app.py

import os
import base64
import json
from datetime import datetime
import traceback
# Removed unused typing import: from typing import Dict, Any

import gradio as gr
from huggingface_hub import HfApi, InferenceClient
# The FastMCP object is automatically initialized when you call app.launch(mcp_server=True)
# You don't need to manually instantiate FastMCP if only using Gradio's integration.
# from fastmcp import FastMCP # Removed manual import/instantiation
from pydantic import BaseModel, 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") # Removed this line

# ---------------------------------------------------
#  Payload Schema (Remains the same as it already expects image_b64)
# ---------------------------------------------------
class RobotWatchPayload(BaseModel):
    hf_token: str = Field(description="Your Hugging Face API token.")
    robot_id: str = Field(description="Robot identifier.", default="unknown")
    image_b64: str = Field(description="Base64 encoded image data.")


# ---------------------------------------------------
# Upload Helper (Remains the same)
# ---------------------------------------------------
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
        )

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

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


# ---------------------------------------------------
# JSON Cleaning Helper (Remains the same)
# ---------------------------------------------------
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


# ---------------------------------------------------
# Core VLM Analysis Logic (Remains the same)
# ---------------------------------------------------
def run_vlm_analysis(payload: RobotWatchPayload):
    # ... (function body remains identical to previous version) ...
    hf_token = payload.hf_token
    image_b64 = payload.image_b64
    robot_id = payload.robot_id

    _, 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 Function (NOW USES BASE64 STRING INPUT)
# ---------------------------------------------------
def robot_watch(
    hf_token_input: str,
    robot_id_input: str,
    image_b64_input: str # Changed input type to a string (base64)
):
    """
    Handles input from individual Gradio components (including base64 string), 
    converts to Pydantic model, and calls the core logic.
    """
    if not image_b64_input:
        return {"error": "Base64 image string is empty."}
    
    # Create the Pydantic model instance manually
    payload_instance = RobotWatchPayload(
        hf_token=hf_token_input,
        robot_id=robot_id_input,
        image_b64=image_b64_input
    )
    
    # Call the core logic
    result = run_vlm_analysis(payload_instance)
    return result


app = gr.Interface(
    fn=robot_watch, # Use the new multi-input function for the UI
    inputs=[
        gr.Textbox(label="Hugging Face Token", lines=1),
        gr.Textbox(label="Robot ID", lines=1, value="unknown"),
        gr.Textbox(label="Image Base64 String", lines=5) # Changed input component to Textbox
    ],
    outputs=gr.Json(label="Tool Output"),
    title="Robot MCP Server (Base64 Inputs)",
    description="Interface for the robot VLM analysis using individual fields, including base64 image string.",
    api_name="predict"
)

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