File size: 6,350 Bytes
87337b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
#
# 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.
#
import json
from ten import (
    AudioFrame,
    VideoFrame,
    AsyncTenEnv,
    Cmd,
    Data,
)
from PIL import Image
from io import BytesIO
from base64 import b64encode

from ten_ai_base.const import CMD_CHAT_COMPLETION_CALL
from ten_ai_base.types import (
    LLMChatCompletionUserMessageParam,
    LLMToolMetadata,
    LLMToolMetadataParameter,
    LLMToolResult,
    LLMToolResultLLMResult,
)
from ten_ai_base.llm_tool import AsyncLLMToolBaseExtension


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 VisionAnalyzeToolExtension(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")

    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

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

    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_chat_completion",
                description="Get the image analyze result 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 in my camera?' or 'Can you see me?'",
                parameters=[
                    LLMToolMetadataParameter(
                        name="query",
                        type="string",
                        description="The vision completion query.",
                        required=True,
                    ),
                ],
            ),
        ]

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

            if "query" not in args:
                raise ValueError("Failed to get property")

            query = args["query"]

            base64_image = rgb2base64jpeg(
                self.image_data, self.image_width, self.image_height
            )
            # return LLMToolResult(message=LLMCompletionArgsMessage(role="user", content=[result]))
            cmd: Cmd = Cmd.create(CMD_CHAT_COMPLETION_CALL)
            message: LLMChatCompletionUserMessageParam = (
                LLMChatCompletionUserMessageParam(
                    role="user",
                    content=[
                        {"type": "text", "text": query},
                        {"type": "image_url", "image_url": {"url": base64_image}},
                    ],
                )
            )
            cmd.set_property_from_json("arguments", json.dumps({"messages": [message]}))
            ten_env.log_info("send_cmd {}".format(message))
            [cmd_result, _] = await ten_env.send_cmd(cmd)
            result = cmd_result.get_property_to_json("response")
            return LLMToolResultLLMResult(
                type="llmresult",
                content=json.dumps(result),
            )