File size: 4,967 Bytes
246d201 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 |
from dataclasses import asdict
from datetime import datetime
from pydantic import BaseModel
from openhands.events import Event, EventSource
from openhands.events.observation.observation import Observation
from openhands.events.serialization.action import action_from_dict
from openhands.events.serialization.observation import observation_from_dict
from openhands.events.serialization.utils import remove_fields
from openhands.events.tool import ToolCallMetadata
# TODO: move `content` into `extras`
TOP_KEYS = [
'id',
'timestamp',
'source',
'message',
'cause',
'action',
'observation',
'tool_call_metadata',
]
UNDERSCORE_KEYS = ['id', 'timestamp', 'source', 'cause', 'tool_call_metadata']
DELETE_FROM_TRAJECTORY_EXTRAS = {
'screenshot',
'dom_object',
'axtree_object',
'active_page_index',
'last_browser_action',
'last_browser_action_error',
'focused_element_bid',
'extra_element_properties',
}
DELETE_FROM_MEMORY_EXTRAS = DELETE_FROM_TRAJECTORY_EXTRAS | {'open_pages_urls'}
def event_from_dict(data) -> 'Event':
evt: Event
if 'action' in data:
evt = action_from_dict(data)
elif 'observation' in data:
evt = observation_from_dict(data)
else:
raise ValueError('Unknown event type: ' + data)
for key in UNDERSCORE_KEYS:
if key in data:
value = data[key]
if key == 'timestamp' and isinstance(value, datetime):
value = value.isoformat()
if key == 'source':
value = EventSource(value)
if key == 'tool_call_metadata':
value = ToolCallMetadata(**value)
setattr(evt, '_' + key, value)
return evt
def _convert_pydantic_to_dict(obj: BaseModel | dict) -> dict:
if isinstance(obj, BaseModel):
return obj.model_dump()
return obj
def event_to_dict(event: 'Event') -> dict:
props = asdict(event)
d = {}
for key in TOP_KEYS:
if hasattr(event, key) and getattr(event, key) is not None:
d[key] = getattr(event, key)
elif hasattr(event, f'_{key}') and getattr(event, f'_{key}') is not None:
d[key] = getattr(event, f'_{key}')
if key == 'id' and d.get('id') == -1:
d.pop('id', None)
if key == 'timestamp' and 'timestamp' in d:
if isinstance(d['timestamp'], datetime):
d['timestamp'] = d['timestamp'].isoformat()
if key == 'source' and 'source' in d:
d['source'] = d['source'].value
if key == 'tool_call_metadata' and 'tool_call_metadata' in d:
d['tool_call_metadata'] = d['tool_call_metadata'].model_dump()
props.pop(key, None)
if 'security_risk' in props and props['security_risk'] is None:
props.pop('security_risk')
if 'action' in d:
d['args'] = props
if event.timeout is not None:
d['timeout'] = event.timeout
elif 'observation' in d:
d['content'] = props.pop('content', '')
# props is a dict whose values can include a complex object like an instance of a BaseModel subclass
# such as CmdOutputMetadata
# we serialize it along with the rest
d['extras'] = {k: _convert_pydantic_to_dict(v) for k, v in props.items()}
# Include success field for CmdOutputObservation
if hasattr(event, 'success'):
d['success'] = event.success
else:
raise ValueError('Event must be either action or observation')
return d
def event_to_trajectory(event: 'Event') -> dict:
d = event_to_dict(event)
if 'extras' in d:
remove_fields(d['extras'], DELETE_FROM_TRAJECTORY_EXTRAS)
return d
def event_to_memory(event: 'Event', max_message_chars: int) -> dict:
d = event_to_dict(event)
d.pop('id', None)
d.pop('cause', None)
d.pop('timestamp', None)
d.pop('message', None)
d.pop('image_urls', None)
# runnable actions have some extra fields used in the BE/FE, which should not be sent to the LLM
if 'args' in d:
d['args'].pop('blocking', None)
d['args'].pop('confirmation_state', None)
if 'extras' in d:
remove_fields(d['extras'], DELETE_FROM_MEMORY_EXTRAS)
if isinstance(event, Observation) and 'content' in d:
d['content'] = truncate_content(d['content'], max_message_chars)
return d
def truncate_content(content: str, max_chars: int) -> str:
"""Truncate the middle of the observation content if it is too long."""
if len(content) <= max_chars or max_chars == -1:
return content
# truncate the middle and include a message to the LLM about it
half = max_chars // 2
return (
content[:half]
+ '\n[... Observation truncated due to length ...]\n'
+ content[-half:]
)
|