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_TOKEN = os.environ.get("HF_CV_ROBOT_TOKEN")
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" 

#if not HF_TOKEN:
#    raise ValueError("HF_TOKEN environment variable not set.")

# Initialize the Hugging Face Inference Client
hf_client = InferenceClient(token=HF_TOKEN)

# --- Helper Functions ---
def save_and_upload_image(image_b64):
    """Save image to /tmp and upload to HF dataset."""
    image_bytes = base64.b64decode(image_b64)
    # Use a unique filename to prevent conflicts in /tmp
    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,
        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:
        hf_token = payload.get("hf_token", HF_TOKEN)
        robot_id = payload.get("robot_id", "unknown")
        image_b64 = payload["image_b64"]

        # 1️⃣ Save & upload image (needed for tracking, but B64 is used for VLM call)
        local_tmp_path, hf_url, path_in_repo, size_bytes = save_and_upload_image(image_b64)
        
        # 2️⃣ Prepare the multimodal message payload for the conversational API
        messages_payload = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "Describe this image in detail."},
                    # Pass the original Base64 string directly in the required format
                    {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_b64}"}},
                ],
            }
        ]

        # 3️⃣ Call VLM using hf_client.chat.completions.create (The correct method for 'conversational' task)
        chat_completion = hf_client.chat.completions.create(
            model=HF_VLM_MODEL,
            messages=messages_payload,
            max_tokens=150, # Use max_tokens instead of max_new_tokens for this method
        )
        
        # Extract the text content from the response object
        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:
        # Added better error handling
        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)