app.py

import os

# =====================================================
# OPTION A: Use ephemeral /tmp cache to avoid 50 GB quota
# ======================================================
os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf_cache"
os.environ["HF_HOME"] = "/tmp/hf_home"
os.environ["HF_DATASETS_CACHE"] = "/tmp/hf_datasets"
os.environ["HF_MODULES_CACHE"] = "/tmp/hf_modules"
# ======================================================

import torch
import gradio as gr
import faiss
import numpy as np
from transformers import AutoTokenizer, AutoModelForCausalLM
from sentence_transformers import SentenceTransformer

# ------------------------------------------------------
# 1️⃣ Model setup
# ------------------------------------------------------
GEN_MODEL = "hackergeek/qwen3-harrison-rag"      # main generation model
EMB_MODEL = "sentence-transformers/all-MiniLM-L6-v2"  # embedding model (lightweight)

tokenizer = AutoTokenizer.from_pretrained(GEN_MODEL)
model = AutoModelForCausalLM.from_pretrained(
    GEN_MODEL,
    cache_dir="/tmp/hf_cache",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device_map="auto",
    low_cpu_mem_usage=True
)

embedder = SentenceTransformer(EMB_MODEL, cache_folder="/tmp/hf_cache")

# ------------------------------------------------------
# 2️⃣ Load and index documents


# ------------------------------------------------------
# 3️⃣ Retrieval + generation logic
# ------------------------------------------------------
def retrieve_context(query, k=5):
    q_emb = embedder.encode([query], convert_to_numpy=True)
    D, I = index.search(q_emb, k)
    return "\n\n".join([chunks[i] for i in I[0]])

def generate_response(query, history):
    context = retrieve_context(query)
    system_prompt = (
        "You are a helpful assistant that uses the retrieved context to answer questions.\n\n"
        f"Context:\n{context}\n\n"
        f"User: {query}\nAssistant:"
    )
    inputs = tokenizer(system_prompt, return_tensors="pt").to(model.device)
    output_ids = model.generate(**inputs, max_new_tokens=300)
    output = tokenizer.decode(output_ids[0], skip_special_tokens=True)
    return output.split("Assistant:")[-1].strip()

def chat_fn(user_message, history):
    response = generate_response(user_message, history)
    history = history + [(user_message, response)]
    return history, history

# ------------------------------------------------------
# 4️⃣ Gradio UI
# ------------------------------------------------------
with gr.Blocks(title="Qwen3-Harrison-RAG Chatbot") as demo:
    gr.Markdown(
        """
        # 🤖 Qwen3-Harrison-RAG Chatbot  
        Ask me anything — I’ll retrieve relevant context and answer!
        """
    )
    chatbot = gr.Chatbot(height=400)
    with gr.Row():
        msg = gr.Textbox(placeholder="Type your message here...", scale=4)
        clear = gr.Button("Clear", scale=1)

    msg.submit(chat_fn, [msg, chatbot], [chatbot, chatbot])
    clear.click(lambda: None, None, chatbot, queue=False)

# ------------------------------------------------------
# 5️⃣ Launch for Hugging Face Spaces
# ------------------------------------------------------
if __name__ == "__main__":
    demo.launch()