- )
-}
\ No newline at end of file
diff --git a/app/src/renderer/src/router/index.tsx b/app/src/renderer/src/router/index.tsx
deleted file mode 100644
index 0ce2742..0000000
--- a/app/src/renderer/src/router/index.tsx
+++ /dev/null
@@ -1,69 +0,0 @@
-import Config from "@renderer/layouts/Config";
-import Home from "@renderer/layouts/Home";
-import { createHashRouter } from "react-router-dom";
-import {Category} from "@renderer/pages/Category";
-import { ContentList } from "@renderer/pages/ContentList";
-import CategoryLoader from "@renderer/pages/Category/CategoryLoader";
-import ContentListLoader from "@renderer/pages/ContentList/ContentListLoader";
-import ContentLoader from "@renderer/pages/Content/ContentLoader";
-import { Content } from "@renderer/pages/Content/Content";
-import ContentAction from "@renderer/pages/Content/ContentAction";
-import { Welcome } from "@renderer/pages/Welcome";
-import ContentListAction from "@renderer/pages/ContentList/ContentListAction";
-import CategoryAction from "@renderer/pages/Category/CategoryAction";
-import { Setting } from "@renderer/pages/Setting";
-import SettingAction from "@renderer/pages/Setting/SettingAction";
-import SettingLoader from "@renderer/pages/Setting/SettingLoader";
-import About from "@renderer/layouts/About";
-
-const router = createHashRouter([
- {
- path: "/",
- element:
- },
- {
- path: "about",
- element:
- },
- {
- path: "config",
- element: ,
- children: [
- {
- index: true,
- element: ,
- loader: SettingLoader,
- action: SettingAction
-
- },
- {
- path: "category",
- // 界面第一次打开时，默认显示的页面
- element: ,
- loader: CategoryLoader,
- action: CategoryAction,
- children: [
- {
- path: "contentList/:cid?",
- loader: ContentListLoader,
- action: ContentListAction,
- element: ,
- children: [
- {
- index: true,
- element:
- },
- {
- path: "content/:id/:search?",
- loader: ContentLoader,
- action: ContentAction,
- element:
- }
- ]
- }
- ]
- }
- ]
- }
-])
-export default router
diff --git a/app/src/renderer/src/store/useStore.ts b/app/src/renderer/src/store/useStore.ts
deleted file mode 100644
index 3fe9c9c..0000000
--- a/app/src/renderer/src/store/useStore.ts
+++ /dev/null
@@ -1,33 +0,0 @@
-import { ChatMessage } from "@ant-design/pro-chat";
-import { create } from "zustand";
-interface StateProps{
- data: ContentType[],
- setData: (data: ContentType[]) => void,
- search: string,
- setSearch: (search: string) => void,
- error: string,
- setError: (error: string) => void,
- selectId: number,
- setSelectId: (selectId: number) => void,
- editCategoryId: number,
- setEditCategoryId: (id: number) => void,
- chatMessages: ChatMessage>[],
- setChatMessage: (chatMessage: ChatMessage>[]) => void,
-
-}
-export const useStore = create((set) => ({
- data: [],
- setData: (data) => set({data}),
- search: "",
- setSearch: (search) => set({search}),
- error: "",
- setError: (error) => set({error}),
- selectId: 0,
- setSelectId: (selectId) => set({selectId}),
- editCategoryId: 0,
- setEditCategoryId: (editCategoryId) => set({editCategoryId}),
- chatMessages: [],
- setChatMessage: (chatMessages) => set({chatMessages}),
- }))
-
-
diff --git a/app/tailwind.config.js b/app/tailwind.config.js
deleted file mode 100644
index 7cb7e37..0000000
--- a/app/tailwind.config.js
+++ /dev/null
@@ -1,77 +0,0 @@
-/** @type {import('tailwindcss').Config} */
-module.exports = {
- darkMode: ["class"],
- content: [
- './pages/**/*.{ts,tsx}',
- './components/**/*.{ts,tsx}',
- './app/**/*.{ts,tsx}',
- './src/**/*.{ts,tsx}',
- ],
- prefix: "",
- theme: {
- container: {
- center: true,
- padding: "2rem",
- screens: {
- "2xl": "1400px",
- },
- },
- extend: {
- colors: {
- border: "hsl(var(--border))",
- input: "hsl(var(--input))",
- ring: "hsl(var(--ring))",
- background: "hsl(var(--background))",
- foreground: "hsl(var(--foreground))",
- primary: {
- DEFAULT: "hsl(var(--primary))",
- foreground: "hsl(var(--primary-foreground))",
- },
- secondary: {
- DEFAULT: "hsl(var(--secondary))",
- foreground: "hsl(var(--secondary-foreground))",
- },
- destructive: {
- DEFAULT: "hsl(var(--destructive))",
- foreground: "hsl(var(--destructive-foreground))",
- },
- muted: {
- DEFAULT: "hsl(var(--muted))",
- foreground: "hsl(var(--muted-foreground))",
- },
- accent: {
- DEFAULT: "hsl(var(--accent))",
- foreground: "hsl(var(--accent-foreground))",
- },
- popover: {
- DEFAULT: "hsl(var(--popover))",
- foreground: "hsl(var(--popover-foreground))",
- },
- card: {
- DEFAULT: "hsl(var(--card))",
- foreground: "hsl(var(--card-foreground))",
- },
- },
- borderRadius: {
- lg: "var(--radius)",
- md: "calc(var(--radius) - 2px)",
- sm: "calc(var(--radius) - 4px)",
- },
- keyframes: {
- "accordion-down": {
- from: { height: "0" },
- to: { height: "var(--radix-accordion-content-height)" },
- },
- "accordion-up": {
- from: { height: "var(--radix-accordion-content-height)" },
- to: { height: "0" },
- },
- },
- animation: {
- "accordion-down": "accordion-down 0.2s ease-out",
- "accordion-up": "accordion-up 0.2s ease-out",
- },
- },
- },
- plugins: [require("tailwindcss-animate")],
-}
\ No newline at end of file
diff --git a/app/tsconfig.json b/app/tsconfig.json
deleted file mode 100644
index 31bac6e..0000000
--- a/app/tsconfig.json
+++ /dev/null
@@ -1,4 +0,0 @@
-{
- "files": [],
- "references": [{ "path": "./tsconfig.node.json" }, { "path": "./tsconfig.web.json" }]
-}
diff --git a/app/tsconfig.node.json b/app/tsconfig.node.json
deleted file mode 100644
index 163a5eb..0000000
--- a/app/tsconfig.node.json
+++ /dev/null
@@ -1,8 +0,0 @@
-{
- "extends": "@electron-toolkit/tsconfig/tsconfig.node.json",
- "include": ["electron.vite.config.*", "src/main/**/*", "src/preload/**/*", "types.d.ts"],
- "compilerOptions": {
- "composite": true,
- "types": ["electron-vite/node"]
- }
-}
diff --git a/app/tsconfig.web.json b/app/tsconfig.web.json
deleted file mode 100644
index 26eaa06..0000000
--- a/app/tsconfig.web.json
+++ /dev/null
@@ -1,20 +0,0 @@
-{
- "extends": "@electron-toolkit/tsconfig/tsconfig.web.json",
- "include": [
- "src/renderer/src/env.d.ts",
- "src/renderer/src/**/*",
- "src/renderer/src/**/*.tsx",
- "src/preload/*.d.ts",
- "types.d.ts"
- ],
- "compilerOptions": {
- "composite": true,
- "jsx": "react-jsx",
- "baseUrl": ".",
- "paths": {
- "@renderer/*": [
- "src/renderer/src/*"
- ]
- }
- }
-}
diff --git a/app/types.d.ts b/app/types.d.ts
deleted file mode 100644
index c7a0d32..0000000
--- a/app/types.d.ts
+++ /dev/null
@@ -1,35 +0,0 @@
-type SqlActionType = 'findAll' | 'findOne' | 'create' | 'update' | 'del'|'config'
-
-type CategoryType = {
- id: number
- name: string
- createdAt: string
-}
-
-type ContentType = {
- id: number
- title: string
- content: string
- category_id: string
- created_at: string
-}
-
-
-type WindowNameType = 'search' | 'config' | 'code' | 'interactive' | 'about'
-
-type ConfigType = {
- id: number
- content: string
-}
-
-
-type ConfigDataType = {
- shortcut: string
- format: string
- llm: {
- model: string
- api_key?: string
- base_url?: string
- api_base?: string
- }
-}
diff --git a/auto.py b/auto.py
new file mode 100644
index 0000000..5343918
--- /dev/null
+++ b/auto.py
@@ -0,0 +1,81 @@
+import os
+import logging
+import argparse
+import shlex
+import subprocess
+from flask import Flask, request, jsonify, send_file
+import threading
+import traceback
+import pyautogui
+from PIL import Image
+from io import BytesIO
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--log_file", help="log file path", type=str,
+ default=os.path.join(os.path.dirname(__file__), "server.log"))
+parser.add_argument("--port", help="port", type=int, default=5000)
+args = parser.parse_args()
+
+logging.basicConfig(filename=args.log_file,level=logging.DEBUG, filemode='w' )
+logger = logging.getLogger('werkzeug')
+
+app = Flask(__name__)
+
+computer_control_lock = threading.Lock()
+
+@app.route('/probe', methods=['GET'])
+def probe_endpoint():
+ return jsonify({"status": "Probe successful", "message": "Service is operational"}), 200
+
+@app.route('/execute', methods=['POST'])
+def execute_command():
+ # Only execute one command at a time
+ with computer_control_lock:
+ data = request.json
+ # The 'command' key in the JSON request should contain the command to be executed.
+ shell = data.get('shell', False)
+ command = data.get('command', "" if shell else [])
+
+ if isinstance(command, str) and not shell:
+ command = shlex.split(command)
+
+ # Expand user directory
+ for i, arg in enumerate(command):
+ if arg.startswith("~/"):
+ command[i] = os.path.expanduser(arg)
+
+ # Execute the command without any safety checks.
+ try:
+ result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, shell=shell, text=True, timeout=120)
+ return jsonify({
+ 'status': 'success',
+ 'output': result.stdout,
+ 'error': result.stderr,
+ 'returncode': result.returncode
+ })
+ except Exception as e:
+ logger.error("\n" + traceback.format_exc() + "\n")
+ return jsonify({
+ 'status': 'error',
+ 'message': str(e)
+ }), 500
+
+@app.route('/screenshot', methods=['GET'])
+def capture_screen_with_cursor():
+ cursor_path = os.path.join(os.path.dirname(__file__), "cursor.png")
+ screenshot = pyautogui.screenshot()
+ cursor_x, cursor_y = pyautogui.position()
+ cursor = Image.open(cursor_path)
+ # make the cursor smaller
+ cursor = cursor.resize((int(cursor.width / 1.5), int(cursor.height / 1.5)))
+ screenshot.paste(cursor, (cursor_x, cursor_y), cursor)
+
+
+ # Convert PIL Image to bytes and send
+ img_io = BytesIO()
+ screenshot.save(img_io, 'PNG')
+ img_io.seek(0)
+ return send_file(img_io, mimetype='image/png')
+
+if __name__ == '__main__':
+ app.run(debug=True, host="0.0.0.0", port=args.port)
\ No newline at end of file
diff --git a/gradio_ui/.gitignore b/gradio_ui/.gitignore
new file mode 100644
index 0000000..c036379
--- /dev/null
+++ b/gradio_ui/.gitignore
@@ -0,0 +1 @@
+tmp/
\ No newline at end of file
diff --git a/gradio_ui/__init__.py b/gradio_ui/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/gradio_ui/agent/anthropic_agent.py b/gradio_ui/agent/anthropic_agent.py
new file mode 100644
index 0000000..ddca387
--- /dev/null
+++ b/gradio_ui/agent/anthropic_agent.py
@@ -0,0 +1,162 @@
+"""
+Agentic sampling loop that calls the Anthropic API and local implenmentation of anthropic-defined computer use tools.
+"""
+import asyncio
+import platform
+from collections.abc import Callable
+from datetime import datetime
+from enum import StrEnum
+from typing import Any, cast
+
+from anthropic import Anthropic, AnthropicBedrock, AnthropicVertex, APIResponse
+from anthropic.types import (
+ ToolResultBlockParam,
+)
+from anthropic.types.beta import (
+ BetaContentBlock,
+ BetaContentBlockParam,
+ BetaImageBlockParam,
+ BetaMessage,
+ BetaMessageParam,
+ BetaTextBlockParam,
+ BetaToolResultBlockParam,
+)
+from anthropic.types import TextBlock
+from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
+
+from gradio_ui.tools import ComputerTool, ToolCollection, ToolResult
+
+from PIL import Image
+from io import BytesIO
+import gradio as gr
+from typing import Dict
+
+BETA_FLAG = "computer-use-2024-10-22"
+
+class APIProvider(StrEnum):
+ ANTHROPIC = "anthropic"
+ BEDROCK = "bedrock"
+ VERTEX = "vertex"
+
+SYSTEM_PROMPT = f"""
+* You are utilizing a Windows system with internet access.
+* The current date is {datetime.today().strftime('%A, %B %d, %Y')}.
+
+"""
+
+class AnthropicActor:
+ def __init__(
+ self,
+ model: str,
+ provider: APIProvider,
+ api_key: str,
+ api_response_callback: Callable[[APIResponse[BetaMessage]], None],
+ max_tokens: int = 4096,
+ only_n_most_recent_images: int | None = None,
+ print_usage: bool = True,
+ ):
+ self.model = model
+ self.provider = provider
+ self.api_key = api_key
+ self.api_response_callback = api_response_callback
+ self.max_tokens = max_tokens
+ self.only_n_most_recent_images = only_n_most_recent_images
+
+ self.tool_collection = ToolCollection(ComputerTool())
+
+ self.system = SYSTEM_PROMPT
+
+ self.total_token_usage = 0
+ self.total_cost = 0
+ self.print_usage = print_usage
+
+ # Instantiate the appropriate API client based on the provider
+ if provider == APIProvider.ANTHROPIC:
+ self.client = Anthropic(api_key=api_key)
+ elif provider == APIProvider.VERTEX:
+ self.client = AnthropicVertex()
+ elif provider == APIProvider.BEDROCK:
+ self.client = AnthropicBedrock()
+
+ def __call__(
+ self,
+ *,
+ messages: list[BetaMessageParam]
+ ):
+ """
+ Generate a response given history messages.
+ """
+ if self.only_n_most_recent_images:
+ _maybe_filter_to_n_most_recent_images(messages, self.only_n_most_recent_images)
+
+ # Call the API synchronously
+ raw_response = self.client.beta.messages.with_raw_response.create(
+ max_tokens=self.max_tokens,
+ messages=messages,
+ model=self.model,
+ system=self.system,
+ tools=self.tool_collection.to_params(),
+ betas=["computer-use-2024-10-22"],
+ )
+
+ self.api_response_callback(cast(APIResponse[BetaMessage], raw_response))
+
+ response = raw_response.parse()
+ print(f"AnthropicActor response: {response}")
+
+ self.total_token_usage += response.usage.input_tokens + response.usage.output_tokens
+ self.total_cost += (response.usage.input_tokens * 3 / 1000000 + response.usage.output_tokens * 15 / 1000000)
+
+ if self.print_usage:
+ print(f"Claude total token usage so far: {self.total_token_usage}, total cost so far: $USD{self.total_cost}")
+
+ return response
+
+
+def _maybe_filter_to_n_most_recent_images(
+ messages: list[BetaMessageParam],
+ images_to_keep: int,
+ min_removal_threshold: int = 10,
+):
+ """
+ With the assumption that images are screenshots that are of diminishing value as
+ the conversation progresses, remove all but the final `images_to_keep` tool_result
+ images in place, with a chunk of min_removal_threshold to reduce the amount we
+ break the implicit prompt cache.
+ """
+ if images_to_keep is None:
+ return messages
+
+ tool_result_blocks = cast(
+ list[ToolResultBlockParam],
+ [
+ item
+ for message in messages
+ for item in (
+ message["content"] if isinstance(message["content"], list) else []
+ )
+ if isinstance(item, dict) and item.get("type") == "tool_result"
+ ],
+ )
+
+ total_images = sum(
+ 1
+ for tool_result in tool_result_blocks
+ for content in tool_result.get("content", [])
+ if isinstance(content, dict) and content.get("type") == "image"
+ )
+
+ images_to_remove = total_images - images_to_keep
+ # for better cache behavior, we want to remove in chunks
+ images_to_remove -= images_to_remove % min_removal_threshold
+
+ for tool_result in tool_result_blocks:
+ if isinstance(tool_result.get("content"), list):
+ new_content = []
+ for content in tool_result.get("content", []):
+ if isinstance(content, dict) and content.get("type") == "image":
+ if images_to_remove > 0:
+ images_to_remove -= 1
+ continue
+ new_content.append(content)
+ tool_result["content"] = new_content
\ No newline at end of file
diff --git a/gradio_ui/agent/llm_utils/groqclient.py b/gradio_ui/agent/llm_utils/groqclient.py
new file mode 100644
index 0000000..c31a502
--- /dev/null
+++ b/gradio_ui/agent/llm_utils/groqclient.py
@@ -0,0 +1,59 @@
+from groq import Groq
+import os
+from .utils import is_image_path
+
+def run_groq_interleaved(messages: list, system: str, model_name: str, api_key: str, max_tokens=256, temperature=0.6):
+ """
+ Run a chat completion through Groq's API, ignoring any images in the messages.
+ """
+ api_key = api_key or os.environ.get("GROQ_API_KEY")
+ if not api_key:
+ raise ValueError("GROQ_API_KEY is not set")
+
+ client = Groq(api_key=api_key)
+ # avoid using system messages for R1
+ final_messages = [{"role": "user", "content": system}]
+
+ if isinstance(messages, list):
+ for item in messages:
+ if isinstance(item, dict):
+ # For dict items, concatenate all text content, ignoring images
+ text_contents = []
+ for cnt in item["content"]:
+ if isinstance(cnt, str):
+ if not is_image_path(cnt): # Skip image paths
+ text_contents.append(cnt)
+ else:
+ text_contents.append(str(cnt))
+
+ if text_contents: # Only add if there's text content
+ message = {"role": "user", "content": " ".join(text_contents)}
+ final_messages.append(message)
+ else: # str
+ message = {"role": "user", "content": item}
+ final_messages.append(message)
+
+ elif isinstance(messages, str):
+ final_messages.append({"role": "user", "content": messages})
+
+ try:
+ completion = client.chat.completions.create(
+ model="deepseek-r1-distill-llama-70b",
+ messages=final_messages,
+ temperature=0.6,
+ max_completion_tokens=max_tokens,
+ top_p=0.95,
+ stream=False,
+ reasoning_format="raw"
+ )
+
+ response = completion.choices[0].message.content
+ final_answer = response.split('\n')[-1] if '' in response else response
+ final_answer = final_answer.replace("", "")
+ token_usage = completion.usage.total_tokens
+
+ return final_answer, token_usage
+ except Exception as e:
+ print(f"Error in interleaved Groq: {e}")
+
+ return str(e), 0
\ No newline at end of file
diff --git a/gradio_ui/agent/llm_utils/oaiclient.py b/gradio_ui/agent/llm_utils/oaiclient.py
new file mode 100644
index 0000000..ad42110
--- /dev/null
+++ b/gradio_ui/agent/llm_utils/oaiclient.py
@@ -0,0 +1,62 @@
+import os
+import logging
+import base64
+import requests
+from .utils import is_image_path, encode_image
+
+def run_oai_interleaved(messages: list, system: str, model_name: str, api_key: str, max_tokens=256, temperature=0, provider_base_url: str = "https://api.openai.com/v1"):
+ headers = {"Content-Type": "application/json",
+ "Authorization": f"Bearer {api_key}"}
+ final_messages = [{"role": "system", "content": system}]
+
+ if type(messages) == list:
+ for item in messages:
+ contents = []
+ if isinstance(item, dict):
+ for cnt in item["content"]:
+ if isinstance(cnt, str):
+ if is_image_path(cnt) and 'o3-mini' not in model_name:
+ # 03 mini does not support images
+ base64_image = encode_image(cnt)
+ content = {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"}}
+ else:
+ content = {"type": "text", "text": cnt}
+ else:
+ # in this case it is a text block from anthropic
+ content = {"type": "text", "text": str(cnt)}
+
+ contents.append(content)
+
+ message = {"role": 'user', "content": contents}
+ else: # str
+ contents.append({"type": "text", "text": item})
+ message = {"role": "user", "content": contents}
+
+ final_messages.append(message)
+
+
+ elif isinstance(messages, str):
+ final_messages = [{"role": "user", "content": messages}]
+
+ payload = {
+ "model": model_name,
+ "messages": final_messages,
+ }
+ if 'o1' in model_name or 'o3-mini' in model_name:
+ payload['reasoning_effort'] = 'low'
+ payload['max_completion_tokens'] = max_tokens
+ else:
+ payload['max_tokens'] = max_tokens
+
+ response = requests.post(
+ f"{provider_base_url}/chat/completions", headers=headers, json=payload
+ )
+
+
+ try:
+ text = response.json()['choices'][0]['message']['content']
+ token_usage = int(response.json()['usage']['total_tokens'])
+ return text, token_usage
+ except Exception as e:
+ print(f"Error in interleaved openAI: {e}. This may due to your invalid API key. Please check the response: {response.json()} ")
+ return response.json()
\ No newline at end of file
diff --git a/gradio_ui/agent/llm_utils/omniparserclient.py b/gradio_ui/agent/llm_utils/omniparserclient.py
new file mode 100644
index 0000000..168effe
--- /dev/null
+++ b/gradio_ui/agent/llm_utils/omniparserclient.py
@@ -0,0 +1,44 @@
+import requests
+import base64
+from pathlib import Path
+from gradio_ui.tools.screen_capture import get_screenshot
+from gradio_ui.agent.llm_utils.utils import encode_image
+
+OUTPUT_DIR = "./tmp/outputs"
+
+class OmniParserClient:
+ def __init__(self,
+ url: str) -> None:
+ self.url = url
+
+ def __call__(self,):
+ screenshot, screenshot_path = get_screenshot()
+ screenshot_path = str(screenshot_path)
+ image_base64 = encode_image(screenshot_path)
+ response = requests.post(self.url, json={"base64_image": image_base64})
+ response_json = response.json()
+ print('omniparser latency:', response_json['latency'])
+
+ som_image_data = base64.b64decode(response_json['som_image_base64'])
+ screenshot_path_uuid = Path(screenshot_path).stem.replace("screenshot_", "")
+ som_screenshot_path = f"{OUTPUT_DIR}/screenshot_som_{screenshot_path_uuid}.png"
+ with open(som_screenshot_path, "wb") as f:
+ f.write(som_image_data)
+
+ response_json['width'] = screenshot.size[0]
+ response_json['height'] = screenshot.size[1]
+ response_json['original_screenshot_base64'] = image_base64
+ response_json['screenshot_uuid'] = screenshot_path_uuid
+ response_json = self.reformat_messages(response_json)
+ return response_json
+
+ def reformat_messages(self, response_json: dict):
+ screen_info = ""
+ for idx, element in enumerate(response_json["parsed_content_list"]):
+ element['idx'] = idx
+ if element['type'] == 'text':
+ screen_info += f'ID: {idx}, Text: {element["content"]}\n'
+ elif element['type'] == 'icon':
+ screen_info += f'ID: {idx}, Icon: {element["content"]}\n'
+ response_json['screen_info'] = screen_info
+ return response_json
\ No newline at end of file
diff --git a/gradio_ui/agent/llm_utils/utils.py b/gradio_ui/agent/llm_utils/utils.py
new file mode 100644
index 0000000..12ab36e
--- /dev/null
+++ b/gradio_ui/agent/llm_utils/utils.py
@@ -0,0 +1,13 @@
+import base64
+
+def is_image_path(text):
+ image_extensions = (".jpg", ".jpeg", ".png", ".gif", ".bmp", ".tiff", ".tif")
+ if text.endswith(image_extensions):
+ return True
+ else:
+ return False
+
+def encode_image(image_path):
+ """Encode image file to base64."""
+ with open(image_path, "rb") as image_file:
+ return base64.b64encode(image_file.read()).decode("utf-8")
\ No newline at end of file
diff --git a/gradio_ui/agent/vlm_agent.py b/gradio_ui/agent/vlm_agent.py
new file mode 100644
index 0000000..ee7ef70
--- /dev/null
+++ b/gradio_ui/agent/vlm_agent.py
@@ -0,0 +1,353 @@
+import json
+from collections.abc import Callable
+from typing import cast, Callable
+import uuid
+from PIL import Image, ImageDraw
+import base64
+from io import BytesIO
+
+from anthropic import APIResponse
+from anthropic.types import ToolResultBlockParam
+from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock, BetaMessageParam, BetaUsage
+
+from gradio_ui.agent.llm_utils.oaiclient import run_oai_interleaved
+from gradio_ui.agent.llm_utils.groqclient import run_groq_interleaved
+from gradio_ui.agent.llm_utils.utils import is_image_path
+import time
+import re
+
+OUTPUT_DIR = "./tmp/outputs"
+
+def extract_data(input_string, data_type):
+ # Regular expression to extract content starting from '```python' until the end if there are no closing backticks
+ pattern = f"```{data_type}" + r"(.*?)(```|$)"
+ # Extract content
+ # re.DOTALL allows '.' to match newlines as well
+ matches = re.findall(pattern, input_string, re.DOTALL)
+ # Return the first match if exists, trimming whitespace and ignoring potential closing backticks
+ return matches[0][0].strip() if matches else input_string
+
+class VLMAgent:
+ def __init__(
+ self,
+ model: str,
+ provider: str,
+ api_key: str,
+ output_callback: Callable,
+ api_response_callback: Callable,
+ max_tokens: int = 4096,
+ only_n_most_recent_images: int | None = None,
+ print_usage: bool = True,
+ ):
+ if model == "omniparser + gpt-4o":
+ self.model = "gpt-4o-2024-11-20"
+ elif model == "omniparser + R1":
+ self.model = "deepseek-r1-distill-llama-70b"
+ elif model == "omniparser + qwen2.5vl":
+ self.model = "qwen2.5-vl-72b-instruct"
+ elif model == "omniparser + o1":
+ self.model = "o1"
+ elif model == "omniparser + o3-mini":
+ self.model = "o3-mini"
+ else:
+ raise ValueError(f"Model {model} not supported")
+
+
+ self.provider = provider
+ self.api_key = api_key
+ self.api_response_callback = api_response_callback
+ self.max_tokens = max_tokens
+ self.only_n_most_recent_images = only_n_most_recent_images
+ self.output_callback = output_callback
+
+ self.print_usage = print_usage
+ self.total_token_usage = 0
+ self.total_cost = 0
+ self.step_count = 0
+
+ self.system = ''
+
+ def __call__(self, messages: list, parsed_screen: list[str, list, dict]):
+ self.step_count += 1
+ image_base64 = parsed_screen['original_screenshot_base64']
+ latency_omniparser = parsed_screen['latency']
+ self.output_callback(f'-- Step {self.step_count}: --', sender="bot")
+ screen_info = str(parsed_screen['screen_info'])
+ screenshot_uuid = parsed_screen['screenshot_uuid']
+ screen_width, screen_height = parsed_screen['width'], parsed_screen['height']
+
+ boxids_and_labels = parsed_screen["screen_info"]
+ system = self._get_system_prompt(boxids_and_labels)
+
+ # drop looping actions msg, byte image etc
+ planner_messages = messages
+ _remove_som_images(planner_messages)
+ _maybe_filter_to_n_most_recent_images(planner_messages, self.only_n_most_recent_images)
+
+ if isinstance(planner_messages[-1], dict):
+ if not isinstance(planner_messages[-1]["content"], list):
+ planner_messages[-1]["content"] = [planner_messages[-1]["content"]]
+ planner_messages[-1]["content"].append(f"{OUTPUT_DIR}/screenshot_{screenshot_uuid}.png")
+ planner_messages[-1]["content"].append(f"{OUTPUT_DIR}/screenshot_som_{screenshot_uuid}.png")
+
+ start = time.time()
+ if "gpt" in self.model or "o1" in self.model or "o3-mini" in self.model:
+ vlm_response, token_usage = run_oai_interleaved(
+ messages=planner_messages,
+ system=system,
+ model_name=self.model,
+ api_key=self.api_key,
+ max_tokens=self.max_tokens,
+ provider_base_url="https://api.openai.com/v1",
+ temperature=0,
+ )
+ print(f"oai token usage: {token_usage}")
+ self.total_token_usage += token_usage
+ if 'gpt' in self.model:
+ self.total_cost += (token_usage * 2.5 / 1000000) # https://openai.com/api/pricing/
+ elif 'o1' in self.model:
+ self.total_cost += (token_usage * 15 / 1000000) # https://openai.com/api/pricing/
+ elif 'o3-mini' in self.model:
+ self.total_cost += (token_usage * 1.1 / 1000000) # https://openai.com/api/pricing/
+ elif "r1" in self.model:
+ vlm_response, token_usage = run_groq_interleaved(
+ messages=planner_messages,
+ system=system,
+ model_name=self.model,
+ api_key=self.api_key,
+ max_tokens=self.max_tokens,
+ )
+ print(f"groq token usage: {token_usage}")
+ self.total_token_usage += token_usage
+ self.total_cost += (token_usage * 0.99 / 1000000)
+ elif "qwen" in self.model:
+ vlm_response, token_usage = run_oai_interleaved(
+ messages=planner_messages,
+ system=system,
+ model_name=self.model,
+ api_key=self.api_key,
+ max_tokens=min(2048, self.max_tokens),
+ provider_base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
+ temperature=0,
+ )
+ print(f"qwen token usage: {token_usage}")
+ self.total_token_usage += token_usage
+ self.total_cost += (token_usage * 2.2 / 1000000) # https://help.aliyun.com/zh/model-studio/getting-started/models?spm=a2c4g.11186623.0.0.74b04823CGnPv7#fe96cfb1a422a
+ else:
+ raise ValueError(f"Model {self.model} not supported")
+ latency_vlm = time.time() - start
+ self.output_callback(f"LLM: {latency_vlm:.2f}s, OmniParser: {latency_omniparser:.2f}s", sender="bot")
+
+ print(f"{vlm_response}")
+
+ if self.print_usage:
+ print(f"Total token so far: {self.total_token_usage}. Total cost so far: $USD{self.total_cost:.5f}")
+
+ vlm_response_json = extract_data(vlm_response, "json")
+ vlm_response_json = json.loads(vlm_response_json)
+
+ img_to_show_base64 = parsed_screen["som_image_base64"]
+ if "Box ID" in vlm_response_json:
+ try:
+ bbox = parsed_screen["parsed_content_list"][int(vlm_response_json["Box ID"])]["bbox"]
+ vlm_response_json["box_centroid_coordinate"] = [int((bbox[0] + bbox[2]) / 2 * screen_width), int((bbox[1] + bbox[3]) / 2 * screen_height)]
+ img_to_show_data = base64.b64decode(img_to_show_base64)
+ img_to_show = Image.open(BytesIO(img_to_show_data))
+
+ draw = ImageDraw.Draw(img_to_show)
+ x, y = vlm_response_json["box_centroid_coordinate"]
+ radius = 10
+ draw.ellipse((x - radius, y - radius, x + radius, y + radius), fill='red')
+ draw.ellipse((x - radius*3, y - radius*3, x + radius*3, y + radius*3), fill=None, outline='red', width=2)
+
+ buffered = BytesIO()
+ img_to_show.save(buffered, format="PNG")
+ img_to_show_base64 = base64.b64encode(buffered.getvalue()).decode("utf-8")
+ except:
+ print(f"Error parsing: {vlm_response_json}")
+ pass
+ self.output_callback(f'', sender="bot")
+ self.output_callback(
+ f''
+ f' Parsed Screen elemetns by OmniParser'
+ f'
{screen_info}
'
+ f'',
+ sender="bot"
+ )
+ vlm_plan_str = ""
+ for key, value in vlm_response_json.items():
+ if key == "Reasoning":
+ vlm_plan_str += f'{value}'
+ else:
+ vlm_plan_str += f'\n{key}: {value}'
+
+ # construct the response so that anthropicExcutor can execute the tool
+ response_content = [BetaTextBlock(text=vlm_plan_str, type='text')]
+ if 'box_centroid_coordinate' in vlm_response_json:
+ move_cursor_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
+ input={'action': 'mouse_move', 'coordinate': vlm_response_json["box_centroid_coordinate"]},
+ name='computer', type='tool_use')
+ response_content.append(move_cursor_block)
+
+ if vlm_response_json["Next Action"] == "None":
+ print("Task paused/completed.")
+ elif vlm_response_json["Next Action"] == "type":
+ sim_content_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
+ input={'action': vlm_response_json["Next Action"], 'text': vlm_response_json["value"]},
+ name='computer', type='tool_use')
+ response_content.append(sim_content_block)
+ else:
+ sim_content_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
+ input={'action': vlm_response_json["Next Action"]},
+ name='computer', type='tool_use')
+ response_content.append(sim_content_block)
+ response_message = BetaMessage(id=f'toolu_{uuid.uuid4()}', content=response_content, model='', role='assistant', type='message', stop_reason='tool_use', usage=BetaUsage(input_tokens=0, output_tokens=0))
+ return response_message, vlm_response_json
+
+ def _api_response_callback(self, response: APIResponse):
+ self.api_response_callback(response)
+
+ def _get_system_prompt(self, screen_info: str = ""):
+ main_section = f"""
+You are using a Windows device.
+You are able to use a mouse and keyboard to interact with the computer based on the given task and screenshot.
+You can only interact with the desktop GUI (no terminal or application menu access).
+
+You may be given some history plan and actions, this is the response from the previous loop.
+You should carefully consider your plan base on the task, screenshot, and history actions.
+
+Here is the list of all detected bounding boxes by IDs on the screen and their description:{screen_info}
+
+Your available "Next Action" only include:
+- type: types a string of text.
+- left_click: move mouse to box id and left clicks.
+- right_click: move mouse to box id and right clicks.
+- double_click: move mouse to box id and double clicks.
+- hover: move mouse to box id.
+- scroll_up: scrolls the screen up to view previous content.
+- scroll_down: scrolls the screen down, when the desired button is not visible, or you need to see more content.
+- wait: waits for 1 second for the device to load or respond.
+
+Based on the visual information from the screenshot image and the detected bounding boxes, please determine the next action, the Box ID you should operate on (if action is one of 'type', 'hover', 'scroll_up', 'scroll_down', 'wait', there should be no Box ID field), and the value (if the action is 'type') in order to complete the task.
+
+Output format:
+```json
+{{
+ "Reasoning": str, # describe what is in the current screen, taking into account the history, then describe your step-by-step thoughts on how to achieve the task, choose one action from available actions at a time.
+ "Next Action": "action_type, action description" | "None" # one action at a time, describe it in short and precisely.
+ "Box ID": n,
+ "value": "xxx" # only provide value field if the action is type, else don't include value key
+}}
+```
+
+One Example:
+```json
+{{
+ "Reasoning": "The current screen shows google result of amazon, in previous action I have searched amazon on google. Then I need to click on the first search results to go to amazon.com.",
+ "Next Action": "left_click",
+ "Box ID": m
+}}
+```
+
+Another Example:
+```json
+{{
+ "Reasoning": "The current screen shows the front page of amazon. There is no previous action. Therefore I need to type "Apple watch" in the search bar.",
+ "Next Action": "type",
+ "Box ID": n,
+ "value": "Apple watch"
+}}
+```
+
+Another Example:
+```json
+{{
+ "Reasoning": "The current screen does not show 'submit' button, I need to scroll down to see if the button is available.",
+ "Next Action": "scroll_down",
+}}
+```
+
+IMPORTANT NOTES:
+1. You should only give a single action at a time.
+
+"""
+ thinking_model = "r1" in self.model
+ if not thinking_model:
+ main_section += """
+2. You should give an analysis to the current screen, and reflect on what has been done by looking at the history, then describe your step-by-step thoughts on how to achieve the task.
+
+"""
+ else:
+ main_section += """
+2. In XML tags give an analysis to the current screen, and reflect on what has been done by looking at the history, then describe your step-by-step thoughts on how to achieve the task. In
-
', sender="bot")
+ self.output_callback(
+ f'