import asyncio
import base64
import hashlib
import io
import json
import logging
import os
import re
import sys
import time
import traceback
import warnings
from typing import (
Any,
AsyncGenerator,
Dict,
List,
Optional,
Sequence,
)
from urllib.parse import quote_plus
import aiofiles
import PIL.Image
from agentopera.chatflow.agents import BaseChatAgent
from agentopera.chatflow.base import Response
from agentopera.chatflow.messages import AgentEvent, ChatMessage, MultiModalMessage, TextMessage
from agentopera.chatflow.utils import content_to_str, remove_images
from agentopera.engine import EVENT_LOGGER_NAME
from agentopera.engine.types.agent import CancellationToken
from agentopera.engine.function_call import FunctionCall
from agentopera.engine.types import Image as AGImage
from agentopera.engine.types.models import (
AssistantMessage,
ChatCompletionClient,
LLMMessage,
ModelFamily,
RequestUsage,
SystemMessage,
UserMessage,
)
from PIL import Image
from playwright.async_api import BrowserContext, Download, Page, Playwright, async_playwright
from pydantic import BaseModel
from typing_extensions import Self
from .events import WebSurferEvent
from .prompts import (
WEB_SURFER_QA_PROMPT,
WEB_SURFER_QA_SYSTEM_MESSAGE,
WEB_SURFER_TOOL_PROMPT_MM,
WEB_SURFER_TOOL_PROMPT_TEXT,
)
from .set_of_mark import add_set_of_mark
from .tool_definitions import (
TOOL_CLICK,
TOOL_HISTORY_BACK,
TOOL_HOVER,
TOOL_READ_PAGE_AND_ANSWER,
TOOL_SCROLL_DOWN,
TOOL_SCROLL_UP,
TOOL_SLEEP,
TOOL_SUMMARIZE_PAGE,
TOOL_TYPE,
TOOL_VISIT_URL,
TOOL_WEB_SEARCH,
)
from .types import InteractiveRegion, UserContent
from .playwright_controller import PlaywrightController
DEFAULT_CONTEXT_SIZE = 128000
[docs]
class MultimodalWebSurfer(BaseChatAgent):
"""
MultimodalWebSurfer is a multimodal agent that acts as a web surfer that can search the web and visit web pages.
Installation:
.. code-block:: bash
pip install "agentopera[web-surfer]"
It launches a chromium browser and allows the playwright to interact with the web browser and can perform a variety of actions. The browser is launched on the first call to the agent and is reused for subsequent calls.
It must be used with a multimodal model client that supports function/tool calling, ideally GPT-4o currently.
When :meth:`on_messages` or :meth:`on_messages_stream` is called, the following occurs:
1) If this is the first call, the browser is initialized and the page is loaded. This is done in :meth:`_lazy_init`. The browser is only closed when :meth:`close` is called.
2) The method :meth:`_generate_reply` is called, which then creates the final response as below.
3) The agent takes a screenshot of the page, extracts the interactive elements, and prepares a set-of-mark screenshot with bounding boxes around the interactive elements.
4) The agent makes a call to the :attr:`model_client` with the SOM screenshot, history of messages, and the list of available tools.
- If the model returns a string, the agent returns the string as the final response.
- If the model returns a list of tool calls, the agent executes the tool calls with :meth:`_execute_tool` using :attr:`_playwright_controller`.
- The agent returns a final response which includes a screenshot of the page, page metadata, description of the action taken and the inner text of the webpage.
5) If at any point the agent encounters an error, it returns the error message as the final response.
.. note::
Please note that using the MultimodalWebSurfer involves interacting with a digital world designed for humans, which carries inherent risks.
Be aware that agents may occasionally attempt risky actions, such as recruiting humans for help or accepting cookie agreements without human involvement. Always ensure agents are monitored and operate within a controlled environment to prevent unintended consequences.
Moreover, be cautious that MultimodalWebSurfer may be susceptible to prompt injection attacks from webpages.
.. note::
On Windows, the event loop policy must be set to `WindowsProactorEventLoopPolicy` to avoid issues with subprocesses.
.. code-block:: python
import sys
import asyncio
if sys.platform == "win32":
asyncio.set_event_loop_policy(asyncio.WindowsProactorEventLoopPolicy())
Args:
name (str): The name of the agent.
model_client (ChatCompletionClient): The model client used by the agent. Must be multimodal and support function calling.
downloads_folder (str, optional): The folder where downloads are saved. Defaults to None, no downloads are saved.
description (str, optional): The description of the agent. Defaults to MultimodalWebSurfer.DEFAULT_DESCRIPTION.
debug_dir (str, optional): The directory where debug information is saved. Defaults to None.
headless (bool, optional): Whether the browser should be headless. Defaults to True.
start_page (str, optional): The start page for the browser. Defaults to MultimodalWebSurfer.DEFAULT_START_PAGE.
animate_actions (bool, optional): Whether to animate actions. Defaults to False.
to_save_screenshots (bool, optional): Whether to save screenshots. Defaults to False.
use_ocr (bool, optional): Whether to use OCR. Defaults to False.
browser_channel (str, optional): The browser channel. Defaults to None.
browser_data_dir (str, optional): The browser data directory. Defaults to None.
to_resize_viewport (bool, optional): Whether to resize the viewport. Defaults to True.
playwright (Playwright, optional): The playwright instance. Defaults to None.
context (BrowserContext, optional): The browser context. Defaults to None.
Example usage:
The following example demonstrates how to create a web surfing agent with
a model client and run it for multiple turns.
.. code-block:: python
import asyncio
from agentopera.chatflow.ui import Console
from agentopera.chatflow.teams import RoundRobinGroupChat
from agentopera.models.openai import OpenAIChatCompletionClient
from agentopera.agents.agents.web_surfer import MultimodalWebSurfer
async def main() -> None:
# Define an agent
web_surfer_agent = MultimodalWebSurfer(
name="MultimodalWebSurfer",
model_client=OpenAIChatCompletionClient(model="gpt-4o-2024-08-06"),
)
# Define a team
agent_team = RoundRobinGroupChat([web_surfer_agent], max_turns=3)
# Run the team and stream messages to the console
stream = agent_team.run_stream(task="Navigate to the agentopera readme on GitHub.")
await Console(stream)
# Close the browser controlled by the agent
await web_surfer_agent.close()
asyncio.run(main())
"""
DEFAULT_DESCRIPTION = """
A helpful assistant with access to a web browser.
Ask them to perform web searches, open pages, and interact with content (e.g., clicking links, scrolling the viewport, filling in form fields, etc.).
It can also summarize the entire page, or answer questions based on the content of the page.
It can also be asked to sleep and wait for pages to load, in cases where the page seems not yet fully loaded.
"""
DEFAULT_START_PAGE = "https://www.bing.com/"
# Viewport dimensions
VIEWPORT_HEIGHT = 900
VIEWPORT_WIDTH = 1440
# Size of the image we send to the MLM
# Current values represent a 0.85 scaling to fit within the GPT-4v short-edge constraints (768px)
MLM_HEIGHT = 765
MLM_WIDTH = 1224
SCREENSHOT_TOKENS = 1105
def __init__(
self,
name: str,
model_client: ChatCompletionClient,
downloads_folder: str | None = None,
description: str = DEFAULT_DESCRIPTION,
debug_dir: str | None = None,
headless: bool = True,
start_page: str | None = DEFAULT_START_PAGE,
animate_actions: bool = False,
to_save_screenshots: bool = False,
use_ocr: bool = False,
browser_channel: str | None = None,
browser_data_dir: str | None = None,
to_resize_viewport: bool = True,
playwright: Playwright | None = None,
context: BrowserContext | None = None,
):
"""
Initialize the MultimodalWebSurfer.
"""
super().__init__(name, description)
if debug_dir is None and to_save_screenshots:
raise ValueError(
"Cannot save screenshots without a debug directory. Set it using the 'debug_dir' parameter. The debug directory is created if it does not exist."
)
if model_client.model_info["function_calling"] is False:
raise ValueError(
"The model does not support function calling. MultimodalWebSurfer requires a model that supports function calling."
)
self._model_client = model_client
self.headless = headless
self.browser_channel = browser_channel
self.browser_data_dir = browser_data_dir
self.start_page = start_page or self.DEFAULT_START_PAGE
self.downloads_folder = downloads_folder
self.debug_dir = debug_dir
self.to_save_screenshots = to_save_screenshots
self.use_ocr = use_ocr
self.to_resize_viewport = to_resize_viewport
self.animate_actions = animate_actions
# Call init to set these in case not set
self._playwright: Playwright | None = playwright
self._context: BrowserContext | None = context
self._page: Page | None = None
self._last_download: Download | None = None
self._prior_metadata_hash: str | None = None
self.logger = logging.getLogger(EVENT_LOGGER_NAME + f".{self.name}.MultimodalWebSurfer")
self._chat_history: List[LLMMessage] = []
# Define the download handler
def _download_handler(download: Download) -> None:
self._last_download = download
self._download_handler = _download_handler
# Define the Playwright controller that handles the browser interactions
self._playwright_controller = PlaywrightController(
animate_actions=self.animate_actions,
downloads_folder=self.downloads_folder,
viewport_width=self.VIEWPORT_WIDTH,
viewport_height=self.VIEWPORT_HEIGHT,
_download_handler=self._download_handler,
to_resize_viewport=self.to_resize_viewport,
)
self.default_tools = [
TOOL_VISIT_URL,
TOOL_WEB_SEARCH,
TOOL_HISTORY_BACK,
TOOL_CLICK,
TOOL_TYPE,
TOOL_READ_PAGE_AND_ANSWER,
TOOL_SUMMARIZE_PAGE,
TOOL_SLEEP,
TOOL_HOVER,
]
self.did_lazy_init = False # flag to check if we have initialized the browser
async def _lazy_init(
self,
) -> None:
"""
On the first call, we initialize the browser and the page.
"""
# Check the current event loop policy if on windows.
if sys.platform == "win32":
current_policy = asyncio.get_event_loop_policy()
if hasattr(asyncio, "WindowsProactorEventLoopPolicy") and not isinstance(
current_policy, asyncio.WindowsProactorEventLoopPolicy
):
warnings.warn(
"The current event loop policy is not WindowsProactorEventLoopPolicy. "
"This may cause issues with subprocesses. "
"Try setting the event loop policy to WindowsProactorEventLoopPolicy. "
"For example: `asyncio.set_event_loop_policy(asyncio.WindowsProactorEventLoopPolicy())`. "
"See https://docs.python.org/3/library/asyncio-eventloop.html#asyncio.ProactorEventLoop.",
stacklevel=2,
)
self._last_download = None
self._prior_metadata_hash = None
# Create the playwright self
launch_args: Dict[str, Any] = {"headless": self.headless}
if self.browser_channel is not None:
launch_args["channel"] = self.browser_channel
if self._playwright is None:
self._playwright = await async_playwright().start()
# Create the context -- are we launching persistent?
if self._context is None:
if self.browser_data_dir is None:
browser = await self._playwright.chromium.launch(**launch_args)
self._context = await browser.new_context(
user_agent="Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/122.0.0.0 Safari/537.36 Edg/122.0.0.0"
)
else:
self._context = await self._playwright.chromium.launch_persistent_context(
self.browser_data_dir, **launch_args
)
# Create the page
self._context.set_default_timeout(60000) # One minute
self._page = await self._context.new_page()
assert self._page is not None
# self._page.route(lambda x: True, self._route_handler)
self._page.on("download", self._download_handler)
if self.to_resize_viewport:
await self._page.set_viewport_size({"width": self.VIEWPORT_WIDTH, "height": self.VIEWPORT_HEIGHT})
await self._page.add_init_script(
path=os.path.join(os.path.abspath(os.path.dirname(__file__)), "page_script.js")
)
await self._page.goto(self.start_page)
await self._page.wait_for_load_state()
# Prepare the debug directory -- which stores the screenshots generated throughout the process
await self._set_debug_dir(self.debug_dir)
self.did_lazy_init = True
[docs]
async def close(self) -> None:
"""
Close the browser and the page.
Should be called when the agent is no longer needed.
"""
if self._page is not None:
await self._page.close()
self._page = None
if self._context is not None:
await self._context.close()
self._context = None
if self._playwright is not None:
await self._playwright.stop()
self._playwright = None
async def _set_debug_dir(self, debug_dir: str | None) -> None:
assert self._page is not None
if self.debug_dir is None:
return
if not os.path.isdir(self.debug_dir):
os.mkdir(self.debug_dir)
if self.to_save_screenshots:
current_timestamp = "_" + int(time.time()).__str__()
screenshot_png_name = "screenshot" + current_timestamp + ".png"
await self._page.screenshot(path=os.path.join(self.debug_dir, screenshot_png_name)) # type: ignore
self.logger.info(
WebSurferEvent(
source=self.name,
url=self._page.url,
message="Screenshot: " + screenshot_png_name,
)
)
@property
def produced_message_types(self) -> Sequence[type[ChatMessage]]:
return (MultiModalMessage,)
[docs]
async def on_reset(self, cancellation_token: CancellationToken) -> None:
if not self.did_lazy_init:
return
assert self._page is not None
self._chat_history.clear()
reset_prior_metadata, reset_last_download = await self._playwright_controller.visit_page(
self._page, self.start_page
)
if reset_last_download and self._last_download is not None:
self._last_download = None
if reset_prior_metadata and self._prior_metadata_hash is not None:
self._prior_metadata_hash = None
if self.to_save_screenshots:
current_timestamp = "_" + int(time.time()).__str__()
screenshot_png_name = "screenshot" + current_timestamp + ".png"
await self._page.screenshot(path=os.path.join(self.debug_dir, screenshot_png_name)) # type: ignore
self.logger.info(
WebSurferEvent(
source=self.name,
url=self._page.url,
message="Screenshot: " + screenshot_png_name,
)
)
self.logger.info(
WebSurferEvent(
source=self.name,
url=self._page.url,
message="Resetting browser.",
)
)
[docs]
async def on_messages(self, messages: Sequence[ChatMessage], cancellation_token: CancellationToken) -> Response:
async for message in self.on_messages_stream(messages, cancellation_token):
if isinstance(message, Response):
return message
raise AssertionError("The stream should have returned the final result.")
[docs]
async def on_messages_stream(
self, messages: Sequence[ChatMessage], cancellation_token: CancellationToken
) -> AsyncGenerator[AgentEvent | ChatMessage | Response, None]:
for chat_message in messages:
if isinstance(chat_message, TextMessage | MultiModalMessage):
self._chat_history.append(UserMessage(content=chat_message.content, source=chat_message.source))
else:
raise ValueError(f"Unexpected message in MultiModalWebSurfer: {chat_message}")
self.inner_messages: List[AgentEvent | ChatMessage] = []
self.model_usage: List[RequestUsage] = []
try:
content = await self._generate_reply(cancellation_token=cancellation_token)
self._chat_history.append(AssistantMessage(content=content_to_str(content), source=self.name))
final_usage = RequestUsage(
prompt_tokens=sum([u.prompt_tokens for u in self.model_usage]),
completion_tokens=sum([u.completion_tokens for u in self.model_usage]),
)
if isinstance(content, str):
yield Response(
chat_message=TextMessage(content=content, source=self.name, models_usage=final_usage),
inner_messages=self.inner_messages,
)
else:
yield Response(
chat_message=MultiModalMessage(content=content, source=self.name, models_usage=final_usage),
inner_messages=self.inner_messages,
)
except BaseException:
content = f"Web surfing error:\n\n{traceback.format_exc()}"
self._chat_history.append(AssistantMessage(content=content, source=self.name))
yield Response(chat_message=TextMessage(content=content, source=self.name))
async def _generate_reply(self, cancellation_token: CancellationToken) -> UserContent:
"""Generates the actual reply. First calls the LLM to figure out which tool to use, then executes the tool."""
# Lazy init, initialize the browser and the page on the first generate reply only
if not self.did_lazy_init:
await self._lazy_init()
assert self._page is not None
# Clone the messages, removing old screenshots
history: List[LLMMessage] = remove_images(self._chat_history)
# Split the history, removing the last message
if len(history):
user_request = history.pop()
else:
user_request = UserMessage(content="Empty request.", source="user")
# Truncate the history for smaller models
if self._model_client.model_info["family"] not in [
ModelFamily.GPT_4O,
ModelFamily.O1,
ModelFamily.O3,
ModelFamily.GPT_4,
ModelFamily.GPT_35,
]:
history = []
# Ask the page for interactive elements, then prepare the state-of-mark screenshot
rects = await self._playwright_controller.get_interactive_rects(self._page)
viewport = await self._playwright_controller.get_visual_viewport(self._page)
screenshot = await self._page.screenshot()
som_screenshot, visible_rects, rects_above, rects_below = add_set_of_mark(screenshot, rects)
if self.to_save_screenshots:
current_timestamp = "_" + int(time.time()).__str__()
screenshot_png_name = "screenshot_som" + current_timestamp + ".png"
som_screenshot.save(os.path.join(self.debug_dir, screenshot_png_name)) # type: ignore
self.logger.info(
WebSurferEvent(
source=self.name,
url=self._page.url,
message="Screenshot: " + screenshot_png_name,
)
)
# What tools are available?
tools = self.default_tools.copy()
# We can scroll up
if viewport["pageTop"] > 5:
tools.append(TOOL_SCROLL_UP)
# Can scroll down
if (viewport["pageTop"] + viewport["height"] + 5) < viewport["scrollHeight"]:
tools.append(TOOL_SCROLL_DOWN)
# Focus hint
focused = await self._playwright_controller.get_focused_rect_id(self._page)
focused_hint = ""
if focused:
name = self._target_name(focused, rects)
if name:
name = f"(and name '{name}') "
else:
name = ""
role = "control"
try:
role = rects[focused]["role"]
except KeyError:
pass
focused_hint = f"\nThe {role} with ID {focused} {name}currently has the input focus.\n\n"
# Everything visible
visible_targets = "\n".join(self._format_target_list(visible_rects, rects)) + "\n\n"
# Everything else
other_targets: List[str] = []
other_targets.extend(self._format_target_list(rects_above, rects))
other_targets.extend(self._format_target_list(rects_below, rects))
if len(other_targets) > 0:
if len(other_targets) > 30:
other_targets = other_targets[0:30]
other_targets.append("...")
other_targets_str = (
"Additional valid interaction targets include (but are not limited to):\n"
+ "\n".join(other_targets)
+ "\n\n"
)
else:
other_targets_str = ""
state_description = "Your " + await self._get_state_description()
tool_names = "\n".join([t["name"] for t in tools])
page_title = await self._page.title()
prompt_message = None
if self._model_client.model_info["vision"]:
text_prompt = WEB_SURFER_TOOL_PROMPT_MM.format(
state_description=state_description,
visible_targets=visible_targets,
other_targets_str=other_targets_str,
focused_hint=focused_hint,
tool_names=tool_names,
title=page_title,
url=self._page.url,
).strip()
# Scale the screenshot for the MLM, and close the original
scaled_screenshot = som_screenshot.resize((self.MLM_WIDTH, self.MLM_HEIGHT))
som_screenshot.close()
if self.to_save_screenshots:
scaled_screenshot.save(os.path.join(self.debug_dir, "screenshot_scaled.png")) # type: ignore
# Create the message
prompt_message = UserMessage(
content=[re.sub(r"(\n\s*){3,}", "\n\n", text_prompt), AGImage.from_pil(scaled_screenshot)],
source=self.name,
)
else:
text_prompt = WEB_SURFER_TOOL_PROMPT_TEXT.format(
state_description=state_description,
visible_targets=visible_targets,
other_targets_str=other_targets_str,
focused_hint=focused_hint,
tool_names=tool_names,
title=page_title,
url=self._page.url,
).strip()
# Create the message
prompt_message = UserMessage(content=re.sub(r"(\n\s*){3,}", "\n\n", text_prompt), source=self.name)
history.append(prompt_message)
history.append(user_request)
# {history[-2].content if isinstance(history[-2].content, str) else history[-2].content[0]}
# print(f"""
# ================={len(history)}=================
# {history[-2].content}
# =====
# {history[-1].content}
# ===================================================
# """)
# Make the request
response = await self._model_client.create(
history, tools=tools, extra_create_args={"tool_choice": "auto"}, cancellation_token=cancellation_token
) # , "parallel_tool_calls": False})
self.model_usage.append(response.usage)
message = response.content
self._last_download = None
if isinstance(message, str):
# Answer directly
self.inner_messages.append(TextMessage(content=message, source=self.name))
return message
elif isinstance(message, list):
# Take an action
return await self._execute_tool(message, rects, tool_names, cancellation_token=cancellation_token)
else:
# Not sure what happened here
raise AssertionError(f"Unknown response format '{message}'")
async def _execute_tool(
self,
message: List[FunctionCall],
rects: Dict[str, InteractiveRegion],
tool_names: str,
cancellation_token: Optional[CancellationToken] = None,
) -> UserContent:
# Execute the tool
name = message[0].name
args = json.loads(message[0].arguments)
action_description = ""
assert self._page is not None
self.logger.info(
WebSurferEvent(
source=self.name,
url=self._page.url,
action=name,
arguments=args,
message=f"{name}( {json.dumps(args)} )",
)
)
self.inner_messages.append(TextMessage(content=f"{name}( {json.dumps(args)} )", source=self.name))
if name == "visit_url":
url = args.get("url")
action_description = f"I typed '{url}' into the browser address bar."
# Check if the argument starts with a known protocol
if url.startswith(("https://", "http://", "file://", "about:")):
reset_prior_metadata, reset_last_download = await self._playwright_controller.visit_page(
self._page, url
)
# If the argument contains a space, treat it as a search query
elif " " in url:
reset_prior_metadata, reset_last_download = await self._playwright_controller.visit_page(
self._page, f"https://www.bing.com/search?q={quote_plus(url)}&FORM=QBLH"
)
# Otherwise, prefix with https://
else:
reset_prior_metadata, reset_last_download = await self._playwright_controller.visit_page(
self._page, "https://" + url
)
if reset_last_download and self._last_download is not None:
self._last_download = None
if reset_prior_metadata and self._prior_metadata_hash is not None:
self._prior_metadata_hash = None
elif name == "history_back":
action_description = "I clicked the browser back button."
await self._playwright_controller.back(self._page)
elif name == "web_search":
query = args.get("query")
action_description = f"I typed '{query}' into the browser search bar."
reset_prior_metadata, reset_last_download = await self._playwright_controller.visit_page(
self._page, f"https://www.bing.com/search?q={quote_plus(query)}&FORM=QBLH"
)
if reset_last_download and self._last_download is not None:
self._last_download = None
if reset_prior_metadata and self._prior_metadata_hash is not None:
self._prior_metadata_hash = None
elif name == "scroll_up":
action_description = "I scrolled up one page in the browser."
await self._playwright_controller.page_up(self._page)
elif name == "scroll_down":
action_description = "I scrolled down one page in the browser."
await self._playwright_controller.page_down(self._page)
elif name == "click":
target_id = str(args.get("target_id"))
target_name = self._target_name(target_id, rects)
if target_name:
action_description = f"I clicked '{target_name}'."
else:
action_description = "I clicked the control."
new_page_tentative = await self._playwright_controller.click_id(self._page, target_id)
if new_page_tentative is not None:
self._page = new_page_tentative
self._prior_metadata_hash = None
self.logger.info(
WebSurferEvent(
source=self.name,
url=self._page.url,
message="New tab or window.",
)
)
elif name == "input_text":
input_field_id = str(args.get("input_field_id"))
text_value = str(args.get("text_value"))
input_field_name = self._target_name(input_field_id, rects)
if input_field_name:
action_description = f"I typed '{text_value}' into '{input_field_name}'."
else:
action_description = f"I input '{text_value}'."
await self._playwright_controller.fill_id(self._page, input_field_id, text_value)
elif name == "scroll_element_up":
target_id = str(args.get("target_id"))
target_name = self._target_name(target_id, rects)
if target_name:
action_description = f"I scrolled '{target_name}' up."
else:
action_description = "I scrolled the control up."
await self._playwright_controller.scroll_id(self._page, target_id, "up")
elif name == "scroll_element_down":
target_id = str(args.get("target_id"))
target_name = self._target_name(target_id, rects)
if target_name:
action_description = f"I scrolled '{target_name}' down."
else:
action_description = "I scrolled the control down."
await self._playwright_controller.scroll_id(self._page, target_id, "down")
elif name == "answer_question":
question = str(args.get("question"))
action_description = f"I answered the following question '{question}' based on the web page."
# Do Q&A on the DOM. No need to take further action. Browser state does not change.
return await self._summarize_page(question=question, cancellation_token=cancellation_token)
elif name == "summarize_page":
# Summarize the DOM. No need to take further action. Browser state does not change.
action_description = "I summarized the current web page"
return await self._summarize_page(cancellation_token=cancellation_token)
elif name == "hover":
target_id = str(args.get("target_id"))
target_name = self._target_name(target_id, rects)
if target_name:
action_description = f"I hovered over '{target_name}'."
else:
action_description = "I hovered over the control."
await self._playwright_controller.hover_id(self._page, target_id)
elif name == "sleep":
action_description = "I am waiting a short period of time before taking further action."
await self._playwright_controller.sleep(self._page, 3)
else:
raise ValueError(f"Unknown tool '{name}'. Please choose from:\n\n{tool_names}")
await self._page.wait_for_load_state()
await self._playwright_controller.sleep(self._page, 3)
# Handle downloads
if self._last_download is not None and self.downloads_folder is not None:
fname = os.path.join(self.downloads_folder, self._last_download.suggested_filename)
await self._last_download.save_as(fname) # type: ignore
page_body = f"<html><head><title>Download Successful</title></head><body style=\"margin: 20px;\"><h1>Successfully downloaded '{self._last_download.suggested_filename}' to local path:<br><br>{fname}</h1></body></html>"
await self._page.goto(
"data:text/html;base64," + base64.b64encode(page_body.encode("utf-8")).decode("utf-8")
)
await self._page.wait_for_load_state()
# Handle metadata
page_metadata = json.dumps(await self._playwright_controller.get_page_metadata(self._page), indent=4)
metadata_hash = hashlib.md5(page_metadata.encode("utf-8")).hexdigest()
if metadata_hash != self._prior_metadata_hash:
page_metadata = (
"\n\nThe following metadata was extracted from the webpage:\n\n" + page_metadata.strip() + "\n"
)
else:
page_metadata = ""
self._prior_metadata_hash = metadata_hash
new_screenshot = await self._page.screenshot()
if self.to_save_screenshots:
current_timestamp = "_" + int(time.time()).__str__()
screenshot_png_name = "screenshot" + current_timestamp + ".png"
async with aiofiles.open(os.path.join(self.debug_dir, screenshot_png_name), "wb") as file: # type: ignore
await file.write(new_screenshot) # type: ignore
self.logger.info(
WebSurferEvent(
source=self.name,
url=self._page.url,
message="Screenshot: " + screenshot_png_name,
)
)
# Return the complete observation
state_description = "The " + await self._get_state_description()
message_content = (
f"{action_description}\n\n" + state_description + page_metadata + "\nHere is a screenshot of the page."
)
return [
re.sub(r"(\n\s*){3,}", "\n\n", message_content), # Removing blank lines
AGImage.from_pil(PIL.Image.open(io.BytesIO(new_screenshot))),
]
async def _get_state_description(self) -> str:
assert self._playwright_controller is not None
assert self._page is not None
# Describe the viewport of the new page in words
viewport = await self._playwright_controller.get_visual_viewport(self._page)
percent_visible = int(viewport["height"] * 100 / viewport["scrollHeight"])
percent_scrolled = int(viewport["pageTop"] * 100 / viewport["scrollHeight"])
if percent_scrolled < 1: # Allow some rounding error
position_text = "at the top of the page"
elif percent_scrolled + percent_visible >= 99: # Allow some rounding error
position_text = "at the bottom of the page"
else:
position_text = str(percent_scrolled) + "% down from the top of the page"
visible_text = await self._playwright_controller.get_visible_text(self._page)
# Return the complete observation
page_title = await self._page.title()
message_content = f"web browser is open to the page [{page_title}]({self._page.url}).\nThe viewport shows {percent_visible}% of the webpage, and is positioned {position_text}\n"
message_content += f"The following text is visible in the viewport:\n\n{visible_text}"
return message_content
def _target_name(self, target: str, rects: Dict[str, InteractiveRegion]) -> str | None:
try:
return rects[target]["aria_name"].strip()
except KeyError:
return None
def _format_target_list(self, ids: List[str], rects: Dict[str, InteractiveRegion]) -> List[str]:
"""
Format the list of targets in the webpage as a string to be used in the agent's prompt.
"""
targets: List[str] = []
for r in list(set(ids)):
if r in rects:
# Get the role
aria_role = rects[r].get("role", "").strip()
if len(aria_role) == 0:
aria_role = rects[r].get("tag_name", "").strip()
# Get the name
aria_name = re.sub(r"[\n\r]+", " ", rects[r].get("aria_name", "")).strip()
# What are the actions?
actions = ['"click", "hover"']
if rects[r]["role"] in ["textbox", "searchbox", "search"]:
actions = ['"input_text"']
actions_str = "[" + ",".join(actions) + "]"
targets.append(f'{{"id": {r}, "name": "{aria_name}", "role": "{aria_role}", "tools": {actions_str} }}')
return targets
async def _summarize_page(
self,
question: str | None = None,
cancellation_token: Optional[CancellationToken] = None,
) -> str:
assert self._page is not None
page_markdown: str = await self._playwright_controller.get_page_markdown(self._page)
title: str = self._page.url
try:
title = await self._page.title()
except Exception:
pass
# Take a screenshot and scale it
screenshot = Image.open(io.BytesIO(await self._page.screenshot()))
scaled_screenshot = screenshot.resize((self.MLM_WIDTH, self.MLM_HEIGHT))
screenshot.close()
ag_image = AGImage.from_pil(scaled_screenshot)
# Prepare the system prompt
messages: List[LLMMessage] = []
messages.append(SystemMessage(content=WEB_SURFER_QA_SYSTEM_MESSAGE))
prompt = WEB_SURFER_QA_PROMPT(title, question)
# Grow the buffer (which is added to the prompt) until we overflow the context window or run out of lines
buffer = ""
# for line in re.split(r"([\r\n]+)", page_markdown):
for line in page_markdown.splitlines():
trial_message = UserMessage(
content=prompt + buffer + line,
source=self.name,
)
try:
remaining = self._model_client.remaining_tokens(messages + [trial_message])
except KeyError:
# Use the default if the model isn't found
remaining = DEFAULT_CONTEXT_SIZE - self._model_client.count_tokens(messages + [trial_message])
if self._model_client.model_info["vision"] and remaining <= 0:
break
if self._model_client.model_info["vision"] and remaining <= self.SCREENSHOT_TOKENS:
break
buffer += line
# Nothing to do
buffer = buffer.strip()
if len(buffer) == 0:
return "Nothing to summarize."
# Append the message
if self._model_client.model_info["vision"]:
# Multimodal
messages.append(
UserMessage(
content=[
prompt + buffer,
ag_image,
],
source=self.name,
)
)
else:
# Text only
messages.append(
UserMessage(
content=prompt + buffer,
source=self.name,
)
)
# Generate the response
response = await self._model_client.create(messages, cancellation_token=cancellation_token)
self.model_usage.append(response.usage)
scaled_screenshot.close()
assert isinstance(response.content, str)
return response.content