app.py

import os
import base64
import requests
import tempfile
import gradio as gr
from huggingface_hub import upload_file, InferenceClient
from PIL import Image

# --- Config ---
HF_DATASET_REPO = "OppaAI/Robot_MCP"
# Model specifically for VLM (image-to-text) tasks on Hugging Face
HF_VLM_MODEL = "Qwen/Qwen2.5-VL-7B-Instruct" 

# --- Helper Functions ---
def save_and_upload_image(image_b64, hf_token):
    """Save image to /tmp and upload to HF dataset."""
    image_bytes = base64.b64decode(image_b64)
    local_tmp_path = f"/tmp/tmp.jpg" 
    with open(local_tmp_path, "wb") as f:
        f.write(image_bytes)

    path_in_repo = f"images/tmp.jpg"
    upload_file(
        path_or_fileobj=local_tmp_path,
        path_in_repo=path_in_repo,
        repo_id=HF_DATASET_REPO,
        token=hf_token,  # ← use token from payload
        repo_type="dataset"
    )

    hf_image_url = f"https://huggingface.co/datasets/{HF_DATASET_REPO}/resolve/main/{path_in_repo}"
    return local_tmp_path, hf_image_url, path_in_repo, len(image_bytes)


# --- Main MCP function ---
def process_and_describe(payload: dict):
    try:
        # 1️⃣ Use robot-sent token if available, otherwise fallback
        hf_token = payload.get("hf_token")
        if not hf_token:
            return {"error": "HF token not provided in payload."}

        robot_id = payload.get("robot_id", "unknown")
        image_b64 = payload["image_b64"]

        # 2️⃣ Save image temporarily (for tracking)
        local_tmp_path, hf_url, path_in_repo, size_bytes = save_and_upload_image(image_b64, hf_token)

        # 3️⃣ Initialize HF client per request
        hf_client = InferenceClient(token=hf_token)

        # 4️⃣ Prepare multimodal message payload
        messages_payload = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "Describe this image in detail."},
                    {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_b64}"}},
                ],
            }
        ]

        # 5️⃣ Call VLM
        chat_completion = hf_client.chat.completions.create(
            model=HF_VLM_MODEL,
            messages=messages_payload,
            max_tokens=150,
        )

        vlm_text = chat_completion.choices[0].message.content.strip()

        return {
            "saved_to_hf_hub": True,
            "repo_id": HF_DATASET_REPO,
            "path_in_repo": path_in_repo,
            "image_url": hf_url,
            "file_size_bytes": size_bytes,
            "robot_id": robot_id,
            "vlm_description": vlm_text
        }

    except Exception as e:
        return {"error": f"An API error occurred: {str(e)}"}

# --- Gradio MCP Interface ---
demo = gr.Interface(
    fn=process_and_describe,
    inputs=gr.JSON(label="Input Payload (Dict format with 'image_b64')"),
    outputs=gr.JSON(label="Reply to Jetson"),
    api_name="predict"
)

if __name__ == "__main__":
    # Ensure you have the latest huggingface-hub: pip install --upgrade huggingface-hub Pillow requests
    demo.launch(mcp_server=True)