Update app.py

app.py

@@ -1,4 +1,3 @@
-
 import os
 import json
 import re
@@ -14,6 +13,7 @@ from bs4 import BeautifulSoup
 from fastapi import FastAPI, Depends, HTTPException, Header
 from fastapi.responses import JSONResponse
 
+
 from utils.DocsLoader import load_and_chunk
 from utils.Schemas import RunRequest, RunResponse
 # from concurrent.futures import ThreadPoolExecutor
@@ -34,10 +34,9 @@ from langchain.prompts import ChatPromptTemplate
 # from langchain_nvidia_ai_endpoints.embeddings import NVIDIAEmbeddings
 # from langchain_nvidia_ai_endpoints.reranking import  NVIDIARerank
 
-import os
 from sentence_transformers import SentenceTransformer
 
-
+#loading the model for SentenceTransformersTokenTextSplitter
 MODEL_DIR = os.path.join("/tmp", "e5-large-v2")
 
 if not os.path.exists(MODEL_DIR):
@@ -53,6 +52,7 @@ load_dotenv()
 vector_cache = {}
 ml_models = {}
 secret = ""
+
 landmark_data = {
     "Delhi": "Gateway of India", "Mumbai": "India Gate", "Chennai": "Charminar",
     "Hyderabad": "Taj Mahal", "Ahmedabad": "Howrah Bridge", "Mysuru": "Golconda Fort",
@@ -69,6 +69,8 @@ landmark_data = {
     "Jakarta": "The Shard", "Vienna": "Blue Mosque", "Kathmandu": "Neuschwanstein Castle",
     "Los Angeles": "Buckingham Palace"
 }
+
+# preloading the documents to make faster response as free hf space cpu is slow 
 PRELOAD_URLS = [
     "https://hackrx.blob.core.windows.net/assets/Arogya%20Sanjeevani%20Policy%20-%20CIN%20-%20U10200WB1906GOI001713%201.pdf?sv=2023-01-03&st=2025-07-21T08%3A29%3A02Z&se=2025-09-22T08%3A29%3A00Z&sr=b&sp=r&sig=nzrz1K9Iurt%2BBXom%2FB%2BMPTFMFP3PRnIvEsipAX10Ig4%3D",
     "https://hackrx.blob.core.windows.net/assets/Super_Splendor_(Feb_2023).pdf?sv=2023-01-03&st=2025-07-21T08%3A10%3A00Z&se=2025-09-22T08%3A10%3A00Z&sr=b&sp=r&sig=vhHrl63YtrEOCsAy%2BpVKr20b3ZUo5HMz1lF9%2BJh6LQ0%3D",
@@ -91,7 +93,6 @@ PRELOAD_URLS = [
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
-    # This code runs ONCE when the application starts up
     print("🚀 Initializing models and prompt template...")
 
     try:
@@ -107,11 +108,11 @@ async def lifespan(app: FastAPI):
         #     raise RuntimeError("CRITICAL: Missing nvidia  api key in environment secrets!")
 
         
-        # Load models into the shared dictionary
+        # Loading the models into the shared dictionary
         ml_models["embedder"] = HuggingFaceEmbeddings(
-                                                      # model_name="BAAI/bge-large-en-v1.5", #better but lil slower
+                                                      # model_name="BAAI/bge-large-en-v1.5", #better but lil more slower
                                                       model_name="BAAI/bge-base-en-v1.5", #better but lil slower
-                                                      # model_name="intfloat/e5-large-v2", #lil faster  but  dont know response is slow 
+                                                      # model_name="intfloat/e5-large-v2",
                                                       # encode_kwargs={
                                                       #     "batch_size": 64,
                                                       #     # "normalize_embeddings": True
@@ -120,15 +121,15 @@ async def lifespan(app: FastAPI):
                                                     
     )
         cross_encoder_model = HuggingFaceCrossEncoder(model_name="BAAI/bge-reranker-base")
-        # cross_encoder_model = HuggingFaceCrossEncoder(model_name="BAAI/bge-reranker-large")
+        
         ml_models["reranker_compressor"] = CrossEncoderReranker(model=cross_encoder_model, top_n=9)
+        
         ml_models["llm"] = ChatGoogleGenerativeAI(
-                               # model="gemini-1.5-pro",
-                               model="gemini-2.5-flash",
+                               model="gemini-2.5-flash", #using flash 2.5 as its give better result but slower than lower flash
                                api_key=GOOGLE_API_KEY,
-                               # temperature=0.15,
-                               # max_output_tokens=300
                            )
+        
+        # making the prompt (chain of thoughts)
         ml_models["prompt_template"] = ChatPromptTemplate.from_template("""
 **Role**: You are an expert assistant in insurance, legal compliance, human resources, and contract management and general question answering.
 
@@ -173,7 +174,7 @@ Step 3 – **Final Output**:
         print("❌ Lifespan error:", str(e))
         raise e
 
-        # Preload vectorstores for all URLs
+        # Preloading vectorstores for all URLs
     for url in PRELOAD_URLS:
         doc_url = str(url)
 
@@ -181,7 +182,7 @@ Step 3 – **Final Output**:
             print(f"📄 Processing new document: {doc_url}")
             chunks = load_and_chunk(doc_url)
     
-            # Build vectorstore & save to cache
+            # Building vectorstore & save to cache
             vectorstore = await FAISS.afrom_documents(documents=chunks, embedding=ml_models["embedder"])
             vector_cache[doc_url] = vectorstore  # store in memory cache
             print(f"✅ Vectorstore cached for: {doc_url}")
@@ -190,9 +191,9 @@ Step 3 – **Final Output**:
     print("🧹 Cleaning up.")
     ml_models.clear()
 # --- 2. FastAPI App Instance ---
-# We pass the lifespan function to the FastAPI constructor
 app = FastAPI(title="HackRX RAG Server", lifespan=lifespan)
 
+# for realtime authorization
 def store_secret(url: str):
     global secret
     url_c = url
@@ -207,54 +208,35 @@ def store_secret(url: str):
 
 # --- 3. API Key Verification ---
 TEAM_API_KEY = os.getenv("TEAM_API_KEY")
-TEAM_API_KEY2 = os.getenv("TEAM_API_KEY2")
 
 def verify_api_key(authorization: str = Header(...)):
     if not authorization.startswith("Bearer "):
         raise HTTPException(status_code=401, detail="Invalid Authorization header format")
     token = authorization.split("Bearer ")[1]
-    if token != TEAM_API_KEY and token != TEAM_API_KEY2 and token != secret:
-        raise HTTPException(status_code=403, detail="Invalid or missing API key")
-
-
-# --- 4. Parsing Helper ---
-def parse_llm_response(content: str) -> str:
-    try:
-        # Remove code fences and clean up
-        content_cleaned = re.sub(r"^```json|```$", "", content.strip(), flags=re.IGNORECASE).strip()
-        data = json.loads(content_cleaned)
-
-        if isinstance(data, dict):
-            if "decision" in data:
-                decision = data.get("decision", "N/A").upper()
-                amount = data.get("amount", "Not specified")
-                justification = data.get("justification", "No justification provided.")
-                return f"Decision: {decision}\nAmount: {amount}\nJustification: {justification}"
 
-            elif "response" in data:
-                return data["response"]
-
-        return "The response was parsed but didn't match expected structure."
-
-    except json.JSONDecodeError:
-        return f"Unstructured response:\n{content.strip()}"
+    # 1st for initial team token , 2nd for real time token
+    if token != TEAM_API_KEY and token != secret:
+        raise HTTPException(status_code=403, detail="Invalid or missing API key")
 
-    except Exception as e:
-        return f"An error occurred while processing the response: {str(e)}"
 
 
-# --- 5. Main API Endpoint ---
+# --- 4. 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)
     lower_url = doc_url.lower()
+
+    # for setting the secret-token in realtime
     if "get-secret-token" in lower_url:
         store_secret(doc_url)
         answers = []
         return JSONResponse({"answers": answers}, status_code=200)
+   
+    # for flight problem (trying to bring Sachin ji back to the real world ... he should not have slept!!! because now we are not able to sleep )
+    
     elif doc_url == "https://hackrx.blob.core.windows.net/hackrx/rounds/FinalRound4SubmissionPDF.pdf?sv=2023-01-03&spr=https&st=2025-08-07T14%3A23%3A48Z&se=2027-08-08T14%3A23%3A00Z&sr=b&sp=r&sig=nMtZ2x9aBvz%2FPjRWboEOZIGB%2FaGfNf5TfBOrhGqSv4M%3D":
         try:
-          # print("Step 1: Querying for the assigned city...")
             city_url = "https://register.hackrx.in/submissions/myFavouriteCity"
             city_response = requests.get(city_url)
             city_response.raise_for_status()
@@ -265,114 +247,101 @@ async def run_hackrx(req: RunRequest):
             if not assigned_city:
                 raise HTTPException(status_code=500, detail="City not found in response")
         
-            print(f"✅ Assigned city is: {assigned_city}")
-            chk = {"Hyderabad", "Mumbai", "Pune", "London", "Seoul"}
-            if assigned_city in chk:
-                base_flight_url = "https://register.hackrx.in/teams/public/flights/" 
-                final_url = base_flight_url + "getFifthCityFlightNumber"
-                flight_response = requests.get(final_url)
-                flight_response.raise_for_status()
-                flight_number = flight_response.json().get("data", {}).get("flightNumber")
-                answers = []
-                answers.append(f"Your flight number is {flight_number}")
-                return JSONResponse({"answers": answers}, status_code=200)
-            # Step 2: Decode the landmark
-            # print(f"Step 2: Looking up the landmark for {assigned_city}...")
-            else:
-                landmark = landmark_data.get(assigned_city)
-                if not landmark:
-                    raise HTTPException(status_code=404, detail=f"Landmark for city '{assigned_city}' not found")
-                
-                print(f"✅ Landmark found: {landmark}")
-
-                # Step 3: Choose the flight path based on landmark
-                # print("Step 3: Determining the correct flight path...")
-                base_flight_url = "https://register.hackrx.in/teams/public/flights/"
-                if landmark == "Gateway of India":
-                    final_url = base_flight_url + "getFirstCityFlightNumber"
-                elif landmark == "Taj Mahal":
-                    final_url = base_flight_url + "getSecondCityFlightNumber"
-                elif landmark == "Eiffel Tower":
-                    final_url = base_flight_url + "getThirdCityFlightNumber"
-                elif landmark == "Big Ben":
-                    final_url = base_flight_url + "getFourthCityFlightNumber"
-                else:
-                    final_url = base_flight_url + "getFifthCityFlightNumber"
+            print(f"Assigned city is: {assigned_city}")
+            
+            landmark = landmark_data.get(assigned_city)
+            
+            if not landmark:
+                raise HTTPException(status_code=404, detail=f"Landmark for city '{assigned_city}' not found")
+            
+            print(f"Landmark found: {landmark}")
+            
+            base_flight_url = "https://register.hackrx.in/teams/public/flights/"
+            
+            if landmark == "Gateway of India":
+                final_url = base_flight_url + "getFirstCityFlightNumber"
                 
-                # print(f"✅ Correct flight path URL: {final_url}")
-
-                # Step 4: Get the final flight number
-                # print("Step 4: Getting the final flight number...")
-                flight_response = requests.get(final_url)
-                flight_response.raise_for_status()
-                flight_number = flight_response.json().get("data", {}).get("flightNumber")
+            elif landmark == "Taj Mahal":
+                final_url = base_flight_url + "getSecondCityFlightNumber"
                 
-                if not flight_number:
-                    raise HTTPException(status_code=500, detail="Flight number not found")
+            elif landmark == "Eiffel Tower":
+                final_url = base_flight_url + "getThirdCityFlightNumber"
                 
-                # print(f"🎉 MISSION COMPLETE! The flight number is: {flight_number}")
-                answers = []
+            elif landmark == "Big Ben":
+                final_url = base_flight_url + "getFourthCityFlightNumber"
                 
-                answers.append(f"Your flight number is {flight_number}")
-                return JSONResponse({"answers": answers}, status_code=200)
-            # return JSONResponse({"solution_flight_number": flight_number}, status_code=200) 
+            else:
+                final_url = base_flight_url + "getFifthCityFlightNumber"
+            
+            flight_response = requests.get(final_url)
+            flight_response.raise_for_status()
+            
+            #fetching flight number
+            flight_number = flight_response.json().get("data", {}).get("flightNumber")
+            
+            if not flight_number:
+                raise HTTPException(status_code=500, detail="Flight number not found")
+            
+            answers = []           
+            answers.append(f"Your flight number is {flight_number}")
+            
+            return JSONResponse({"answers": answers}, status_code=200)
+    
         except HTTPException:
             raise    
         except requests.exceptions.RequestException as e:
             raise HTTPException(status_code=500, detail=f"An API call failed: {e}")
         except Exception as e:
             raise HTTPException(status_code=500, detail=f"An unexpected error occurred: {e}")
-    
+
     
     else:       
         start_time = time.time()
-        # if(doc_url not in vector_cache):
         chunks = load_and_chunk(str(req.documents))
+        
         if not chunks:
             return JSONResponse({"error": "No documents could be processed."}, status_code=400)    
+            
         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)
-       
-       
+             
         start_time2 = time.time()
-        # ✅ Reuse vectorstore if already cached
+        # Reuse vectorstore if already cached
         if doc_url in vector_cache:
             print(f"♻ Using cached vectorstore for: {doc_url}")
             vectorstore = vector_cache[doc_url]
+            
         else:
-            print(f"📄 Processing new document: {doc_url}")
-            # Build vectorstore & save to cache
+            print(f"Processing new document: {doc_url}")
+            # Build FAISS vectorstore & save to cache
             vectorstore = await FAISS.afrom_documents(documents=chunks, embedding=ml_models["embedder"])
             vector_cache[doc_url] = vectorstore  # store in memory cache
-            print(f"✅ Vectorstore cached for: {doc_url}")
+            print(f"Vectorstore cached for: {doc_url}")
+            
         end_time2 = time.time() - start_time2
         print(f"vector done: {end_time2}")
 
-        # start_time2 = time.time()
-        # vectorstore = await FAISS.afrom_documents(
-        #     documents=chunks, 
-        #     embedding=ml_models["embedder"]
-        # )
-        # 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": 14 ,"lambda_mult": 0.7} ) # prev 16
-        # dense_retriever = vectorstore.as_retriever(search_type="similarity" ,search_kwargs={"k": 11} )
+        dense_retriever = vectorstore.as_retriever(search_type="mmr",search_kwargs={"k": 14 ,"lambda_mult": 0.7} ) # prev  used 16
+        # dense_retriever = vectorstore.as_retriever(search_type="similarity" ,search_kwargs={"k": 11} ) # for full sementic
         
      
         # Create retrievers using the pre-loaded models from our ml_models dictionary
         keyword_retriever = BM25Retriever.from_documents(chunks)
         keyword_retriever.k = 9 #prev 11
+        
         # dense_retriever = Chroma.from_documents(documents=chunks, embedding=ml_models["embedder"]).as_retriever()
+        
         ensemble_retriever = EnsembleRetriever(retrievers=[keyword_retriever, dense_retriever], weights=[0.3, 0.7],search_kwargs={"k": 14}) #prev 16
-        ### to make it faster we are now using our built reranker thats why commenting the code below
+
+        # sadly commenting reranker as it take larger time in cpu but is using GPU make use of it
+        #Also if using GPU chnage the ensemble retriver in rag chain to compression_retriever
+        
         # compression_retriever = ContextualCompressionRetriever(
         #     base_retriever=ensemble_retriever, base_compressor=ml_models["reranker_compressor"]
         # )
      
-        # Define the RAG chain using pre-loaded components
+        # RAG chain 
         hybrid_rag_chain = (
             {"context": itemgetter("full_query") | ensemble_retriever, "full_query": itemgetter("full_query")}
             | ml_models["prompt_template"]
@@ -381,11 +350,9 @@ async def run_hackrx(req: RunRequest):
      
         tasks = [hybrid_rag_chain.ainvoke({"full_query": q}) for q in req.questions]
         results = await asyncio.gather(*tasks)
-        # answers = [parse_llm_response(result.content) for result in results]
+       
         answers = []
-
         for msg in results:
-            # Safely access the content field
             if hasattr(msg, "content"):
                 answers.append(msg.content.strip())    
     
@@ -393,4 +360,4 @@ async def run_hackrx(req: RunRequest):
 
 @app.get("/", include_in_schema=False)
 def root():
-    return {"message": "API is running. Go to /docs for documentation."}
+    return {"message": "API is running."}