Skyvern/skyvern/forge/agent_functions.py

429 lines
16 KiB
Python

import asyncio
import copy
import hashlib
from datetime import timedelta
from typing import Dict, List
import structlog
from playwright.async_api import Frame, Page
from skyvern.config import settings
from skyvern.constants import SKYVERN_ID_ATTR
from skyvern.exceptions import StepUnableToExecuteError
from skyvern.forge import app
from skyvern.forge.async_operations import AsyncOperation
from skyvern.forge.prompts import prompt_engine
from skyvern.forge.sdk.api.llm.exceptions import LLMProviderError
from skyvern.forge.sdk.models import Step, StepStatus
from skyvern.forge.sdk.schemas.organizations import Organization
from skyvern.forge.sdk.schemas.tasks import Task, TaskStatus
from skyvern.webeye.browser_factory import BrowserState
from skyvern.webeye.scraper.scraper import ELEMENT_NODE_ATTRIBUTES, CleanupElementTreeFunc, json_to_html
LOG = structlog.get_logger()
USELESS_SHAPE_ATTRIBUTE = [SKYVERN_ID_ATTR, "id", "aria-describedby"]
SVG_SHAPE_CONVERTION_ATTEMPTS = 3
CSS_SHAPE_CONVERTION_ATTEMPTS = 1
INVALID_SHAPE = "N/A"
def _remove_rect(element: dict) -> None:
if "rect" in element:
del element["rect"]
def _should_css_shape_convert(element: Dict) -> bool:
if "id" not in element:
return False
tag_name = element.get("tagName")
if tag_name not in ["a", "span", "i"]:
return False
# should be without children
if len(element.get("children", [])) > 0:
return False
# should be no text
if element.get("text"):
return False
# if <span> and <i> we try to convert the shape
if tag_name in ["span", "i"]:
return True
# if <a>, it should be no text, no href/target attribute
if tag_name == "a":
attributes = element.get("attributes", {})
if "href" in attributes:
return False
if "target" in attributes:
return False
return True
return False
def _get_svg_cache_key(hash: str) -> str:
return f"skyvern:svg:{hash}"
def _get_shape_cache_key(hash: str) -> str:
return f"skyvern:shape:{hash}"
def _remove_skyvern_attributes(element: Dict) -> Dict:
"""
To get the original HTML element without skyvern attributes
"""
element_copied = copy.deepcopy(element)
for attr in ELEMENT_NODE_ATTRIBUTES:
if element_copied.get(attr):
del element_copied[attr]
if "attributes" in element_copied:
attributes: dict = copy.deepcopy(element_copied.get("attributes", {}))
for key in attributes.keys():
if key in USELESS_SHAPE_ATTRIBUTE:
del element_copied["attributes"][key]
children: List[Dict] | None = element_copied.get("children", None)
if children is None:
return element_copied
trimmed_children = []
for child in children:
trimmed_children.append(_remove_skyvern_attributes(child))
element_copied["children"] = trimmed_children
return element_copied
async def _convert_svg_to_string(
element: Dict,
task: Task | None = None,
step: Step | None = None,
) -> None:
if element.get("tagName") != "svg":
return
if element.get("isDropped", False):
return
task_id = task.task_id if task else None
step_id = step.step_id if step else None
element_id = element.get("id", "")
svg_element = _remove_skyvern_attributes(element)
svg_html = json_to_html(svg_element)
hash_object = hashlib.sha256()
hash_object.update(svg_html.encode("utf-8"))
svg_hash = hash_object.hexdigest()
svg_key = _get_svg_cache_key(svg_hash)
svg_shape: str | None = None
try:
svg_shape = await app.CACHE.get(svg_key)
except Exception:
LOG.warning(
"Failed to loaded SVG cache",
task_id=task_id,
step_id=step_id,
exc_info=True,
key=svg_key,
)
if svg_shape:
LOG.debug("SVG loaded from cache", element_id=element_id, key=svg_key, shape=svg_shape)
else:
if len(svg_html) > settings.SVG_MAX_LENGTH:
# TODO: implement a fallback solution for "too large" case, maybe convert by screenshot
LOG.warning(
"SVG element is too large to convert, going to drop the svg element.",
element_id=element_id,
task_id=task_id,
step_id=step_id,
length=len(svg_html),
key=svg_key,
)
del element["children"]
element["isDropped"] = True
return
LOG.debug("call LLM to convert SVG to string shape", element_id=element_id)
svg_convert_prompt = prompt_engine.load_prompt("svg-convert", svg_element=svg_html)
for retry in range(SVG_SHAPE_CONVERTION_ATTEMPTS):
try:
json_response = await app.SECONDARY_LLM_API_HANDLER(prompt=svg_convert_prompt, step=step)
svg_shape = json_response.get("shape", "")
recognized = json_response.get("recognized", False)
if not svg_shape or not recognized:
raise Exception("Empty or unrecognized SVG shape replied by secondary llm")
LOG.info("SVG converted by LLM", element_id=element_id, key=svg_key, shape=svg_shape)
await app.CACHE.set(svg_key, svg_shape)
break
except LLMProviderError:
LOG.info(
"Failed to convert SVG to string due to llm error. Will retry if haven't met the max try attempt after 3s.",
exc_info=True,
task_id=task_id,
step_id=step_id,
element_id=element_id,
key=svg_key,
retry=retry,
)
if retry == SVG_SHAPE_CONVERTION_ATTEMPTS - 1:
# set the invalid css shape to cache to avoid retry in the near future
await app.CACHE.set(svg_key, INVALID_SHAPE, ex=timedelta(hours=1))
await asyncio.sleep(3)
except Exception:
LOG.info(
"Failed to convert SVG to string shape by secondary llm. Will retry if haven't met the max try attempt after 3s.",
exc_info=True,
task_id=task_id,
step_id=step_id,
element_id=element_id,
retry=retry,
)
if retry == SVG_SHAPE_CONVERTION_ATTEMPTS - 1:
# set the invalid css shape to cache to avoid retry in the near future
await app.CACHE.set(svg_key, INVALID_SHAPE, ex=timedelta(weeks=1))
await asyncio.sleep(3)
else:
LOG.warning(
"Reaching the max try to convert svg element, going to drop the svg element.",
element_id=element_id,
task_id=task_id,
step_id=step_id,
key=svg_key,
length=len(svg_html),
)
del element["children"]
element["isDropped"] = True
return
element["attributes"] = dict()
if svg_shape != INVALID_SHAPE:
# refresh the cache expiration
await app.CACHE.set(svg_key, svg_shape)
element["attributes"]["alt"] = svg_shape
del element["children"]
return
async def _convert_css_shape_to_string(
frame: Page | Frame,
element: Dict,
task: Task | None = None,
step: Step | None = None,
) -> None:
element_id: str = element.get("id", "")
task_id = task.task_id if task else None
step_id = step.step_id if step else None
shape_element = _remove_skyvern_attributes(element)
svg_html = json_to_html(shape_element)
hash_object = hashlib.sha256()
hash_object.update(svg_html.encode("utf-8"))
shape_hash = hash_object.hexdigest()
shape_key = _get_shape_cache_key(shape_hash)
css_shape: str | None = None
try:
css_shape = await app.CACHE.get(shape_key)
except Exception:
LOG.warning(
"Failed to loaded CSS shape cache",
task_id=task_id,
step_id=step_id,
exc_info=True,
key=shape_key,
)
if css_shape:
LOG.debug("CSS shape loaded from cache", element_id=element_id, key=shape_key, shape=css_shape)
else:
# FIXME: support element in iframe
locater = frame.locator(f'[{SKYVERN_ID_ATTR}="{element_id}"]')
if await locater.count() == 0:
LOG.info(
"No locater found to convert css shape",
task_id=task_id,
step_id=step_id,
element_id=element_id,
)
return None
if await locater.count() > 1:
LOG.info(
"multiple locaters found to convert css shape",
task_id=task_id,
step_id=step_id,
element_id=element_id,
)
return None
try:
LOG.debug("call LLM to convert css shape to string shape", element_id=element_id)
screenshot = await locater.screenshot(timeout=settings.BROWSER_SCREENSHOT_TIMEOUT_MS)
prompt = prompt_engine.load_prompt("css-shape-convert")
# TODO: we don't retry the css shape conversion today
for retry in range(CSS_SHAPE_CONVERTION_ATTEMPTS):
try:
json_response = await app.SECONDARY_LLM_API_HANDLER(
prompt=prompt, screenshots=[screenshot], step=step
)
css_shape = json_response.get("shape", "")
recognized = json_response.get("recognized", False)
if not css_shape or not recognized:
raise Exception("Empty or unrecognized css shape replied by secondary llm")
LOG.info("CSS Shape converted by LLM", element_id=element_id, key=shape_key, shape=css_shape)
await app.CACHE.set(shape_key, css_shape)
break
except LLMProviderError:
LOG.info(
"Failed to convert css shape due to llm error. Will retry if haven't met the max try attempt after 3s.",
exc_info=True,
task_id=task_id,
step_id=step_id,
element_id=element_id,
retry=retry,
key=shape_key,
)
if retry == CSS_SHAPE_CONVERTION_ATTEMPTS - 1:
# set the invalid css shape to cache to avoid retry in the near future
await app.CACHE.set(shape_key, INVALID_SHAPE, ex=timedelta(hours=1))
await asyncio.sleep(3)
except Exception:
LOG.info(
"Failed to convert css shape to string shape by secondary llm. Will retry if haven't met the max try attempt after 3s.",
exc_info=True,
task_id=task_id,
step_id=step_id,
element_id=element_id,
retry=retry,
key=shape_key,
)
if retry == CSS_SHAPE_CONVERTION_ATTEMPTS - 1:
# set the invalid css shape to cache to avoid retry in the near future
await app.CACHE.set(shape_key, INVALID_SHAPE, ex=timedelta(weeks=1))
await asyncio.sleep(3)
else:
LOG.info(
"Max css shape convertion retry, going to abort the convertion.",
task_id=task_id,
step_id=step_id,
element_id=element_id,
key=shape_key,
)
return None
except Exception:
LOG.warning(
"Failed to convert css shape to string shape by LLM",
key=shape_key,
task_id=task_id,
step_id=step_id,
element_id=element_id,
exc_info=True,
)
return None
if "attributes" not in element:
element["attributes"] = dict()
if css_shape != INVALID_SHAPE:
# refresh the cache expiration
await app.CACHE.set(shape_key, css_shape)
element["attributes"]["shape-description"] = css_shape
return None
class AgentFunction:
async def validate_step_execution(
self,
task: Task,
step: Step,
) -> None:
"""
Checks if the step can be executed. It is called before the step is executed.
:return: A tuple of whether the step can be executed and a list of reasons why it can't be executed.
"""
reasons = []
# can't execute if task status is not running
has_valid_task_status = task.status == TaskStatus.running
if not has_valid_task_status:
reasons.append(f"invalid_task_status:{task.status}")
# can't execute if the step is already running or completed
has_valid_step_status = step.status in [StepStatus.created, StepStatus.failed]
if not has_valid_step_status:
reasons.append(f"invalid_step_status:{step.status}")
# can't execute if the task has another step that is running
steps = await app.DATABASE.get_task_steps(task_id=task.task_id, organization_id=task.organization_id)
has_no_running_steps = not any(step.status == StepStatus.running for step in steps)
if not has_no_running_steps:
reasons.append(f"another_step_is_running_for_task:{task.task_id}")
can_execute = has_valid_task_status and has_valid_step_status and has_no_running_steps
if not can_execute:
raise StepUnableToExecuteError(step_id=step.step_id, reason=f"Cannot execute step. Reasons: {reasons}")
async def prepare_step_execution(
self,
organization: Organization | None,
task: Task,
step: Step,
browser_state: BrowserState,
) -> None:
"""
Get prepared for the step execution. It's called at the first beginning when step running.
"""
return
async def generate_async_operations(
self,
organization: Organization,
task: Task,
page: Page,
) -> list[AsyncOperation]:
return []
def cleanup_element_tree_factory(
self,
task: Task | None = None,
step: Step | None = None,
) -> CleanupElementTreeFunc:
async def cleanup_element_tree_func(frame: Page | Frame, url: str, element_tree: list[dict]) -> list[dict]:
"""
Remove rect and attribute.unique_id from the elements.
The reason we're doing it is to
1. reduce unnecessary data so that llm get less distrction
TODO later: 2. reduce tokens sent to llm to save money
:param elements: List of elements to remove xpaths from.
:return: List of elements without xpaths.
"""
queue = []
for element in element_tree:
queue.append(element)
while queue:
queue_ele = queue.pop(0)
_remove_rect(queue_ele)
await _convert_svg_to_string(queue_ele, task, step)
if _should_css_shape_convert(element=queue_ele):
await _convert_css_shape_to_string(
frame=frame,
element=queue_ele,
task=task,
step=step,
)
# TODO: we can come back to test removing the unique_id
# from element attributes to make sure this won't increase hallucination
# _remove_unique_id(queue_ele)
if "children" in queue_ele:
queue.extend(queue_ele["children"])
return element_tree
return cleanup_element_tree_func