utils.py

import os
from dotenv import load_dotenv
from transformers import pipeline
from PyPDF2 import PdfReader
import numpy as np


#Load .env file
load_dotenv()

#Getting the model name from environment (default if not found)
MODEL_NAME = os.getenv("MODEL_NAME", "google/flan-t5-small")

def load_handbook_text(pdf_path):
    reader = PdfReader(pdf_path)
    return " ".join(page.extract_text() for page in reader.pages if page.extract_text())

def split_text(text, max_len=800):
    words = text.split()
    return [" ".join(words[i:i+max_len]) for i in range(0, len(words), max_len)]

def find_relevant_chunk(query, chunks):
    # Very lightweight keyword overlap retrieval
    q_tokens = set(query.lower().split())
    scores = [len(q_tokens & set(c.lower().split())) for c in chunks]
    return chunks[int(np.argmax(scores))]

def answer_question(query, pdf_path="STUDENT-HANDBOOK-2021-EDITION.pdf"):
    """
    Answering questions based on the Student Handbook with concise and relevant responses.
    """
    text = load_handbook_text(pdf_path)
    chunks = split_text(text)
    context = find_relevant_chunk(query, chunks)

    #Loading the model, which is cached automatically by Hugging Face.
    qa = pipeline(
        "text2text-generation",
        model=MODEL_NAME,
        tokenizer=MODEL_NAME
    )

    #Improved prompts for more precise answers.
    prompt = (
        f"Use only the context below to answer concisely and clearly.\n\n"
        f"Question: {query}\n\n"
        f"Context:\n{context[:700]}\n\n"
        f"Answer in 2-3 sentences only."
    )

    result = qa(prompt, max_new_tokens=100, temperature=0.3, num_beams=4)

    #Cleaning up the output by trimming the redundant text
    answer = result[0]["generated_text"].strip()
    return answer