app.py

# app.py
# RAG app for chatting with research papers (optimized for Hugging Face Spaces)

import os, sys, subprocess, re, json, uuid, gc
from typing import List, Dict, Tuple

# -----------------------------
# Auto-install deps if missing
# -----------------------------
def ensure(pkg, pip_name=None):
    try:
        __import__(pkg)
    except ImportError:
        subprocess.check_call([sys.executable, "-m", "pip", "install", pip_name or pkg])

ensure("torch")
ensure("transformers")
ensure("accelerate")
ensure("gradio")
ensure("faiss", "faiss-cpu")
ensure("sentence_transformers", "sentence-transformers")
ensure("pypdf")
ensure("docx", "python-docx")

import torch
from transformers import (
    AutoTokenizer, 
    AutoModelForCausalLM, 
    TextIteratorStreamer
)
from sentence_transformers import SentenceTransformer
import faiss, gradio as gr
from pypdf import PdfReader

# -----------------------------
# Config
# -----------------------------
DATA_DIR = "rag_data"
os.makedirs(DATA_DIR, exist_ok=True)
INDEX_PATH = os.path.join(DATA_DIR, "faiss.index")
DOCS_PATH  = os.path.join(DATA_DIR, "docs.jsonl")

# Default Models
default_emb_model = "allenai/specter2_base"
default_llm_model = "microsoft/Phi-3-mini-4k-instruct"

EMB_MODEL_ID = os.environ.get("EMB_MODEL_ID", default_emb_model)
LLM_MODEL_ID = os.environ.get("LLM_MODEL_ID", default_llm_model)

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# -----------------------------
# File loaders
# -----------------------------
def read_txt(path):
    return open(path, "r", encoding="utf-8", errors="ignore").read()

def read_pdf(path):
    r = PdfReader(path)
    return "\n".join([p.extract_text() or "" for p in r.pages])

def read_docx(path):
    import docx
    d = docx.Document(path)
    return "\n".join([p.text for p in d.paragraphs])

def load_file(path):
    ext = os.path.splitext(path)[1].lower()
    if ext in [".txt", ".md"]:
        return read_txt(path)
    if ext == ".pdf":
        return read_pdf(path)
    if ext == ".docx":
        return read_docx(path)
    return read_txt(path)

# -----------------------------
# Chunking
# -----------------------------
def normalize_ws(s: str):
    return re.sub(r"\s+", " ", s).strip()

def chunk_text(text, chunk_size=900, overlap=150):
    text = normalize_ws(text)
    chunks = []
    for i in range(0, len(text), chunk_size - overlap):
        chunks.append(text[i:i+chunk_size])
    return chunks

# -----------------------------
# VectorStore
# -----------------------------
class VectorStore:
    def __init__(self, emb_model):
        self.emb_model = emb_model
        self.dim = emb_model.get_sentence_embedding_dimension()
        if os.path.exists(INDEX_PATH):
            self.index = faiss.read_index(INDEX_PATH)
            self.meta = [json.loads(l) for l in open(DOCS_PATH, "r", encoding="utf-8")]
        else:
            self.index = faiss.IndexFlatIP(self.dim)
            self.meta = []

    def _embed(self, texts):
        embs = self.emb_model.encode(texts, convert_to_tensor=True, normalize_embeddings=True)
        return embs.cpu().numpy()

    def add(self, chunks, source):
        if not chunks: return 0
        embs = self._embed(chunks)
        faiss.normalize_L2(embs)
        self.index.add(embs)
        recs = []
        for c in chunks:
            rec = {"id": str(uuid.uuid4()), "source": source, "text": c}
            self.meta.append(rec)
            recs.append(json.dumps(rec))
        with open(DOCS_PATH, "a", encoding="utf-8") as f:
            f.write("\n".join(recs) + "\n")
        faiss.write_index(self.index, INDEX_PATH)
        return len(chunks)

    def search(self, query, k=5):
        q = self._embed([query])
        faiss.normalize_L2(q)
        D, I = self.index.search(q, k)
        return [(float(d), self.meta[i]) for d, i in zip(D[0], I[0]) if i != -1]

    def clear(self):
        self.index = faiss.IndexFlatIP(self.dim)
        self.meta = []
        if os.path.exists(INDEX_PATH): os.remove(INDEX_PATH)
        if os.path.exists(DOCS_PATH): os.remove(DOCS_PATH)

# -----------------------------
# Load models
# -----------------------------
print(f"[RAG] Loading embeddings: {EMB_MODEL_ID}")
EMB = SentenceTransformer(EMB_MODEL_ID, device=DEVICE)
VEC = VectorStore(EMB)

print(f"[RAG] Loading LLM: {LLM_MODEL_ID}")
bnb_config = None
if DEVICE == "cuda":
    from transformers import BitsAndBytesConfig
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_use_double_quant=True,
        bnb_4bit_quant_type="nf4"
    )

TOKENIZER = AutoTokenizer.from_pretrained(LLM_MODEL_ID, use_fast=True, trust_remote_code=True)
LLM = AutoModelForCausalLM.from_pretrained(
    LLM_MODEL_ID,
    device_map="auto",
    quantization_config=bnb_config,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
    low_cpu_mem_usage=True,
    trust_remote_code=True,
)

# -----------------------------
# Prompt + Generate
# -----------------------------
SYSTEM_PROMPT = "You are a helpful assistant. Use the provided context from research papers to answer questions."

def build_prompt(query, history, retrieved):
    ctx = "\n\n".join([f"[{i+1}] {m['text']}" for i, (_, m) in enumerate(retrieved)])
    # Try to use chat template if available
    if hasattr(TOKENIZER, "apply_chat_template"):
        messages = [{"role": "system", "content": SYSTEM_PROMPT + "\nContext:\n" + ctx}]
        for u, a in history[-3:]:
            messages.append({"role": "user", "content": u})
            messages.append({"role": "assistant", "content": a})
        messages.append({"role": "user", "content": query})
        return TOKENIZER.apply_chat_template(messages, tokenize=False)
    else:
        # Fallback manual prompt
        hist = "".join([f"<user>{u}</user><assistant>{a}</assistant>" for u, a in history[-3:]])
        return f"<system>{SYSTEM_PROMPT}\nContext:\n{ctx}</system>{hist}<user>{query}</user><assistant>"

@torch.inference_mode()
def generate_answer(prompt, temperature=0.3, max_new_tokens=512):
    streamer = TextIteratorStreamer(TOKENIZER, skip_prompt=True, skip_special_tokens=True)
    inputs = TOKENIZER([prompt], return_tensors="pt").to(LLM.device)
    kwargs = dict(
        **inputs,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        do_sample=temperature > 0,
        streamer=streamer
    )
    import threading
    t = threading.Thread(target=LLM.generate, kwargs=kwargs)
    t.start()
    out = ""
    for token in streamer:
        out += token
        yield out
    t.join()

# -----------------------------
# Gradio UI
# -----------------------------
def ui_ingest(files, chunk_size, overlap):
    total = 0
    names = []
    for f in files or []:
        text = load_file(f.name)
        chunks = chunk_text(text, chunk_size, overlap)
        n = VEC.add(chunks, os.path.basename(f.name))
        total += n; names.append(f.name)
    return f"Added {total} chunks", "\n".join(names) or "—", VEC.index.ntotal

def ui_clear():
    VEC.clear()
    gc.collect()
    return "Index cleared", "—", 0

def ui_chat(msg, history, top_k, temperature, max_tokens):
    if not msg.strip():
        return history, ""
    retrieved = VEC.search(msg, top_k)
    prompt = build_prompt(msg, history, retrieved)
    reply = ""
    for partial in generate_answer(prompt, temperature, max_tokens):
        reply = partial
        yield history + [(msg, reply)], ""
    yield history + [(msg, reply)], ""

with gr.Blocks() as demo:
    gr.Markdown("# 🔎📚 Research Paper RAG Chat (Phi-3-mini + Specter2)")
    with gr.Row():
        with gr.Column(scale=2):
            chatbot = gr.Chatbot(height=500)
            msg = gr.Textbox(placeholder="Ask a question...")
            with gr.Row():
                send = gr.Button("Send", variant="primary")
                clearc = gr.Button("Clear Chat")
        with gr.Column():
            files = gr.File(label="Upload PDFs/DOCX/TXT", file_types=[".pdf", ".docx", ".txt", ".md"], file_count="multiple")
            chunk_size = gr.Slider(200,2000,900,step=50,label="Chunk Size")
            overlap = gr.Slider(0,400,150,step=10,label="Overlap")
            ingest_btn = gr.Button("Index Documents")
            status = gr.Textbox(label="Status", value="—")
            added = gr.Textbox(label="Files", value="—")
            total = gr.Number(label="Total Chunks", value=VEC.index.ntotal)
            clear_idx = gr.Button("Clear Index", variant="stop")
            top_k = gr.Slider(1,10,5,1,label="Top-K")
            temperature = gr.Slider(0.0,1.5,0.3,0.1,label="Temperature")
            max_tokens = gr.Slider(64,2048,512,64,label="Max New Tokens")

    ingest_btn.click(ui_ingest, [files, chunk_size, overlap], [status, added, total])
    clear_idx.click(ui_clear, [], [status, added, total])
    send.click(ui_chat, [msg, chatbot, top_k, temperature, max_tokens], [chatbot, msg])
    msg.submit(ui_chat, [msg, chatbot, top_k, temperature, max_tokens], [chatbot, msg])
    clearc.click(lambda: ([], ""), [], [chatbot, msg])

if __name__ == "__main__":
    demo.queue().launch(server_name="0.0.0.0", server_port=int(os.getenv("PORT", 7860)))