IAGO / owl /camel /toolkits /web_toolkit.py
lazychih114
update
c413558
raw
history blame
44.5 kB
from playwright.sync_api import sync_playwright, Page, BrowserContext, Browser
from typing import Union, List, Dict, Any, Tuple, cast, Optional, Literal
from playwright._impl._errors import Error as PlaywrightError
from playwright._impl._errors import TimeoutError
from loguru import logger
from typing import Any, Dict, List, TypedDict, Union, BinaryIO
from PIL import Image, ImageDraw, ImageFont
from firecrawl import FirecrawlApp
from html2text import html2text
from retry import retry
from copy import deepcopy
from camel.toolkits.base import BaseToolkit
from camel.toolkits import FunctionTool, VideoAnalysisToolkit
from camel.messages import BaseMessage
from camel.agents import ChatAgent
from camel.models import ModelFactory
from camel.types import ModelType, ModelPlatformType
import io
import random
import os
import json
import shutil
import datetime
import time
import requests
import re
TOP_NO_LABEL_ZONE = 20
AVAILABLE_ACTIONS_PROMPT = """
1. `fill_input_id(identifier: Union[str, int], text: str)`: Fill an input field (e.g. search box) with the given text and press Enter.
2. `click_id(identifier: Union[str, int])`: Click an element with the given ID.
3. `hover_id(identifier: Union[str, int])`: Hover over an element with the given ID.
4. `download_file_id(identifier: Union[str, int])`: Download a file with the given ID. It returns the path to the downloaded file. If the file is successfully downloaded, you can stop the simulation and report the path to the downloaded file for further processing.
5. `scroll_to_bottom()`: Scroll to the bottom of the page.
6. `scroll_to_top()`: Scroll to the top of the page.
7. `scroll_up()`: Scroll up the page. It is suitable when you want to see the elements above the current viewport.
8. `scroll_down()`: Scroll down the page. It is suitable when you want to see the elements below the current viewport. If the webpage does not change, It means that the webpage has scrolled to the bottom.
9. `back()`: Navigate back to the previous page. This is useful when you want to go back to the previous page, as current page is not useful.
10. `stop()`: Stop the action process, because the task is completed or failed (impossible to find the answer). In this situation, you should provide your answer in your output.
11. `get_url()`: Get the current URL of the current page.
12. `find_text_on_page(search_text: str)`: Find the next given text on the current whole page, and scroll the page to the targeted text. It is equivalent to pressing Ctrl + F and searching for the text, and is powerful when you want to fast-check whether the current page contains some specific text.
13. `visit_page(url: str)`: Go to the specific url page.
14. `click_blank_area()`: Click a blank area of the page to unfocus the current element. It is useful when you have clicked an element but it cannot unfocus itself (e.g. Menu bar) to automatically render the updated webpage.
15. `ask_question_about_video(question: str)`: Ask a question about the current webpage which contains video, e.g. youtube websites.
"""
ACTION_WITH_FEEDBACK_LIST = [
'ask_question_about_video',
'download_file_id'
]
# codes from magentic-one
class DOMRectangle(TypedDict):
x: Union[int, float]
y: Union[int, float]
width: Union[int, float]
height: Union[int, float]
top: Union[int, float]
right: Union[int, float]
bottom: Union[int, float]
left: Union[int, float]
class VisualViewport(TypedDict):
height: Union[int, float]
width: Union[int, float]
offsetLeft: Union[int, float]
offsetTop: Union[int, float]
pageLeft: Union[int, float]
pageTop: Union[int, float]
scale: Union[int, float]
clientWidth: Union[int, float]
clientHeight: Union[int, float]
scrollWidth: Union[int, float]
scrollHeight: Union[int, float]
class InteractiveRegion(TypedDict):
tag_name: str
role: str
aria_name: str
v_scrollable: bool
rects: List[DOMRectangle]
def _get_str(d: Any, k: str) -> str:
val = d[k]
assert isinstance(val, str)
return val
def _get_number(d: Any, k: str) -> Union[int, float]:
val = d[k]
assert isinstance(val, int) or isinstance(val, float)
return val
def _get_bool(d: Any, k: str) -> bool:
val = d[k]
assert isinstance(val, bool)
return val
def _parse_json_output(text: str) -> Dict[str, Any]:
# judge if text is markdown format (```json ````)
if text.startswith("```json") and text.endswith("```"):
text = text.replace("```json", "").replace("```", "").strip()
try:
return json.loads(text)
except json.JSONDecodeError:
logger.warning(f"Failed to parse JSON output: {text}")
return {}
def _reload_image(image: Image.Image):
buffer = io.BytesIO()
image.save(buffer, format="PNG")
buffer.seek(0)
return Image.open(buffer)
def domrectangle_from_dict(rect: Dict[str, Any]) -> DOMRectangle:
return DOMRectangle(
x=_get_number(rect, "x"),
y=_get_number(rect, "y"),
width=_get_number(rect, "width"),
height=_get_number(rect, "height"),
top=_get_number(rect, "top"),
right=_get_number(rect, "right"),
bottom=_get_number(rect, "bottom"),
left=_get_number(rect, "left"),
)
def interactiveregion_from_dict(region: Dict[str, Any]) -> InteractiveRegion:
typed_rects: List[DOMRectangle] = []
for rect in region["rects"]:
typed_rects.append(domrectangle_from_dict(rect))
return InteractiveRegion(
tag_name=_get_str(region, "tag_name"),
role=_get_str(region, "role"),
aria_name=_get_str(region, "aria-name"),
v_scrollable=_get_bool(region, "v-scrollable"),
rects=typed_rects,
)
def visualviewport_from_dict(viewport: Dict[str, Any]) -> VisualViewport:
return VisualViewport(
height=_get_number(viewport, "height"),
width=_get_number(viewport, "width"),
offsetLeft=_get_number(viewport, "offsetLeft"),
offsetTop=_get_number(viewport, "offsetTop"),
pageLeft=_get_number(viewport, "pageLeft"),
pageTop=_get_number(viewport, "pageTop"),
scale=_get_number(viewport, "scale"),
clientWidth=_get_number(viewport, "clientWidth"),
clientHeight=_get_number(viewport, "clientHeight"),
scrollWidth=_get_number(viewport, "scrollWidth"),
scrollHeight=_get_number(viewport, "scrollHeight"),
)
def add_set_of_mark(
screenshot: bytes | Image.Image | io.BufferedIOBase, ROIs: Dict[str, InteractiveRegion]
) -> Tuple[Image.Image, List[str], List[str], List[str]]:
if isinstance(screenshot, Image.Image):
return _add_set_of_mark(screenshot, ROIs)
if isinstance(screenshot, bytes):
screenshot = io.BytesIO(screenshot)
# TODO: Not sure why this cast was needed, but by this point screenshot is a binary file-like object
image = Image.open(cast(BinaryIO, screenshot))
comp, visible_rects, rects_above, rects_below = _add_set_of_mark(image, ROIs)
image.close()
return comp, visible_rects, rects_above, rects_below
def _add_set_of_mark(
screenshot: Image.Image, ROIs: Dict[str, InteractiveRegion]
) -> Tuple[Image.Image, List[str], List[str], List[str]]:
visible_rects: List[str] = list()
rects_above: List[str] = list() # Scroll up to see
rects_below: List[str] = list() # Scroll down to see
fnt = ImageFont.load_default(14)
base = screenshot.convert("L").convert("RGBA")
overlay = Image.new("RGBA", base.size)
draw = ImageDraw.Draw(overlay)
for r in ROIs:
for rect in ROIs[r]["rects"]:
# Empty rectangles
if not rect:
continue
if rect["width"] * rect["height"] == 0:
continue
mid = ((rect["right"] + rect["left"]) / 2.0, (rect["top"] + rect["bottom"]) / 2.0)
if 0 <= mid[0] and mid[0] < base.size[0]:
if mid[1] < 0:
rects_above.append(r)
elif mid[1] >= base.size[1]:
rects_below.append(r)
else:
visible_rects.append(r)
_draw_roi(draw, int(r), fnt, rect)
comp = Image.alpha_composite(base, overlay)
overlay.close()
return comp, visible_rects, rects_above, rects_below
def _draw_roi(
draw: ImageDraw.ImageDraw, idx: int, font: ImageFont.FreeTypeFont | ImageFont.ImageFont, rect: DOMRectangle
) -> None:
color = _color(idx)
luminance = color[0] * 0.3 + color[1] * 0.59 + color[2] * 0.11
text_color = (0, 0, 0, 255) if luminance > 90 else (255, 255, 255, 255)
roi = ((rect["left"], rect["top"]), (rect["right"], rect["bottom"]))
label_location = (rect["right"], rect["top"])
label_anchor = "rb"
if label_location[1] <= TOP_NO_LABEL_ZONE:
label_location = (rect["right"], rect["bottom"])
label_anchor = "rt"
draw.rectangle(roi, outline=color, fill=(color[0], color[1], color[2], 48), width=2)
bbox = draw.textbbox(label_location, str(idx), font=font, anchor=label_anchor, align="center") # type: ignore
bbox = (bbox[0] - 3, bbox[1] - 3, bbox[2] + 3, bbox[3] + 3)
draw.rectangle(bbox, fill=color)
draw.text(label_location, str(idx), fill=text_color, font=font, anchor=label_anchor, align="center") # type: ignore
def _color(identifier: int) -> Tuple[int, int, int, int]:
rnd = random.Random(int(identifier))
color = [rnd.randint(0, 255), rnd.randint(125, 255), rnd.randint(0, 50)]
rnd.shuffle(color)
color.append(255)
return cast(Tuple[int, int, int, int], tuple(color))
class BaseBrowser:
def __init__(self,
headless=True,
cache_dir: Optional[str] = None,
page_script_path: Optional[str] = None):
r"""Initialize the WebBrowserToolkit instance.
Args:
headless (bool): Whether to run the browser in headless mode.
cache_dir (Union[str, None]): The directory to store cache files.
Returns:
None
"""
self.history = []
self.playwright = sync_playwright().start()
self.browser: Browser = self.playwright.chromium.launch(headless=headless)
self.context: BrowserContext = self.browser.new_context(accept_downloads=True)
self.page: Page = self.context.new_page()
self.page_url: str = None # stores the current page URL
self.page_script: str = None
# self.page_content: str = None # stores the current page content
self.page_history = [] # stores the history of visited pages
# set the cache directory
self.cache_dir = "tmp/"
os.makedirs(self.cache_dir, exist_ok=True)
if cache_dir is not None:
self.cache_dir = cache_dir
# load the page script
if page_script_path is None:
abs_dir_path = os.path.dirname(os.path.abspath(__file__))
page_script_path = os.path.join(abs_dir_path, "page_script.js")
try:
with open(page_script_path, "r", encoding='utf-8') as f:
self.page_script = f.read()
f.close()
except FileNotFoundError:
logger.warning(f"Page script file not found: {page_script_path}")
# initialize the page into google search
# self.visit_page("https://www.google.com")
def clean_cache(self):
r"""delete the cache directory and its contents."""
if os.path.exists(self.cache_dir):
shutil.rmtree(self.cache_dir)
def _wait_for_load(self, timeout: int = 20):
r"""Wait for a certain amount of time for the page to load."""
timeout_ms = timeout * 1000
self.page.wait_for_load_state("load", timeout=timeout_ms)
# self.page.wait_for_load_state("networkidle", timeout=timeout_ms)
# self.page.wait_for_load_state("domcontentloaded", timeout=timeout_ms)
time.sleep(2)
def click_blank_area(self):
r"""Click a blank area of the page to unfocus the current element."""
self.page.mouse.click(0, 0)
self._wait_for_load()
def visit_page(self, url: str):
r"""Visit a page with the given URL."""
self.page.goto(url)
self._wait_for_load()
self.page_url = url
def ask_question_about_video(self, question: str) -> str:
r"""Ask a question about the video on the current page. It is suitable to process youtube video.
Args:
question (str): The question to ask.
Returns:
str: The answer to the question.
"""
video_analyzer = VideoAnalysisToolkit()
result = video_analyzer.ask_question_about_video(self.page_url, question)
return result
@retry(PlaywrightError, delay=1, logger=logger)
def get_screenshot(self, save_image: bool = False) -> Tuple[Image.Image, Union[str, None]]:
r"""Get a screenshot of the current page.
Args:
save_image (bool): Whether to save the image to the cache directory.
Returns:
Tuple[Image.Image, str]: A tuple containing the screenshot image and the path to the image file.
"""
image_data = self.page.screenshot(timeout=60000)
image = Image.open(io.BytesIO(image_data))
file_path = None
if save_image:
# get url name to form a file name
url_name = self.page_url.split("/")[-1].replace(".", "_").replace(":", "_")
# get formatted time: mmddhhmmss
timestamp = datetime.datetime.now().strftime("%m%d%H%M%S")
file_path = os.path.join(self.cache_dir, f"{url_name}_{timestamp}.png")
with open(file_path, "wb") as f:
image.save(f, "PNG")
f.close()
return image, file_path
def capture_full_page_screenshots(self, scroll_ratio: float = 0.8) -> List[str]:
r"""Capture full page screenshots by scrolling the page with a buffer zone.
Args:
scroll_ratio (float): The ratio of viewport height to scroll each step (default: 0.7).
Returns:
List[str]: A list of paths to the screenshot files.
"""
screenshots = []
scroll_height = self.page.evaluate("document.body.scrollHeight")
viewport_height = self.page.viewport_size["height"]
current_scroll = 0
screenshot_index = 1
url_name = self.page.url.split("/")[-1].replace(".", "_")
timestamp = datetime.datetime.now().strftime("%m%d%H%M%S")
base_file_path = os.path.join(self.cache_dir, f"{url_name}_{timestamp}")
max_height = scroll_height - viewport_height
scroll_step = int(viewport_height * scroll_ratio)
last_height = 0
while True:
logger.debug(f"Current scroll: {current_scroll}, max_height: {max_height}, step: {scroll_step}")
file_path = f"{base_file_path}_{screenshot_index}.png"
_, file_path = self.get_screenshot(save_image=True)
screenshots.append(file_path)
self.page.evaluate(f"window.scrollBy(0, {scroll_step})")
time.sleep(0.5)
current_scroll = self.page.evaluate("window.scrollY")
if abs(current_scroll - last_height) < viewport_height * 0.1:
break
last_height = current_scroll
screenshot_index += 1
return screenshots
def get_visual_viewport(self) -> VisualViewport:
r"""Get the visual viewport of the current page.
Returns:
VisualViewport: The visual viewport of the current page.
"""
try:
self.page.evaluate(self.page_script)
except Exception as e:
logger.warning(f"Error evaluating page script: {e}")
return visualviewport_from_dict(
self.page.evaluate("MultimodalWebSurfer.getVisualViewport();")
)
def get_interactive_elements(self) -> List[Dict[str, Any]]:
# codes from magentic-one
try:
self.page.evaluate(self.page_script)
except Exception as e:
logger.warning(f"Error evaluating page script: {e}")
result = cast(
Dict[str, Dict[str, Any]], self.page.evaluate("MultimodalWebSurfer.getInteractiveRects();")
)
typed_results: Dict[str, InteractiveRegion] = {}
for k in result:
typed_results[k] = interactiveregion_from_dict(result[k])
return typed_results
def get_som_screenshot(self, save_image: bool = False) -> Tuple[Image.Image, Union[str, None]]:
r"""Get a screenshot of the current viewport with interactive elements marked.
Args:
save_image (bool): Whether to save the image to the cache directory.
Returns:
Tuple[Image.Image, str]: A tuple containing the screenshot image and the path to the image file.
"""
self._wait_for_load()
screenshot, _ = self.get_screenshot(save_image=False)
rects = self.get_interactive_elements()
file_path = None
comp, visible_rects, rects_above, rects_below = add_set_of_mark(screenshot, rects)
if save_image:
url_name = self.page_url.split("/")[-1].replace(".", "_").replace(":", "_")
timestamp = datetime.datetime.now().strftime("%m%d%H%M%S")
file_path = os.path.join(self.cache_dir, f"{url_name}_{timestamp}.png")
with open(file_path, "wb") as f:
comp.save(f, "PNG")
f.close()
return comp, file_path
def scroll_up(self) -> None:
self.page.keyboard.press("PageUp")
def scroll_down(self) -> None:
self.page.keyboard.press("PageDown")
def get_url(self) -> str:
return self.page.url
def click_id(self, identifier: Union[str, int]):
if isinstance(identifier, int):
identifier = str(identifier)
target = self.page.locator(f"[__elementId='{identifier}']")
# See if it exists
try:
target.wait_for(timeout=5000)
except TimeoutError:
raise ValueError("No such element.") from None
# Click it
target.scroll_into_view_if_needed()
new_page = None
try:
with self.page.expect_event("popup", timeout=1000) as page_info:
box = cast(Dict[str, Union[int, float]], target.bounding_box())
self.page.mouse.click(box["x"] + box["width"] / 2, box["y"] + box["height"] / 2)
new_page = page_info.value
# If a new page is opened, switch to it
if new_page:
self.page_history.append(deepcopy(self.page.url))
self.page = new_page
except PlaywrightError:
pass
# new_page = self
self._wait_for_load()
def extract_url_content(self):
r"""Extract the content of the current page."""
# TODO: update it using firecrawl
content = self.page.content()
return content
def download_file_id(self, identifier: Union[str, int]) -> str:
r"""Download a file with the given selector.
Args:
identifier (str): The identifier of the file to download.
file_path (str): The path to save the downloaded file.
Returns:
str: The result of the action.
"""
if isinstance(identifier, int):
identifier = str(identifier)
try:
target = self.page.locator(f"[__elementId='{identifier}']")
except TimeoutError:
logger.warning(f"Element with identifier '{identifier}' not found.")
return f"Element with identifier '{identifier}' not found."
target.scroll_into_view_if_needed()
file_path = os.path.join(self.cache_dir)
self._wait_for_load()
try:
with self.page.expect_download() as download_info:
target.click()
download = download_info.value
file_name = download.suggested_filename
file_path = os.path.join(file_path, file_name)
download.save_as(file_path)
return f"Downloaded file to path '{file_path}'."
except PlaywrightError:
return f"Failed to download file with identifier '{identifier}'."
def fill_input_id(self, identifier: Union[str, int], text: str) -> str:
r""" Fill an input field with the given text, and then press Enter.
Args:
identifier (str): The identifier of the input field.
text (str): The text to fill.
Returns:
str: The result of the action.
"""
if isinstance(identifier, int):
identifier = str(identifier)
try:
target = self.page.locator(f"[__elementId='{identifier}']")
except TimeoutError:
logger.warning(f"Element with identifier '{identifier}' not found.")
return f"Element with identifier '{identifier}' not found."
target.scroll_into_view_if_needed()
target.focus()
try:
target.fill(text)
except PlaywrightError:
target.press_sequentially(text)
target.press("Enter")
self._wait_for_load()
return f"Filled input field '{identifier}' with text '{text}' and pressed Enter."
def scroll_to_bottom(self) -> str:
self.page.evaluate("window.scrollTo(0, document.body.scrollHeight);")
self._wait_for_load()
return "Scrolled to the bottom of the page."
def scroll_to_top(self) -> str:
self.page.evaluate("window.scrollTo(0, 0);")
self._wait_for_load()
return "Scrolled to the top of the page."
def hover_id(self, identifier: Union[str, int]) -> str:
r""" Hover over an element with the given identifier.
Args:
identifier (str): The identifier of the element to hover over.
Returns:
str: The result of the action.
"""
if isinstance(identifier, int):
identifier = str(identifier)
try:
target = self.page.locator(f"[__elementId='{identifier}']")
except TimeoutError:
logger.warning(f"Element with identifier '{identifier}' not found.")
return f"Element with identifier '{identifier}' not found."
target.scroll_into_view_if_needed()
target.hover()
self._wait_for_load()
return f"Hovered over element with identifier '{identifier}'."
def find_text_on_page(self, search_text: str) -> str:
r"""Find the next given text on the page, and scroll the page to the targeted text.
It is equivalent to pressing Ctrl + F and searching for the text."""
script = f"""
(function() {{
let text = "{search_text}";
let found = window.find(text);
if (!found) {{
let elements = document.querySelectorAll("*:not(script):not(style)");
for (let el of elements) {{
if (el.innerText && el.innerText.includes(text)) {{
el.scrollIntoView({{behavior: "smooth", block: "center"}});
el.style.backgroundColor = "yellow";
el.style.border = '2px solid red';
return true;
}}
}}
return false;
}}
return true;
}})();
"""
found = self.page.evaluate(script)
self._wait_for_load()
if found:
return f"Found text '{search_text}' on the page."
else:
return f"Text '{search_text}' not found on the page."
def back(self):
r"""Navigate back to the previous page."""
page_url_before = self.page.url
self.page.go_back()
page_url_after = self.page.url
if page_url_after == "about:blank":
self.visit_page(page_url_before)
if page_url_before == page_url_after:
# If the page is not changed, try to use the history
if len(self.page_history) > 0:
self.visit_page(self.page_history.pop())
time.sleep(1)
self._wait_for_load()
def close(self):
self.browser.close()
self.playwright.stop()
def show_interactive_elements(self):
r"""Show simple interactive elements on the current page."""
self.page.evaluate(self.page_script)
self.page.evaluate("""
() => {
document.querySelectorAll('a, button, input, select, textarea, [tabindex]:not([tabindex="-1"]), [contenteditable="true"]').forEach(el => {
el.style.border = '2px solid red';
});
}
""")
@retry(requests.RequestException)
def get_webpage_content(self) -> str:
self._wait_for_load()
html_content = self.page.content()
markdown_content = html2text(html_content)
return markdown_content
class WebToolkit(BaseToolkit):
def __init__(self,
headless=True,
cache_dir: Optional[str] = None,
page_script_path: Optional[str] = None,
model: Literal['gpt-4o', 'gpt-4o-mini'] = 'gpt-4o',
history_window: int = 5
):
self.browser = BaseBrowser(
headless=headless,
cache_dir=cache_dir,
page_script_path=page_script_path
)
self.history_window = history_window
self.history = []
# self.search_toolkit = SearchToolkit()
self.web_agent, self.planning_agent = self._initialize_agent(model)
def _reset(self):
self.web_agent.reset()
self.planning_agent.reset()
self.history = []
os.makedirs(self.browser.cache_dir, exist_ok=True)
def _initialize_agent(self, model: Literal['gpt-4o', 'gpt-4o-mini']) -> Tuple[ChatAgent, ChatAgent]:
r"""Initialize the agent."""
if model == 'gpt-4o':
web_agent_model = ModelFactory.create(
model_platform=ModelPlatformType.OPENAI,
model_type=ModelType.GPT_4O,
model_config_dict={"temperature": 0, "top_p": 1}
)
elif model == 'gpt-4o-mini':
web_agent_model = ModelFactory.create(
model_platform=ModelPlatformType.OPENAI,
model_type=ModelType.GPT_4O_MINI,
model_config_dict={"temperature": 0, "top_p": 1}
)
else:
raise ValueError("Invalid model type.")
planning_model = ModelFactory.create(
model_platform=ModelPlatformType.OPENAI,
model_type=ModelType.O3_MINI,
)
system_prompt = """
You are a helpful web agent that can assist users in browsing the web.
Given a high-level task, you can leverage predefined browser tools to help users achieve their goals.
"""
web_agent = ChatAgent(
system_message=system_prompt,
model=web_agent_model,
)
planning_system_prompt = """
You are a helpful planning agent that can assist users in planning complex tasks which need multi-step browser interaction.
"""
planning_agent = ChatAgent(
system_message=planning_system_prompt,
model=planning_model
)
return web_agent, planning_agent
def _observe(self, task_prompt: str, detailed_plan: Optional[str] = None) -> Tuple[str, str, str]:
r"""Let agent observe the current environment, and get the next action."""
detailed_plan_prompt = ""
if detailed_plan is not None:
detailed_plan_prompt = f"""
Here is a plan about how to solve the task step-by-step which you must follow: <detailed_plan>{detailed_plan}<detailed_plan>
"""
observe_prompt = f"""
Please act as a web agent to help me complete the following high-level task: <task>{task_prompt}</task>
Now, I have made screenshot (only the current viewport, not the full webpage) based on the current browser state, and marked interactive elements in the webpage.
Please carefully examine the requirements of the task, and current state of the browser, and provide the next appropriate action to take.
{detailed_plan_prompt}
Here are the current available browser functions you can use:
{AVAILABLE_ACTIONS_PROMPT}
Here are the latest {self.history_window} trajectory (at most) you have taken:
<history>
{self.history[-self.history_window:]}
</history>
Your output should be in json format, including the following fields:
- `observation`: The detailed image description about the current viewport. Do not over-confident about the correctness of the history actions. You should always check the current viewport to make sure the correctness of the next action.
- `reasoning`: The reasoning about the next action you want to take, and the possible obstacles you may encounter, and how to solve them. Do not forget to check the history actions to avoid the same mistakes.
- `action_code`: The action code you want to take. It is only one step action code, without any other texts (such as annotation)
Here are an example of the output:
```json
{{
"observation": [IMAGE_DESCRIPTION],
"reasoning": [YOUR_REASONING],
"action_code": `fill_input_id([ID], [TEXT])`
}}
Here are some tips for you:
- Never forget the overall question: **{task_prompt}**
- Maybe after a certain operation (e.g. click_id), the page content has not changed. You can check whether the action step is successful by looking at the `success` of the action step in the history. If successful, it means that the page content is indeed the same after the click. You need to try other methods.
- If using one way to solve the problem is not successful, try other ways. Make sure your provided ID is correct!
- Some cases are very complex and need to be achieve by an iterative process. You can use the `back()` function to go back to the previous page to try other methods.
- There are many links on the page, which may be useful for solving the problem. You can use the `click_id()` function to click on the link to see if it is useful.
- Always keep in mind that your action must be based on the ID shown in the current image or viewport, not the ID shown in the history.
- Do not use `stop()` lightly. Always remind yourself that the image only shows a part of the full page. If you cannot find the answer, try to use functions like `scroll_up()` and `scroll_down()` to check the full content of the webpage before doing anything else, because the answer or next key step may be hidden in the content below.
- If the webpage needs human verification, you must avoid processing it. Please use `back()` to go back to the previous page, and try other ways.
- If you have tried everything and still cannot resolve the issue, please stop the simulation, and report issues you have encountered.
- Check the history actions carefully, detect whether you have repeatedly made the same actions or not.
- When dealing with wikipedia revision history related tasks, you need to think about the solution flexibly. First, adjust the browsing history displayed on a single page to the maximum, and then make use of the find_text_on_page function. This is extremely useful which can quickly locate the text you want to find and skip massive amount of useless information.
- Flexibly use interactive elements like slide down selection bar to filter out the information you need. Sometimes they are extremely useful.
```
"""
# get current state
som_screenshot, som_screenshot_path = self.browser.get_som_screenshot(save_image=True)
img = _reload_image(som_screenshot)
message = BaseMessage.make_user_message(
role_name='user',
content=observe_prompt,
image_list=[img]
)
resp = self.web_agent.step(message)
resp_content = resp.msgs[0].content
resp_dict = _parse_json_output(resp_content)
observation_result: str = resp_dict.get("observation", "")
reasoning_result: str = resp_dict.get("reasoning", "")
action_code: str = resp_dict.get("action_code", "")
action_code = action_code.replace("`", "").strip()
return observation_result, reasoning_result, action_code
def _act(self, action_code: str) -> Tuple[bool, str]:
r"""Let agent act based on the given action code.
Args:
action_code (str): The action code to act.
Returns:
Tuple[bool, str]: A tuple containing a boolean indicating whether the action was successful, and the information to be returned.
"""
def _check_if_with_feedback(action_code: str) -> bool:
r"""Check if the action code needs feedback."""
for action_with_feedback in ACTION_WITH_FEEDBACK_LIST:
if action_with_feedback in action_code:
return True
return False
prefix = "self.browser."
code = f"{prefix}{action_code}"
try:
if _check_if_with_feedback(action_code):
# execute code, and get the executed result
result = eval(code)
time.sleep(1)
return True, result
else:
exec(code)
time.sleep(1)
return True, "Action was successful."
except Exception as e:
time.sleep(1)
return False, f"Error while executing the action {action_code}: {e}. If timeout, please recheck whether you have provided the correct identifier."
def _get_final_answer(self, task_prompt: str) -> str:
r"""Get the final answer based on the task prompt and current browser state.
It is used when the agent thinks that the task can be completed without any further action, and answer can be directly found in the current viewport.
"""
# screenshot, _ = self.browser.get_screenshot()
# img = _reload_image(screenshot)
prompt = f"""
We are solving a complex web task which needs multi-step browser interaction. After the multi-step observation, reasoning and acting with web browser, we think that the task is currently solved.
Here are all trajectory we have taken:
<history>{self.history}</history>
Please find the final answer, or give valuable insights and founds (e.g. if previous actions contain downloading files, your output should include the path of the downloaded file) about the overall task: <task>{task_prompt}</task>
"""
message = BaseMessage.make_user_message(
role_name='user',
content=prompt,
)
resp = self.web_agent.step(message)
return resp.msgs[0].content
def _make_reflection(self, task_prompt: str) -> str:
r"""Make a reflection about the current state and the task prompt."""
reflection_prompt = f"""
Now we are working on a complex task that requires multi-step browser interaction. The task is: <task>{task_prompt}</task>
To achieve this goal, we have made a series of observations, reasonings, and actions. We have also made a reflection on previous states.
Here are the global available browser functions we can use:
{AVAILABLE_ACTIONS_PROMPT}
Here are the latest {self.history_window} trajectory (at most) we have taken:
<history>{self.history[-self.history_window:]}</history>
The image provided is the current state of the browser, where we have marked interactive elements.
Please carefully examine the requirements of the task, and the current state of the browser, and then make reflections on the previous steps, thinking about whether they are helpful or not, and why, offering detailed feedback and suggestions for the next steps.
Your output should be in json format, including the following fields:
- `reflection`: The reflection about the previous steps, thinking about whether they are helpful or not, and why, offering detailed feedback.
- `suggestion`: The suggestion for the next steps, offering detailed suggestions, including the common solutions to the overall task based on the current state of the browser.
"""
som_image, _ = self.browser.get_som_screenshot()
img = _reload_image(som_image)
message = BaseMessage.make_user_message(
role_name='user',
content=reflection_prompt,
image_list=[img]
)
resp = self.web_agent.step(message)
return resp.msgs[0].content
def _task_planning(self, task_prompt: str, start_url: str) -> str:
r"""Plan the task based on the given task prompt."""
# Here are the available browser functions we can use: {AVAILABLE_ACTIONS_PROMPT}
planning_prompt = f"""
<task>{task_prompt}</task>
According to the problem above, if we use browser interaction, what is the general process of the interaction after visiting the webpage `{start_url}`?
Please note that it can be viewed as Partially Observable MDP. Do not over-confident about your plan.
Please first restate the task in detail, and then provide a detailed plan to solve the task.
"""
# Here are some tips for you: Please note that we can only see a part of the full page because of the limited viewport after an action. Thus, do not forget to use methods like `scroll_up()` and `scroll_down()` to check the full content of the webpage, because the answer or next key step may be hidden in the content below.
message = BaseMessage.make_user_message(
role_name='user',
content=planning_prompt
)
resp = self.planning_agent.step(message)
return resp.msgs[0].content
def _task_replanning(self, task_prompt: str, detailed_plan: str) -> Tuple[bool, str]:
r"""Replan the task based on the given task prompt.
Args:
task_prompt (str): The original task prompt.
detailed_plan (str): The detailed plan to replan.
Returns:
Tuple[bool, str]: A tuple containing a boolean indicating whether the task needs to be replanned, and the replanned schema.
"""
# Here are the available browser functions we can use: {AVAILABLE_ACTIONS_PROMPT}
replanning_prompt = f"""
We are using browser interaction to solve a complex task which needs multi-step actions.
Here are the overall task:
<overall_task>{task_prompt}</overall_task>
In order to solve the task, we made a detailed plan previously. Here is the detailed plan:
<detailed plan>{detailed_plan}</detailed plan>
According to the task above, we have made a series of observations, reasonings, and actions. Here are the latest {self.history_window} trajectory (at most) we have taken:
<history>{self.history[-self.history_window:]}</history>
However, the task is not completed yet. As the task is partially observable, we may need to replan the task based on the current state of the browser if necessary.
Now please carefully examine the current task planning schema, and our history actions, and then judge whether the task needs to be fundamentally replanned. If so, please provide a detailed replanned schema (including the restated overall task).
Your output should be in json format, including the following fields:
- `if_need_replan`: bool, A boolean value indicating whether the task needs to be fundamentally replanned.
- `replanned_schema`: str, The replanned schema for the task, which should not be changed too much compared with the original one. If the task does not need to be replanned, the value should be an empty string.
"""
resp = self.planning_agent.step(replanning_prompt)
resp_dict = _parse_json_output(resp.msgs[0].content)
if_need_replan = resp_dict.get("if_need_replan", False)
replanned_schema = resp_dict.get("replanned_schema", "")
if if_need_replan:
return True, replanned_schema
else:
return False, replanned_schema
def browser_simulation(self,
task_prompt: str,
start_url: str,
) -> str:
r"""A powerful toolkit which can simulate the browser interaction to solve the task which needs multi-step actions.
Args:
task_prompt (str): The task prompt to solve.
start_url (str): The start URL to visit.
Returns:
str: The simulation result to the task.
"""
ROUND_LIMIT = 12
self._reset()
task_completed = False
detailed_plan = self._task_planning(task_prompt, start_url)
logger.debug(f"Detailed plan: {detailed_plan}")
self.browser.visit_page(start_url)
for i in range(ROUND_LIMIT):
observation, reasoning, action_code = self._observe(task_prompt, detailed_plan)
logger.debug(f"Observation: {observation}")
logger.debug(f"Reasoning: {reasoning}")
logger.debug(f"Action code: {action_code}")
# breakpoint()
if "stop" in action_code:
task_completed = True
trajectory_info = {
"round": i,
"observation": observation,
"thought": reasoning,
"action": action_code,
"action_if_success": True,
"info": None,
"current_url": self.browser.get_url()
}
self.history.append(trajectory_info)
break
else:
success, info = self._act(action_code)
if not success:
logger.warning(f"Error while executing the action: {info}")
trajectory_info = {
"round": i,
"observation": observation,
"thought": reasoning,
"action": action_code,
"action_if_success": success,
"info": info,
"current_url": self.browser.get_url()
}
self.history.append(trajectory_info)
# replan the task if necessary
if_need_replan, replanned_schema = self._task_replanning(task_prompt, detailed_plan)
if if_need_replan:
detailed_plan = replanned_schema
logger.debug(f"Replanned schema: {replanned_schema}")
if not task_completed:
simulation_result = f"""
The task is not completed within the round limit. Please check the last round {self.history_window} information to see if there is any useful information:
<history>{self.history[-self.history_window:]}</history>
"""
else:
simulation_result = self._get_final_answer(task_prompt)
return simulation_result
def get_tools(self) -> List[FunctionTool]:
return [FunctionTool(self.browser_simulation)]