app.py

import os
import gradio as gr
import numpy as np
import faiss

from groq import Groq
from pypdf import PdfReader
from sentence_transformers import SentenceTransformer
from langchain_text_splitters import RecursiveCharacterTextSplitter

# =====================================================
# Configuration
# =====================================================
RELEVANCE_THRESHOLD = 1.2   # lower = stricter relevance

# =====================================================
# Initialize Groq Client
# =====================================================
client = Groq(api_key=os.environ.get("RAG-GROQ"))
# =====================================================
# Load Embedding Model
# =====================================================
embedding_model = SentenceTransformer("all-MiniLM-L6-v2")

# =====================================================
# Global Vector Store
# =====================================================
vector_store = None
stored_chunks = []

# =====================================================
# PDF Processing Function
# =====================================================
def process_pdf(pdf_file):
    global vector_store, stored_chunks

    reader = PdfReader(pdf_file)
    full_text = ""

    for page in reader.pages:
        if page.extract_text():
            full_text += page.extract_text() + "\n"

    splitter = RecursiveCharacterTextSplitter(
        chunk_size=500,
        chunk_overlap=100
    )
    chunks = splitter.split_text(full_text)

    embeddings = embedding_model.encode(chunks)

    dimension = embeddings.shape[1]
    vector_store = faiss.IndexFlatL2(dimension)
    vector_store.add(np.array(embeddings))

    stored_chunks = chunks

    return "✅ PDF processed successfully. You can now ask questions."

# =====================================================
# Question Answering Function
# =====================================================
def answer_question(question):
    if vector_store is None:
        return "⚠️ Please upload and process a PDF first."

    question_embedding = embedding_model.encode([question])
    distances, indices = vector_store.search(
        np.array(question_embedding), k=3
    )

    avg_distance = distances[0].mean()

    context = ""
    for idx in indices[0]:
        context += stored_chunks[idx] + "\n"

    # Relevance feedback
    if avg_distance > RELEVANCE_THRESHOLD:
        relevance_note = (
            "⚠️ **Note:** This question is not directly answered in the document.\n"
            "The response below is based on loosely related context.\n\n"
        )
    else:
        relevance_note = ""

    prompt = f"""
You are an honest and careful AI assistant.

Instructions:
- Answer ONLY using the provided context.
- If the answer is not explicitly stated, say:
  "This is not directly mentioned in the document, but based on related context..."

Context:
{context}

Question:
{question}
"""

    response = client.chat.completions.create(
        model="llama-3.3-70b-versatile",
        messages=[
            {"role": "user", "content": prompt}
        ]
    )

    return relevance_note + response.choices[0].message.content

# =====================================================
# Gradio UI
# =====================================================
with gr.Blocks() as app:
    gr.Markdown("## 📄 RAG-based PDF Question Answering (Groq + FAISS)")
    gr.Markdown(
        "Upload a PDF and ask questions. "
        "The system will clearly tell you if an answer is not directly mentioned."
    )

    pdf_file = gr.File(label="Upload PDF", file_types=[".pdf"])
    process_btn = gr.Button("Process PDF")
    status_box = gr.Textbox(label="Status", interactive=False)

    question_box = gr.Textbox(label="Ask a Question")
    answer_box = gr.Textbox(label="Answer", lines=8)

    process_btn.click(
        process_pdf,
        inputs=pdf_file,
        outputs=status_box
    )

    question_box.submit(
        answer_question,
        inputs=question_box,
        outputs=answer_box
    )

app.launch()