurance, legal compliance, human resources, and contract management.
 **Instructions**:
-- If the query contains multiple questions, split them into sub-questions.
+- 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; do not copy large clauses unless needed.
+- 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**:
@@ -229,14 +229,15 @@ 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": 10 ,"lambda_mult": 0.80} )
+    # dense_retriever = vectorstore.as_retriever(search_type="mmr",search_kwargs={"k": 10 ,"lambda_mult": 0.80} )
+    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 = 7
+    keyword_retriever.k = 8
     # 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=[0.3, 0.75],search_kwargs={"k": 9})
     ### 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"]