app.py

import os
import copy
import base64
import requests
import tempfile
import secrets
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 = "llava-hf/llava-interleave-qwen-0.5b-hf" # A suitable VLM model

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/uploaded_image_{secrets.token_hex(8)}.jpg" 
    with open(local_tmp_path, "wb") as f:
        f.write(image_bytes)

    path_in_repo = f"images/uploaded_image_{secrets.token_hex(8)}.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:
        robot_id = payload.get("robot_id", "unknown")
        image_b64 = payload["image_b64"]

        # 1️⃣ Save & upload image
        local_tmp_path, hf_url, path_in_repo, size_bytes = save_and_upload_image(image_b64)
        
        # --- NEW VLM CALL LOGIC ---
        # 2️⃣ Prepare prompt in the Qwen specific format (using Markdown for image embedding)
        # The API expects an image embedded in the prompt using Markdown syntax.
        prompt_text = "Describe this image in detail."
        
        # Base64 encode the image for embedding in the JSON payload
        with open(local_tmp_path, "rb") as f:
            image_b64_encoded_string = base64.b64encode(f.read()).decode("utf-8")

        # The full prompt format required by Qwen in the API
        full_prompt = f'<img src="data:image/jpeg;base64,{image_b64_encoded_string}">{prompt_text}'

        # 3️⃣ Call VLM using hf_client.post (low-level API call for specific models)
        # We use the 'text-generation' task endpoint as indicated by the error message.
        api_response = hf_client.post(
            json={"inputs": full_prompt, "parameters": {"max_new_tokens": 150}},
            model=HF_VLM_MODEL,
            task="text-generation"
        )
        
        # The response is usually a list of dicts, extract the generated text
        # Example response format: [{'generated_text': '... description ...'}]
        if isinstance(api_response, list) and len(api_response) > 0:
            vlm_text = api_response[0].get('generated_text', '').strip()
        else:
            vlm_text = "Failed to parse VLM response."

        # --- END NEW VLM CALL LOGIC ---

        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 as suggested previously
        return {"error": str(e)}

# ... (Gradio Interface code remains the same) ...

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