functions.py

import os
import tempfile
import logging
import streamlit as st
from langchain.vectorstores import FAISS
from langchain.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.embeddings import HuggingFaceEmbeddings
from huggingface_hub import InferenceClient
from sentence_transformers import CrossEncoder
from dotenv import load_dotenv
load_dotenv()
import datetime

# Initialize Logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Retrieve HF API Key and initialize HF client
# HF_API_KEY = os.getenv("HF_API_KEY")
# if not HF_API_KEY:
#     raise ValueError("Hugging Face API key not found. Please set HF_API_KEY in the .env file.")
client = InferenceClient()

FALLBACK_MESSAGE = "Sorry, I didn’t understand your question. Do you want to connect with a live agent?"
FEEDBACK_ERROR = "Our servers are busy. Please try again later."

# Retrieval parameters
FAISS_TOP_K = 20        # Number of top chunks to retrieve from FAISS
RERANK_TOP_K = 6      # Number of top chunks to use for generating the answer
CHUNK_SIZE = 1500
CHUNK_OVERLAP = 100

# Threshold for re-ranker score
MIN_SCORE_THRESHOLD = -8.0

# PDF upload constraints
MAX_PDF_SIZE = 200 * 1024 * 1024


def build_vectorstore(pdf_file):
    """Loads and processes a single PDF file, then builds a new FAISS vectorstore."""

    if pdf_file is None:
        logger.warning("No PDF file provided to build_vectorstore.")
        return 0, {}

    # Clear existing vectorstore if any
    st.session_state.vectorstore = None

    # Check PDF size
    if pdf_file.size > MAX_PDF_SIZE:
        st.error(f"PDF exceeds the maximum allowed size of {MAX_PDF_SIZE / (1024 * 1024)} MB.")
        logger.warning("Uploaded PDF exceeds the maximum allowed size.")
        return 0, {}

    try:
        with tempfile.NamedTemporaryFile(delete=False, suffix=".pdf") as tmp_file:
            tmp_file.write(pdf_file.getbuffer())
            tmp_path = tmp_file.name
            logger.info(f"Temporary PDF saved at {tmp_path}")

        loader = PyPDFLoader(tmp_path)
        docs = loader.load()

        # Extract metadata
        metadata = loader.metadata if hasattr(loader, 'metadata') else {}
        os.remove(tmp_path)
        logger.info("PDF loaded and temporary file removed.")

        text_splitter = RecursiveCharacterTextSplitter(
            chunk_size=CHUNK_SIZE,
            chunk_overlap=CHUNK_OVERLAP,
            length_function=len
        )
        chunks = text_splitter.split_documents(docs)
        num_chunks = len(chunks)
        logger.info(f"PDF split into {num_chunks} chunks.")

        # Build FAISS vectorstore
        st.session_state.vectorstore = FAISS.from_documents(chunks, st.session_state.embeddings)
        logger.info("FAISS vectorstore built successfully.")
        return num_chunks, metadata

    except Exception as e:
        st.error("An error occurred while processing the PDF.")
        logger.error(f"Error in build_vectorstore: {e}", exc_info=True)
        return 0, {}


def retrieve(question, k=FAISS_TOP_K):
    """Retrieves FAISS_TOP_K number of documents based on vector similarity search from database."""

    vs = st.session_state.vectorstore
    retrieved_docs = vs.similarity_search(question, k=k)
    if not retrieved_docs:
        logger.info("No documents retrieved for the question.")
        return []
    logger.info(f"Retrieved {len(retrieved_docs)} documents from FAISS.")
    return retrieved_docs

def sanitize_input(user_input):
    """Validates user input. """
    if not user_input:
        st.error("Input cannot be empty.")
        return False

    if len(user_input) > 1000:
        st.error("Input exceeds the maximum allowed length of 1000 characters.")
        return False
    return True

def rerank_documents(question, documents):
    """Re-rank the retrieved documents using a cross-encoder reranker and returns RERANK_TOP_K number of documents according to relevance order"""
    if not documents:
        logger.info("No documents to rerank.")
        return []

    reranker = st.session_state.reranker
    model_inputs = [(question, doc.page_content) for doc in documents]

    try:
        rerank_scores = reranker.predict(model_inputs, batch_size=16)
        logger.info("Re-ranker scores generated successfully.")
    except Exception as e:
        logger.error(f"Error during reranking: {e}")
        return []

    # Pair docs with reranker scores and sort by reranker score descending
    reranked_pairs = sorted(zip(documents, rerank_scores), key=lambda x: x[1], reverse=True)
    logger.info("Documents reranked successfully.")
    return reranked_pairs

def call_hf_api(prompt):
    """Calls the Hugging Face InferenceClient API for chat completion and returns the generated response from the LLM"""
    messages = [
        {"role": "system", "content": (
             "You are a knowledgeable and helpful assistant. When answering, please:\n"
            "- Use headings and subheadings for organization. Utilize bullet points or numbered lists for clarity.\n"
            "- Provide factual information based solely on the provided context.\n"
            "- Mention that schedules or dates may change if applicable.\n"
            "- Clearly list accessibility or facilities details when relevant.\n"
            "- Your output should only contain the relevant answer to the question.\n "
            "- If the answer is not found in the context or the user input is not a valid question, respond only with: 'Sorry, I didn’t understand your question. Do you want to connect with a live agent?'.\n"
            "- End responses by asking if the user needs further assistance.\n"
            "- Maintain a professional and concise tone."
        )},
        {"role": "user", "content": prompt}
    ]

    try:
        completion = client.chat.completions.create(
            model="mistralai/Mistral-7B-Instruct-v0.3",
            messages=messages,
            max_tokens=2000
        )
        response = completion.choices[0].message["content"].strip()
        logger.info("LLM response generated successfully.")
        return response
    except Exception as e:
        logger.error(f"Error during LLM call: {e}")
        return FALLBACK_ERROR


def get_answer(question):
    """Main Logical block. Question -> Retrives chunks -> Re-ranks them -> Gives them to LLM to answer"""

    if len(question.split()) < 3:
        # Question has fewer than 3 words. Not a valid question.
        return FALLBACK_MESSAGE

    logger.info(f"Processing question: {question}")
    retrieved_docs = retrieve(question, k=FAISS_TOP_K)
    if not retrieved_docs:
        logger.info("No relevant documents found. Triggering fallback.")
        return FALLBACK_MESSAGE

    reranked_pairs = rerank_documents(question, retrieved_docs)
    if not reranked_pairs:
        logger.info("Re-ranking failed or no documents after reranking. Triggering fallback.")
        return FALLBACK_MESSAGE

    # Check the top re-ranker score
    top_doc, top_score = reranked_pairs[0]
    logger.info(f"Top re-ranker score: {top_score}")
    if top_score < MIN_SCORE_THRESHOLD:
        logger.info("Top re-ranker score below threshold. Triggering fallback.")
        return FALLBACK_MESSAGE

    # Select top 5 chunks for the final context
    top_chunks = reranked_pairs[:RERANK_TOP_K]
    combined_context = "\n\n".join([doc.page_content for doc, _ in top_chunks])

    # Construct the final prompt
    final_prompt = (
        f"Context:\n{combined_context}\n\n"
        f"User Question:\n{question}\n\n"
        "Instructions:\n"
        "Answer the user's question based only on the provided context. Ensure your response is:\n"
        "- Clear and concise.\n"
        "- Well-structured with headings or subheadings.\n"
        "- Formatted using bullet points, numbered lists, or short paragraphs.\n"
        "- Inclusive of logical inferences if implied by the context.\n"
        "- Free from unrelated information.\n"
        "- Concluding with an offer for further assistance.\n\n"
        "Answer:"
    )

    final_answer = call_hf_api(final_prompt).strip()
    logger.info("Final answer generated.")
    return final_answer if final_answer else FALLBACK_MESSAGE