app.py

import os
import gradio as gr
import faiss
import numpy as np
from groq import Groq
from sentence_transformers import SentenceTransformer
from PyPDF2 import PdfReader
import requests
from io import BytesIO

# -------- SETTINGS --------
EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
INDEX_FILE = "faiss_index.bin"
CHUNKS_FILE = "chunks.npy"
CHUNK_SIZE = 500
TOP_K = 3

# ASME PDF URLs (fixed for this app)
PDF_URLS = [
    "https://www.asme.org/wwwasmeorg/media/resourcefiles/aboutasme/who%20we%20are/standards_and_certification/asme_codes_and_standards-examples_of_use_for_mechanical_engineering_students.pdf",
    "https://www.asme.org/wwwasmeorg/media/resourcefiles/campaigns/marketing/2012/the-state-of-mechanical-engineering-survey.pdf"
]

# Load embedding model
embedder = SentenceTransformer(EMBED_MODEL)
embed_dim = embedder.get_sentence_embedding_dimension()

# Initialize or load FAISS + chunks
if os.path.exists(INDEX_FILE) and os.path.exists(CHUNKS_FILE):
    index = faiss.read_index(INDEX_FILE)
    chunks = np.load(CHUNKS_FILE, allow_pickle=True).tolist()
else:
    index = faiss.IndexFlatL2(embed_dim)
    chunks = []

# Groq client
client = Groq(api_key=os.environ.get("GROQ_API_KEY"))

# -------- FUNCTIONS --------
def pdf_url_to_chunks(pdf_url, chunk_size=CHUNK_SIZE):
    resp = requests.get(pdf_url)
    resp.raise_for_status()
    pdf_bytes = BytesIO(resp.content)
    reader = PdfReader(pdf_bytes)
    text_all = ""
    for page in reader.pages:
        page_text = page.extract_text()
        if page_text:
            text_all += page_text + " "
    words = text_all.split()
    return [
        " ".join(words[i:i+chunk_size])
        for i in range(0, len(words), chunk_size)
        if len(words[i:i+chunk_size]) > 20
    ]

def build_vector_db():
    global index, chunks
    if len(chunks) > 0:
        return f"✅ Knowledge base already built with {len(chunks)} chunks."
    new_chunks = []
    for url in PDF_URLS:
        url_chunks = pdf_url_to_chunks(url)
        new_chunks.extend(url_chunks)
    embeddings = embedder.encode(new_chunks, convert_to_numpy=True)
    index.add(embeddings)
    chunks.extend(new_chunks)
    faiss.write_index(index, INDEX_FILE)
    np.save(CHUNKS_FILE, np.array(chunks, dtype=object))
    return f"✅ Knowledge base built with {len(chunks)} chunks."

def retrieve_chunks(query, top_k=TOP_K):
    if len(chunks) == 0:
        return []
    query_vec = embedder.encode([query], convert_to_numpy=True)
    distances, indices = index.search(query_vec, top_k)
    return [chunks[i] for i in indices[0] if i < len(chunks)]

def ask_with_rag(query):
    if len(chunks) == 0:
        return "⚠️ Please build the knowledge base first."
    retrieved = retrieve_chunks(query)
    context = "\n\n".join(retrieved)
    prompt = f"""You are an assistant knowledgeable in ASME Standards.
Context:
{context}

User Query:
{query}

Answer using the context. If not found, reply: “I could not find it in the provided ASME documents.”"""

    chat_completion = client.chat.completions.create(
        messages=[{"role": "user", "content": prompt}],
        model="llama-3.3-70b-versatile",
    )
    return chat_completion.choices[0].message.content

# -------- GRADIO UI --------
with gr.Blocks() as demo:
    gr.Markdown("# 📘 ASME RAG Assistant")
    gr.Markdown("This app is powered by FAISS + Groq LLM. Knowledge base is built from 2 official ASME PDFs.")

    build_btn = gr.Button("Build Knowledge Base")
    build_status = gr.Textbox(label="Status")

    build_btn.click(build_vector_db, inputs=None, outputs=build_status)

    query_input = gr.Textbox(label="Ask a Question", lines=1)
    answer_output = gr.Textbox(label="Answer", lines=5)
    query_btn = gr.Button("Ask")

    query_btn.click(ask_with_rag, inputs=query_input, outputs=answer_output)

demo.launch()