app.py

"""
Hugging Face Space: Multilingual Document Assistant
Run this as a Gradio app on Hugging Face Spaces.
Set HF_MODEL_ID to your Hub model (e.g. your-username/multilingual-doc-assistant).
"""
import os
import torch
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

# Model: Hub id (e.g. your-username/multilingual-doc-assistant) or local path.
# On Spaces set HF_MODEL_ID in Settings → Variables. Local: use trained folder if present.
_SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
_LOCAL_MODEL = os.path.join(_SCRIPT_DIR, "multilingual-doc-model")
HF_MODEL_ID = os.environ.get("HF_MODEL_ID") or (_LOCAL_MODEL if os.path.isdir(_LOCAL_MODEL) else "bigscience/bloom-560m")

def load_pipeline():
    tokenizer = AutoTokenizer.from_pretrained(HF_MODEL_ID)
    model = AutoModelForCausalLM.from_pretrained(HF_MODEL_ID)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    device = 0 if torch.cuda.is_available() else -1
    return pipeline(
        "text-generation",
        model=model,
        tokenizer=tokenizer,
        device=device,
    )

# Load once at startup (Spaces will cache)
pipe = load_pipeline()

def _get_text(content):
    """Extract plain text from Gradio message content (str or list of parts)."""
    if isinstance(content, str):
        return content
    if isinstance(content, list):
        for part in content:
            if isinstance(part, dict) and part.get("type") == "text":
                return part.get("text", "")
            if isinstance(part, str):
                return part
    return ""

def build_prompt(history, message):
    parts = []
    for turn in history:
        if isinstance(turn, (list, tuple)) and len(turn) >= 2:
            user_msg, assistant_msg = str(turn[0] or ""), str(turn[1] or "")
        elif isinstance(turn, dict):
            role = turn.get("role", "")
            content = _get_text(turn.get("content", ""))
            if role == "user":
                user_msg, assistant_msg = content, ""
            else:
                user_msg, assistant_msg = "", content
            if not user_msg and not assistant_msg:
                continue
        else:
            continue
        if user_msg:
            parts.append(f"User: {user_msg}\nAssistant: {assistant_msg}")
    parts.append(f"User: {message}\nAssistant:")
    return "\n".join(parts)

def chat(message, history):
    if not message.strip():
        return ""
    prompt = build_prompt(history, message)
    out = pipe(
        prompt,
        max_new_tokens=150,
        do_sample=True,
        temperature=0.7,
        pad_token_id=pipe.tokenizer.pad_token_id,
    )
    full = out[0]["generated_text"]
    # Return only the new Assistant part (after the last "Assistant:")
    if "Assistant:" in full:
        reply = full.split("Assistant:")[-1].strip()
    else:
        reply = full[len(prompt):].strip()
    # Stop at next "User:" or double newline
    for stop in ["\nUser:", "\n\nUser:"]:
        if stop in reply:
            reply = reply.split(stop)[0].strip()
    return reply

with gr.Blocks(
    title="Multilingual Document Assistant",
    theme=gr.themes.Soft(),
) as demo:
    gr.Markdown("""
    # Multilingual Document Assistant
    **Supports:** Spanish · Chinese · Vietnamese · Portuguese  
    Ask about documents, get explanations, or chat. *(Agent-style responses)*
    """)
    gr.ChatInterface(
        fn=chat,
        type="messages",
        examples=[
            ["Explícame este documento: La IA mejora la productividad."],
            ["总结这段文字: 人工智能正在改变世界。"],
            ["Giải thích đoạn này: Công nghệ giúp cuộc sống dễ dàng hơn."],
        ],
        retry_btn="Retry",
        undo_btn="Undo",
        clear_btn="Clear",
    )
    gr.Markdown(f"*Model: `{HF_MODEL_ID}`*")

if __name__ == "__main__":
    demo.launch()