src/chatbot.py

import os

# --- add these 3 lines before anything Hugging Face runs ---
os.environ["HF_HOME"] = "/tmp/hf_cache"
os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf_cache"
os.environ["SENTENCE_TRANSFORMERS_HOME"] = "/tmp/hf_cache"

from typing import List
from dotenv import load_dotenv

from langchain_groq import ChatGroq
from langchain.schema import HumanMessage, AIMessage
from langchain_community.vectorstores import FAISS
from langchain_huggingface import HuggingFaceEmbeddings
from langchain.prompts import PromptTemplate
from langchain.chains import RetrievalQA

# ---------------------------
# Load environment variables
# ---------------------------
load_dotenv()
GROQ_API_KEY = os.getenv("GROQ_API_KEY")

# ---------------------------
# Settings / Tuning
# ---------------------------
DB_FAISS_PATH = "vectorStore"
EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
K = 5                    # how many candidates to check for pre-filter
MAX_DISTANCE = 1.0       # FAISS distance threshold (lower = better). 
MAX_CHAT_HISTORY = 50    # cap chat history to avoid unbounded growth

# ---------------------------
# Load FAISS VectorStore
# ---------------------------
embeddings = HuggingFaceEmbeddings(
    model_name=EMBED_MODEL,
    cache_folder="/tmp/hf_cache"   # <--- new
)
db = FAISS.load_local(DB_FAISS_PATH, embeddings, allow_dangerous_deserialization=True)

# ---------------------------
# ChatBot Class
# ---------------------------
class RAGChatBot:
    def __init__(self):
        # LLM
        if not GROQ_API_KEY:
            raise ValueError("GROQ_API_KEY not set in environment")
        self.llm = ChatGroq(
            groq_api_key=GROQ_API_KEY,
            model="llama-3.1-8b-instant",
            temperature=0
        )
        self.chat_history: List = []

        # Retriever used by RetrievalQA (kept, but we will pre-filter before calling the chain)
        self.retriever = db.as_retriever(search_type="similarity", search_kwargs={"k": 3})

        # Custom Prompt (dynamic fallback included)
        custom_prompt = """
Use the following context to answer the user’s question.
If the answer cannot be found in the context, reply exactly with:
"I'm trained only on Odisha disaster management reports (i.e,OSDMA, NDMA, IMD, Research papers). I don't have any information about: '{question}'"

Context:
{context}

Question:
{question}

Answer:
"""
        self.prompt = PromptTemplate(template=custom_prompt, input_variables=["context", "question"])

        # Retrieval QA Chain (keeps structured QA behavior)
        self.qa_chain = RetrievalQA.from_chain_type(
            llm=self.llm,
            retriever=self.retriever,
            return_source_documents=True,
            chain_type_kwargs={"prompt": self.prompt}
        )

    # ---------------------------
    # NEW: Rewrite function
    # ---------------------------
    def rewrite_query(self, user_input: str) -> str:
        """Rewrite query into formal disaster-management style language using LLM."""
        rewrite_prompt = f"""
        Rewrite the following user query into clear, formal disaster management language
        as used in government reports (OSDMA, NDMA, IMD).
        If it is not disaster-related, just return it unchanged.

        Query: {user_input}
        """
        try:
            response = self.llm.invoke([HumanMessage(content=rewrite_prompt)])
            return response.content.strip()
        except Exception as e:
            print("⚠ Rewrite error:", e)
            return user_input  # fallback to original

    def _prefilter_by_distance(self, query: str, k: int = K, max_distance: float = MAX_DISTANCE) -> bool:
        """Check if query is in-domain using FAISS distance."""
        results = db.similarity_search_with_score(query, k=k)
        if not results:
            return False
        best_score = results[0][1]  # (Document, score)
        return best_score <= max_distance

    def chat(self, user_input: str) -> str:
        # 1) Rewrite user query
        rewritten_query = self.rewrite_query(user_input)
        # print(f"[debug] rewritten query: {rewritten_query}")

        # 2) Quick in-domain prefilter
        try:
            in_domain = self._prefilter_by_distance(rewritten_query)
        except Exception as e:
            print("⚠ prefilter error:", e)
            in_domain = True

        if not in_domain:
            return (
                f"I’m trained only on Odisha disaster management reports "
                f"(OSDMA, NDMA, IMD, research). I don’t have any information about: '{user_input}'."
            )

        # 3) Retrieval + QA
        try:
            response = self.qa_chain.invoke({"query": rewritten_query})
            answer = response.get("result") if isinstance(response, dict) else str(response)
        except Exception as e:
            print("⚠ LLM / chain error:", e)
            answer = "Sorry, I encountered an error while generating the answer."

        # 4) Update memory (bounded)
        self.chat_history.append(HumanMessage(content=user_input))
        self.chat_history.append(AIMessage(content=answer))
        if len(self.chat_history) > MAX_CHAT_HISTORY * 2:
            self.chat_history = self.chat_history[-MAX_CHAT_HISTORY * 2 :]

        return answer


# ---------------------------
# Run Chatbot (CLI)
# ---------------------------
if __name__ == "__main__":
    bot = RAGChatBot()
    print("🤖 Odisha Disaster Management ChatBot ready! Type 'exit' to quit.")
    while True:
        query = input("You: ")
        if query.lower() in ["exit", "quit"]:
            break
        print("Bot:", bot.chat(query))