app.py

# app.py  –  PolicyGPT 🇮🇳  (error-free)

import pathlib, tempfile, textwrap, traceback, requests
from functools import lru_cache

import gradio as gr
from langchain_community.embeddings import HuggingFaceEmbeddings  # new import
from langchain_community.vectorstores import FAISS               # new import
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.docstore.document import Document
from transformers import pipeline
import pypdf

# ---------- 1. Policy corpus ----------
POLICY_URLS = {
    "DPDP Act 2023":
        "https://www.meity.gov.in/static/uploads/2024/06/2bf1f0e9f04e6fb4f8fef35e82c42aa5.pdf",
    "Responsible AI (NITI Aayog)":
        "https://www.niti.gov.in/sites/default/files/2021-08/Part2-Responsible-AI-12082021.pdf",
    # … keep the rest …
}

INDUSTRY_MAP = {
    "Health Care": ["DPDP Act 2023", "Responsible AI (NITI Aayog)"],
    "All": list(POLICY_URLS.keys()),
}

# ---------- 2. Helpers ----------
def download(url: str, path: pathlib.Path):
    if not path.exists():
        path.parent.mkdir(parents=True, exist_ok=True)
        r = requests.get(url, timeout=120)
        r.raise_for_status()
        path.write_bytes(r.content)
    return path

def pdf_text(path: pathlib.Path) -> str:
    out = []
    with path.open("rb") as f:
        for p in pypdf.PdfReader(f).pages:
            out.append(p.extract_text() or "")
    return "\n".join(out)

@lru_cache(maxsize=1)
def store(srcs=tuple(POLICY_URLS.keys())):
    splitter = RecursiveCharacterTextSplitter(chunk_size=1024, chunk_overlap=128)
    docs = []
    for name in srcs:
        path = pathlib.Path(tempfile.gettempdir()) / "policygpt" / f"{name}.pdf"
        try:
            for chunk in splitter.split_text(pdf_text(download(POLICY_URLS[name], path))):
                docs.append(Document(page_content=chunk, metadata={"src": name}))
        except Exception as e:
            print("❌", name, e)
    embed = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
    return FAISS.from_documents(docs, embed)

GEN = pipeline(                 # ✅ use text2text-generation
    "text2text-generation",
    model="google/flan-t5-small",
    max_new_tokens=200,
    do_sample=False,
)

def rag(q: str, industry: str):
    db = store(tuple(POLICY_URLS.keys()) if industry == "All" else tuple(INDUSTRY_MAP[industry]))
    ctx = "\n\n".join(d.page_content for d in db.similarity_search(q, k=4))[:3500]
    prompt = textwrap.dedent(f"""
        You are PolicyGPT. Using CONTEXT, answer QUESTION (≤150 words)
        and cite source names in brackets. If unsure, say I don’t know.

        CONTEXT:
        {ctx}

        QUESTION: {q}
        ANSWER:
    """)
    try:
        return GEN(prompt)[0]["generated_text"].strip() or "I don’t know."
    except Exception as e:
        return f"⚠️ Generation error: {e}"

def risk(text: str):
    low = text.lower()
    if any(k in low for k in ("violation", "prohibited", "penalty")):
        return "High"
    if any(k in low for k in ("must", "should", "shall")):
        return "Medium"
    return "Low"

# ---------- 3. Gradio UI ----------
def answer_fn(q, ind):
    a = rag(q, ind)
    return a, f"**Estimated compliance risk:** {risk(a)}", gr.update(interactive=True)

with gr.Blocks(title="PolicyGPT 🇮🇳") as demo:
    gr.Markdown("# PolicyGPT 🇮🇳 — ask about AI & Data-governance laws")
    ind = gr.Dropdown(list(INDUSTRY_MAP.keys()), label="Select industry", value="All")
    qbox = gr.Textbox(lines=2, label="Your question",
                      placeholder="e.g. What PII rules apply to hospitals?")
    ask = gr.Button("Ask")
    ans = gr.Markdown(); rsk = gr.Markdown()

    # Disable button while processing
    ask.click(lambda: gr.update(interactive=False), None, ask, queue=False)
    ask.click(answer_fn, [qbox, ind], [ans, rsk, ask])
    qbox.submit(lambda: gr.update(interactive=False), None, ask, queue=False)
    qbox.submit(answer_fn, [qbox, ind], [ans, rsk, ask])

# Gradio 4+: no concurrency_count param
if __name__ == "__main__":
    demo.queue().launch()