Update app.py

app.py

@@ -70,7 +70,7 @@ 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:
@@ -105,23 +105,24 @@ async def lifespan(app: FastAPI):
                                max_output_tokens=300
                            )
         ml_models["prompt_template"] = ChatPromptTemplate.from_template("""
-**Role**: You are an expert assistant in insurance, legal compliance, human resources, and contract management. 
-Your job is to answer using ONLY the provided context, in a way that feels natural, clear, and helpful — as if speaking to a client. 
-Do not invent or assume information that is not explicitly stated in the context just say. Do not use line breaks between the statements of the response. 
-If the answer is not found in the context, politely state that Please ask query related to the Document only.
-
+**Role**: You are an expert assistant in insurance, legal compliance, human resources, and contract management.
+**Instructions**:
+- Answer using ONLY the provided context.
+- Identify each distinct question in the query and give exactly one concise, complete sentence for each , Also do not use line breakers in between the sentence.
+- Summarize the relevant part of the context without losing meaning.
+- Do not copy large clauses verbatim unless absolutely necessary.
+- If the answer is not in the context, respond: "Please ask query related to the Document only."
+- Do not use unnecessary filler or legal jargon.
 **Tone & Style**:
-- Be professional yet approachable.
-- Use complete sentences and natural phrasing.
-- Avoid legal jargon unless it is in the context.
-- Keep answers concise but complete.
-- do not make it unnecessary lengthy.
-
-**Context**: 
+- Professional yet approachable.
+- Clear and direct.
+---
+**Context**:
 {context}
-
-**Query**: {full_query}
-
+---
+**Query**:
+{full_query}
+---
 **Response**:
 """
 )
@@ -192,16 +193,15 @@ def parse_llm_response(content: str) -> str:
 # --- 5. Main API Endpoint ---
 @app.post("/api/v1/hackrx/run", response_model=RunResponse, dependencies=[Depends(verify_api_key)])
 async def run_hackrx(req: RunRequest):
-
+    doc_url = str(req.documents) 
     start_time = time.time()
-    chunks = load_and_chunk(str(req.documents))
+    if(doc_url not in vector_cache):
+        chunks = load_and_chunk(str(req.documents))
     end_time = time.time() - start_time
     print(f"chunking done: {end_time}")
     if not chunks:
         return JSONResponse({"error": "No documents could be processed."}, status_code=400)
 
-    doc_url = str(req.documents)  # Assuming it's a URL or unique path
-
 
     start_time2 = time.time()
     # ✅ Reuse vectorstore if already cached
@@ -210,9 +210,6 @@ async def run_hackrx(req: RunRequest):
         vectorstore = vector_cache[doc_url]
     else:
         print(f"📄 Processing new document: {doc_url}")
-        if not chunks:
-            return JSONResponse({"error": "No documents could be processed."}, status_code=400)
-
         # Build vectorstore & save to cache
         vectorstore = await FAISS.afrom_documents(documents=chunks, embedding=ml_models["embedder"])
         vector_cache[doc_url] = vectorstore  # store in memory cache
@@ -235,15 +232,15 @@ async def run_hackrx(req: RunRequest):
     keyword_retriever = BM25Retriever.from_documents(chunks)
     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": 11})
+    ensemble_retriever = EnsembleRetriever(retrievers=[keyword_retriever, dense_retriever], weights=[0.35, 0.65],search_kwargs={"k": 20})
     ### 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"]
-    # )
+    compression_retriever = ContextualCompressionRetriever(
+        base_retriever=ensemble_retriever, base_compressor=ml_models["reranker_compressor"]
+    )
 
     # Define the RAG chain using pre-loaded components
     hybrid_rag_chain = (
-        {"context": itemgetter("full_query") | ensemble_retriever, "full_query": itemgetter("full_query")}
+        {"context": itemgetter("full_query") | compression_retriever, "full_query": itemgetter("full_query")}
         | ml_models["prompt_template"]
         | ml_models["llm"]
     )