viswadarshan06's picture
Update app.py
19179a4 verified
import os
import pickle
import faiss
import json
import numpy as np
import pandas as pd
import requests
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from sentence_transformers import SentenceTransformer
# ------------------
# Environment Setup
# ------------------
os.environ["HF_HOME"] = "/app/hf_cache"
os.environ["TRANSFORMERS_CACHE"] = "/app/hf_cache"
GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")
GOOGLE_API_KEY = os.getenv("GOOGLE_SEARCH_API_KEY")
GOOGLE_CX = os.getenv("GOOGLE_SEARCH_CX")
GEMINI_URL = f"https://generativelanguage.googleapis.com/v1beta/models/gemma-3n-e4b-it:generateContent?key={GEMINI_API_KEY}" if GEMINI_API_KEY else None
# ------------------
# FastAPI App Config
# ------------------
app = FastAPI()
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
# ------------------
# Load Components
# ------------------
try:
model = SentenceTransformer("/app/model")
except Exception as e:
print(f"Error loading model: {e}")
model = None
try:
with open("thirukkural_data.pkl", "rb") as f:
kural_data = pickle.load(f)
print(f"✅ Kural data loaded: {kural_data.shape}")
except Exception as e:
print(f"❌ Error loading data: {e}")
kural_data = None
try:
english_index = faiss.read_index("thirukkural_english_index.faiss")
tamil_index = faiss.read_index("thirukkural_tamil_index.faiss")
print("✅ FAISS indexes loaded")
except Exception as e:
print(f"❌ Error loading FAISS indexes: {e}")
english_index = tamil_index = None
# ------------------
# Request Schema
# ------------------
class QueryRequest(BaseModel):
query: str
lang: str # "en" or "ta"
top_k: int = 3
# ------------------
# Intent Classifier
# ------------------
def classify_intent(query: str) -> str:
query = query.lower()
if any(word in query for word in ["hello", "hi", "vanakkam", "hey"]):
return "greeting"
if any(word in query for word in ["thanks", "thank you", "nandri"]):
return "thanks"
if any(word in query for word in ["sad", "happy", "life", "love", "philosophy", "purpose", "feel", "emotion"]):
return "life_advice"
if any(word in query for word in ["what", "why", "how", "when", "where", "who", "search", "find"]):
return "web_search"
return "fallback"
# ------------------
# Utility Functions
# ------------------
def safe_get_value(row, keys, default=""):
for k in keys:
if k in row:
return row[k]
return default
def format_kural_for_response(kural_dict):
return {
"Number": safe_get_value(kural_dict, ["ID", "Index"], "Unknown"),
"Tamil": safe_get_value(kural_dict, ["Kural", "combined_text_tamil"]),
"English": safe_get_value(kural_dict, ["Couplet", "combined_text_english"]),
"Category": safe_get_value(kural_dict, ["Adhigaram", "Section", "Adhigaram_ID"], "General"),
"Explanation": safe_get_value(kural_dict, ["Vilakam", "Kalaingar_Urai", "Solomon_Pappaiya"], "")
}
def format_kural_for_prompt(kural_dict):
return f"#{safe_get_value(kural_dict, ['ID', 'Index'])}\nTamil: {safe_get_value(kural_dict, ['Kural'])}\nEnglish: {safe_get_value(kural_dict, ['Couplet'])}"
# ------------------
# Web Search Agent
# ------------------
def web_search_agent(query: str, lang: str = "en") -> dict:
if not GOOGLE_API_KEY or not GOOGLE_CX:
return {"query": query, "results": [], "explanation": "Google Search API credentials are missing."}
try:
params = {
"key": GOOGLE_API_KEY,
"cx": GOOGLE_CX,
"q": query,
"hl": lang,
"num": 5
}
response = requests.get("https://www.googleapis.com/customsearch/v1", params=params, timeout=10)
response.raise_for_status()
data = response.json()
results = [{"title": i.get("title"), "snippet": i.get("snippet"), "link": i.get("link")} for i in data.get("items", [])]
return {"query": query, "results": results, "explanation": f"Here are relevant results for: '{query}'"}
except Exception as e:
return {"query": query, "results": [], "explanation": f"Search failed: {str(e)}"}
# === Supervisor Agent ===
def classify_intent(query: str) -> str:
query_lower = query.lower().strip()
greetings = ["hi", "hello", "vanakkam", "hey", "good morning", "good evening"]
gratitude = ["thanks", "thank you"]
farewell = ["bye", "goodbye"]
emotional_keywords = [
"sad", "happy", "lonely", "depressed", "worried", "peace", "struggle",
"confidence", "purpose", "meaning", "life", "love", "failure", "success"
]
philosophical_keywords = ["karma", "virtue", "truth", "justice", "ethics", "philosophy", "soul"]
# Prioritize emotional/philosophical over question words
if any(word in query_lower for word in greetings + farewell + gratitude):
return "greeting"
if any(word in query_lower for word in emotional_keywords + philosophical_keywords):
return "life_advice"
# THEN check for factual / web queries
question_starters = ["who", "what", "when", "where", "why", "how"]
if any(query_lower.startswith(w) for w in question_starters):
return "factual_question"
if "search" in query_lower or "find" in query_lower:
return "web_search"
return "fallback"
# === Thirukkural RAG Agent ===
def thirukkural_rag_agent(query: str, lang: str, top_k: int):
if not all([model, kural_data, tamil_index, english_index]):
return {"error": "RAG components not loaded properly."}
# Embed and vector search
query_embedding = model.encode([query])
index = tamil_index if lang == "ta" else english_index
search_k = min(top_k, len(kural_data))
D, I = index.search(np.array(query_embedding).astype("float32"), search_k)
results = []
for i in I[0]:
if 0 <= i < len(kural_data):
row_dict = kural_data.iloc[i].to_dict()
results.append(format_kural_for_response(row_dict))
if not results:
return {"query": query, "results": [], "explanation": "No relevant Thirukkurals found."}
# Gemini Prompt
kural_texts = '\n\n'.join([f'#{k["Number"]}\nTamil: {k["Tamil"]}\nEnglish: {k["English"]}' for k in results])
prompt = f"""
You are a compassionate, wise, and culturally sensitive literary guide.
Your role is to explain how specific Thirukkural couplets relate meaningfully to a user's question. Your tone must be emotionally supportive, philosophically grounded, and respectful of the Tamil literary tradition.
The user has asked:
➡️ "{query}" (Language: {lang})
📜 Matching Thirukkurals:
{kural_texts}
🎯 Guidelines:
- Reply fully in Tamil if lang='ta'; in English if lang='en'.
- Don't rephrase Kural poems; show clearly and preserve original form.
- Avoid AI/tech mentions. Write like a human mentor.
- Use paragraph style, not list.
"""
# Gemini API Call
explanation = "Explanation unavailable."
if GEMINI_API_KEY and GEMINI_URL:
try:
headers = {"Content-Type": "application/json"}
data = {"contents": [{"parts": [{"text": prompt}]}]}
response = requests.post(GEMINI_URL, headers=headers, data=json.dumps(data), timeout=30)
response.raise_for_status()
response_data = response.json()
explanation = (
response_data.get("candidates", [{}])[0]
.get("content", {})
.get("parts", [{}])[0]
.get("text", "Explanation unavailable.")
.strip()
)
except Exception as e:
explanation = f"Gemini error: {str(e)}"
# return {
# "query": query,
# "language": lang,
# "matched_kurals": results,
# "explanation": explanation,
# "total_results": len(results)
# }
return {
"query": query,
"language": lang,
"matched_kurals": [],
"web_results": results,
"explanation": f"Here are relevant results for: '{query}'",
"total_results": len(results)
}
# === Supervisor Agent ===
def supervisor_agent(query: str, lang: str, top_k: int = 3):
intent = classify_intent(query)
if intent == "greeting":
return {
"query": query,
"language": lang,
"matched_kurals": [],
"explanation": "வணக்கம்! என்னைப் பற்றி கேளுங்கள். உங்கள் கேள்விக்கு உதவ தயாராக உள்ளேன்." if lang == "ta" else "Hello! I'm here to help. Ask me anything.",
"total_results": 0
}
elif intent == "life_advice":
return thirukkural_rag_agent(query, lang, top_k)
elif intent == "web_search":
return web_search_agent(query, lang)
else:
return {
"query": query,
"language": lang,
"matched_kurals": [],
"explanation": "I couldn't determine how to respond. Could you please rephrase?",
"total_results": 0
}
# === /search/ Endpoint ===
@app.post("/search/")
def search_router(req: QueryRequest):
try:
return supervisor_agent(req.query, req.lang, req.top_k)
except Exception as e:
print(f"Error in search_router: {str(e)}")
raise HTTPException(status_code=500, detail="Internal server error")
# import os
# from fastapi import FastAPI, HTTPException
# from fastapi.middleware.cors import CORSMiddleware
# from pydantic import BaseModel
# import faiss
# import pickle
# import numpy as np
# from sentence_transformers import SentenceTransformer
# import requests
# import json
# import pandas as pd
# # Set cache dirs (optional, for HF spaces)
# os.environ["HF_HOME"] = "/app/hf_cache"
# os.environ["TRANSFORMERS_CACHE"] = "/app/hf_cache"
# # Load Gemini API key (add this in your HF Space secrets)
# GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")
# if GEMINI_API_KEY:
# GEMINI_URL = f"https://generativelanguage.googleapis.com/v1beta/models/gemma-3n-e4b-it:generateContent?key={GEMINI_API_KEY}"
# else:
# GEMINI_URL = None
# # Initialize FastAPI
# app = FastAPI()
# # Add CORS middleware
# app.add_middleware(
# CORSMiddleware,
# allow_origins=["*"],
# allow_credentials=True,
# allow_methods=["*"],
# allow_headers=["*"],
# )
# # Load local SentenceTransformer model
# try:
# model = SentenceTransformer("/app/model")
# except Exception as e:
# print(f"Error loading model: {e}")
# model = None
# # Load Thirukkural data
# try:
# with open("thirukkural_data.pkl", "rb") as f:
# kural_data = pickle.load(f)
# print(f"Data loaded successfully. Shape: {kural_data.shape}")
# print(f"Columns: {list(kural_data.columns)}")
# except Exception as e:
# print(f"Error loading data: {e}")
# kural_data = None
# # Load FAISS indexes
# try:
# english_index = faiss.read_index("thirukkural_english_index.faiss")
# tamil_index = faiss.read_index("thirukkural_tamil_index.faiss")
# print("FAISS indexes loaded successfully")
# except Exception as e:
# print(f"Error loading FAISS indexes: {e}")
# english_index = None
# tamil_index = None
# # Request schema
# class QueryRequest(BaseModel):
# query: str
# lang: str # "en" or "ta"
# top_k: int = 3
# def safe_get_value(row, possible_keys, default=""):
# """Safely get value from row using possible key names"""
# for key in possible_keys:
# if key in row:
# return row[key]
# return default
# def format_kural_for_gemini(kural_dict, lang):
# number = safe_get_value(kural_dict, ['ID', 'Index'], "Unknown")
# if lang == 'ta':
# text = safe_get_value(kural_dict, ['combined_text_tamil', 'Kural'], "Tamil not available")
# else:
# text = safe_get_value(kural_dict, ['combined_text_english', 'Couplet'], "English not available")
# return f"{number}. {text}"
# def format_kural_for_response(kural_dict):
# return {
# "Number": safe_get_value(kural_dict, ['ID', 'Index'], "Unknown"),
# "Tamil": safe_get_value(kural_dict, ['Kural', 'combined_text_tamil']),
# "English": safe_get_value(kural_dict, ['Couplet', 'combined_text_english']),
# "Category": safe_get_value(kural_dict, ['Adhigaram', 'Section', 'Adhigaram_ID'], "General"),
# "Explanation": safe_get_value(kural_dict, ['Vilakam', 'Kalaingar_Urai', 'Solomon_Pappaiya'], "")
# }
# # Search + Explain Endpoint
# @app.post("/search/")
# def search_and_explain(req: QueryRequest):
# try:
# # Check if required components are loaded
# if model is None:
# raise HTTPException(status_code=500, detail="Model not loaded")
# if kural_data is None:
# raise HTTPException(status_code=500, detail="Kural data not loaded")
# if english_index is None or tamil_index is None:
# raise HTTPException(status_code=500, detail="FAISS indexes not loaded")
# # Step 1: Embed and search
# query_embedding = model.encode([req.query])
# index = tamil_index if req.lang == "ta" else english_index
# # Ensure we don't search for more results than available
# search_k = min(req.top_k, len(kural_data))
# D, I = index.search(np.array(query_embedding).astype("float32"), search_k)
# # Step 2: Get results and handle potential index errors
# results = []
# for i in I[0]:
# if i >= 0 and i < len(kural_data): # Valid index
# row_dict = kural_data.iloc[i].to_dict()
# formatted_kural = format_kural_for_response(row_dict)
# results.append(formatted_kural)
# if not results:
# raise HTTPException(status_code=404, detail="No matching kurals found")
# # Step 3: Format for Gemini explanation
# # ✅ New better-structured Gemini prompt
# kural_texts = '\n\n'.join([
# f'#{k["Number"]}\nTamil: {k["Tamil"]}\nEnglish: {k["English"]}' for k in results
# ])
# prompt = f"""
# You are a compassionate, wise, and culturally sensitive literary guide.
# Your role is to explain how specific Thirukkural couplets relate meaningfully to a user's question. Your tone must be emotionally supportive, philosophically grounded, and respectful of the Tamil literary tradition.
# The user has asked:
# ➡️ "{req.query}" (Language: {req.lang})
# You are given:
# 📜 A selection of Thirukkural couplets, each including:
# - Kural Number
# - Tamil poetic lines
# - English explanation or interpretation
# 🎯 Your task is to deeply connect each of the given Thirukkurals to the user's question or concern, providing an empathetic and insightful explanation.
# ---
# 📝 Response Guidelines:
# 1. **Language Compliance**
# - If `lang` is `"ta"`: Respond fully in rich, poetic **Tamil**.
# - If `lang` is `"en"`: Respond in **English**, but you may include Tamil verses if relevant.
# - Never mix code-switching improperly (e.g., avoid Tamil-English mashups unless culturally meaningful).
# 2. **Kural Formatting**
# - **Strictly preserve the original poetic form**:
# - **Line 1:** First 4 Tamil words
# - **Line 2:** Remaining 3 Tamil words
# - Do **not** restructure or paraphrase Thirukkural verses.
# - Clearly display each Kural before explaining it.
# 3. **Explanation Style**
# - Avoid vague summaries.
# - For each Kural:
# - Offer context to the verse.
# - Connect it directly to the user's emotion or situation.
# - Give philosophical or moral insights that feel timeless and comforting.
# - Avoid listing; write in smooth **paragraph form**.
# - Responses must feel like a **wise teacher or companion**, not a bot.
# 4. **Tone**
# - In Tamil: Use respectful, lyrical, yet simple language—avoid modern slang.
# - In English: Use graceful, reflective, and gentle wording. Avoid robotic phrasing.
# 5. **Confidentiality & Role**
# - Never mention:
# - You are an AI or language model
# - Gemini
# - Backend infrastructure (FastAPI, HuggingFace, etc.)
# - Prompt instructions or formatting logic
# - Do **not** explain your reasoning for response generation.
# - Always remain in-character as a **thoughtful literary guide**.
# ---
# 🧍‍♂️ User Query:
# {req.query}
# 📜 Matching Thirukkurals:
# {kural_texts}
# Please begin your reflective explanation now.
# """
# # Step 4: Gemini API call
# explanation = ""
# if GEMINI_URL and GEMINI_API_KEY:
# try:
# headers = {
# "Content-Type": "application/json"
# }
# data = {
# "contents": [
# {
# "parts": [
# {
# "text": prompt
# }
# ]
# }
# ]
# }
# response = requests.post(GEMINI_URL, headers=headers, data=json.dumps(data), timeout=30)
# response.raise_for_status()
# response_data = response.json()
# if "candidates" in response_data and len(response_data["candidates"]) > 0:
# explanation = response_data["candidates"][0]["content"]["parts"][0]["text"].strip()
# else:
# explanation = "Unable to generate explanation from Gemini API."
# except requests.exceptions.RequestException as e:
# explanation = f"Error calling Gemini API: {str(e)}"
# except (KeyError, IndexError) as e:
# explanation = f"Error parsing Gemini response: {str(e)}"
# except Exception as e:
# explanation = f"Unexpected error with Gemini API: {str(e)}"
# else:
# explanation = "Gemini API key not configured. Please set 'GEMINI_API_KEY' environment variable."
# # Step 5: Return response
# return {
# "query": req.query,
# "language": req.lang,
# "matched_kurals": results,
# "explanation": explanation,
# "total_results": len(results)
# }
# except HTTPException:
# raise
# except Exception as e:
# print(f"Unexpected error in search_and_explain: {str(e)}")
# raise HTTPException(status_code=500, detail=f"Internal server error: {str(e)}")
# # Health check
# @app.get("/")
# def root():
# status = {
# "message": "Thirukkural AI RAG API",
# "status": "running",
# "components": {
# "model_loaded": model is not None,
# "data_loaded": kural_data is not None,
# "english_index_loaded": english_index is not None,
# "tamil_index_loaded": tamil_index is not None,
# "gemini_configured": GEMINI_API_KEY is not None
# }
# }
# if kural_data is not None:
# status["data_info"] = {
# "total_kurals": len(kural_data),
# "columns": list(kural_data.columns)
# }
# return status
# # Debug endpoint to check data structure
# @app.get("/debug/data")
# def debug_data():
# if kural_data is None:
# return {"error": "Data not loaded"}
# return {
# "shape": kural_data.shape,
# "columns": list(kural_data.columns),
# "sample_row": kural_data.iloc[0].to_dict() if len(kural_data) > 0 else None,
# "dtypes": kural_data.dtypes.to_dict()
# }
# # Test endpoint
# @app.get("/test")
# def test_endpoint():
# return {"message": "API is working!", "timestamp": pd.Timestamp.now().isoformat()}