agents/ten_packages/extension/vision_tool_python/extension.py

#
# This file is part of TEN Framework, an open source project.
# Licensed under the Apache License, Version 2.0.
# See the LICENSE file for more information.
#
from ten_ai_base.types import LLMChatCompletionContentPartImageParam, LLMToolMetadata, LLMToolResult, LLMToolResultRequery
from ten_ai_base.llm_tool import AsyncLLMToolBaseExtension
from ten import (
    AudioFrame,
    VideoFrame,
    AsyncTenEnv,
    Cmd,
    Data,
)
from PIL import Image
from io import BytesIO
from base64 import b64encode


def rgb2base64jpeg(rgb_data, width, height):
    # Convert the RGB image to a PIL Image
    pil_image = Image.frombytes("RGBA", (width, height), bytes(rgb_data))
    pil_image = pil_image.convert("RGB")

    # Resize the image while maintaining its aspect ratio
    pil_image = resize_image_keep_aspect(pil_image, 512)

    # Save the image to a BytesIO object in JPEG format
    buffered = BytesIO()
    pil_image.save(buffered, format="JPEG")
    # pil_image.save("test.jpg", format="JPEG")

    # Get the byte data of the JPEG image
    jpeg_image_data = buffered.getvalue()

    # Convert the JPEG byte data to a Base64 encoded string
    base64_encoded_image = b64encode(jpeg_image_data).decode("utf-8")

    # Create the data URL
    mime_type = "image/jpeg"
    base64_url = f"data:{mime_type};base64,{base64_encoded_image}"
    return base64_url


def resize_image_keep_aspect(image, max_size=512):
    """
    Resize an image while maintaining its aspect ratio, ensuring the larger dimension is max_size.
    If both dimensions are smaller than max_size, the image is not resized.

    :param image: A PIL Image object
    :param max_size: The maximum size for the larger dimension (width or height)
    :return: A PIL Image object (resized or original)
    """
    # Get current width and height
    width, height = image.size

    # If both dimensions are already smaller than max_size, return the original image
    if width <= max_size and height <= max_size:
        return image

    # Calculate the aspect ratio
    aspect_ratio = width / height

    # Determine the new dimensions
    if width > height:
        new_width = max_size
        new_height = int(max_size / aspect_ratio)
    else:
        new_height = max_size
        new_width = int(max_size * aspect_ratio)

    # Resize the image with the new dimensions
    resized_image = image.resize((new_width, new_height))

    return resized_image


class VisionToolExtension(AsyncLLMToolBaseExtension):
    image_data = None
    image_width = 0
    image_height = 0

    async def on_init(self, ten_env: AsyncTenEnv) -> None:
        ten_env.log_debug("on_init")
        await super().on_init(ten_env)

    async def on_start(self, ten_env: AsyncTenEnv) -> None:
        ten_env.log_debug("on_start")
        await super().on_start(ten_env)

    async def on_stop(self, ten_env: AsyncTenEnv) -> None:
        ten_env.log_debug("on_stop")

        # TODO: clean up resources

        await super().on_stop(ten_env)

    async def on_cmd(self, ten_env: AsyncTenEnv, cmd: Cmd) -> None:
        cmd_name = cmd.get_name()
        ten_env.log_debug("on_cmd name {}".format(cmd_name))

        await super().on_cmd(ten_env, cmd)

    async def on_data(self, ten_env: AsyncTenEnv, data: Data) -> None:
        data_name = data.get_name()
        ten_env.log_debug("on_data name {}".format(data_name))

    async def on_audio_frame(
        self, ten_env: AsyncTenEnv, audio_frame: AudioFrame
    ) -> None:
        audio_frame_name = audio_frame.get_name()
        ten_env.log_debug("on_audio_frame name {}".format(audio_frame_name))

    async def on_video_frame(
        self, ten_env: AsyncTenEnv, video_frame: VideoFrame
    ) -> None:
        video_frame_name = video_frame.get_name()
        ten_env.log_debug("on_video_frame name {}".format(video_frame_name))

        self.image_data = video_frame.get_buf()
        self.image_width = video_frame.get_width()
        self.image_height = video_frame.get_height()

    def get_tool_metadata(self, ten_env: AsyncTenEnv) -> list[LLMToolMetadata]:
        return [
            LLMToolMetadata(
                name="get_vision_tool",
                description="Get the image from camera. Call this whenever you need to understand the input camera image like you have vision capability, for example when user asks 'What can you see?' or 'Can you see me?'",
                parameters=[],
            ),
        ]

    async def run_tool(
        self, ten_env: AsyncTenEnv, name: str, args: dict
    ) -> LLMToolResult | None:
        if name == "get_vision_tool":
            if self.image_data is None:
                raise ValueError("No image data available")

            base64_image = rgb2base64jpeg(
                self.image_data, self.image_width, self.image_height
            )
            return LLMToolResultRequery(
                type="requery",
                content=[
                    LLMChatCompletionContentPartImageParam(
                        type="image_url",
                        image_url={
                            "url": base64_image,
                        },
                    )
                ],
            )