app.py

import os
import base64
import json
from datetime import datetime
import traceback
from typing import Dict, Any

import gradio as gr
from huggingface_hub import HfApi, InferenceClient
from fastmcp import FastMCP
from pydantic import BaseModel, Field # Import Pydantic BaseModel and Field

HF_DATASET_REPO = os.environ.get("HF_DATASET_REPO", "OppaAI/Robot_MCP")
HF_VLM_MODEL = os.environ.get("HF_VLM_MODEL", "Qwen/Qwen2.5-VL-7B-Instruct")

mcp = FastMCP("Robot_MCP_Server")   # <-- Important

# ---------------------------------------------------
#  Define Pydantic Schema for the input payload
# ---------------------------------------------------
# This defines the expected structure and automatically generates the valid JSON schema
class RobotWatchPayload(BaseModel):
    hf_token: str = Field(description="Your Hugging Face API token.")
    robot_id: str = Field(description="The unique identifier for the robot.", default="unknown")
    image_b64: str = Field(description="Base64 encoded image data.")


def upload_image(image_b64: str, hf_token: str):
    try:
        image_bytes = base64.b64decode(image_b64)

        os.makedirs("/tmp", exist_ok=True)
        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
        local_path = f"/tmp/robot_img_{timestamp}.jpg"

        with open(local_path, "wb") as f:
            f.write(image_bytes)

        filename = f"robot_{timestamp}.jpg"
        api = HfApi()
        api.upload_file(
            path_or_fileobj=local_path,
            path_in_repo=f"tmp/{filename}",
            repo_id=HF_DATASET_REPO,
            repo_type="dataset",
            token=hf_token
        )

        url = f"https://huggingface.co/datasets/{HF_DATASET_REPO}/resolve/main/tmp/{filename}"
        return local_path, url, filename, len(image_bytes)

    except Exception:
        traceback.print_exc()
        return None, None, None, 0


def safe_parse_json_from_text(text: str):
    if not text:
        return None
    try:
        return json.loads(text)
    except:
        pass

    cleaned = text.strip().strip("`").strip()
    if cleaned.lower().startswith("json"):
        cleaned = cleaned[4:].strip()

    try:
        start = cleaned.find("{")
        end = cleaned.rfind("}")
        return json.loads(cleaned[start:end + 1])
    except:
        return None


# ---------------------------------------------------
#  TRUE MCP TOOL — THIS must be exposed to MCP client
# ---------------------------------------------------
@mcp.tool(
    name="robot_watch",
    description="Analyze a base64 image using Qwen VLM and return structured JSON.",
)
def robot_watch(payload: RobotWatchPayload): # <-- Type hint with Pydantic model
    # The payload is already validated and typed correctly by fastmcp/pydantic
    hf_token = payload.hf_token
    image_b64 = payload.image_b64
    robot_id = payload.robot_id

    if not hf_token:
        # This check is technically redundant if the schema demands it, but safe.
        return {"error": "Missing hf_token"}
    # image_b64 existence is also guaranteed by the schema

    _, hf_url, _, size_bytes = upload_image(image_b64, hf_token)
    if not hf_url:
        return {"error": "Image upload failed"}

    system_prompt = """
Respond in STRICT JSON ONLY:
{
 "description": "...",
 "human": "...",
 "environment": "...",
 "objects": []
}
"""

    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": [
            {"type": "text", "text": "Analyze the image."},
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_b64}"}}
        ]}
    ]

    client = InferenceClient(token=hf_token)
    try:
        resp = client.chat.completions.create(
            model=HF_VLM_MODEL,
            messages=messages,
            max_tokens=500,
            temperature=0.1
        )
    except Exception as e:
        return {"status": "error", "message": str(e)}

    vlm_output = resp.choices[0].message.content.strip()
    parsed = safe_parse_json_from_text(vlm_output) or {}

    return {
        "status": "success",
        "robot_id": robot_id,
        "file_size_bytes": size_bytes,
        "image_url": hf_url,
        "result": parsed,
        "vlm_raw": vlm_output
    }


# ---------------------------------------------------
# Gradio UI — separate from MCP tool layer
# ---------------------------------------------------
# The process_json function will still work with the Pydantic model input
def process_json(payload: Dict[str, Any]):
    # When called via Gradio UI (not MCP), input will be Dict, so handle type conversion
    pydantic_payload = RobotWatchPayload(**payload)
    return robot_watch(pydantic_payload)


app = gr.Interface(
    fn=process_json,
    inputs=gr.JSON(),
    outputs=gr.JSON(),
    title="Robot MCP Server",
    api_name="predict"
)

if __name__ == "__main__":
    # Gradio will use the 'mcp' instance defined globally to host the MCP server endpoints
    app.launch(mcp_server=True)