File size: 15,410 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 |
from __future__ import annotations
from abc import ABC, abstractmethod
from contextlib import contextmanager
from typing import Any
from litellm import supports_response_schema
from pydantic import BaseModel
from typing_extensions import override
from openhands.controller.state.state import State
from openhands.core.config.condenser_config import (
AmortizedForgettingCondenserConfig,
CondenserConfig,
LLMAttentionCondenserConfig,
LLMSummarizingCondenserConfig,
NoOpCondenserConfig,
ObservationMaskingCondenserConfig,
RecentEventsCondenserConfig,
)
from openhands.core.logger import openhands_logger as logger
from openhands.events.event import Event
from openhands.events.observation import AgentCondensationObservation, Observation
from openhands.llm.llm import LLM
CONDENSER_METADATA_KEY = 'condenser_meta'
"""Key identifying where metadata is stored in a `State` object's `extra_data` field."""
def get_condensation_metadata(state: State) -> list[dict[str, Any]]:
"""Utility function to retrieve a list of metadata batches from a `State`.
Args:
state: The state to retrieve metadata from.
Returns:
list[dict[str, Any]]: A list of metadata batches, each representing a condensation.
"""
if CONDENSER_METADATA_KEY in state.extra_data:
return state.extra_data[CONDENSER_METADATA_KEY]
return []
class Condenser(ABC):
"""Abstract condenser interface.
Condensers take a list of `Event` objects and reduce them into a potentially smaller list.
Agents can use condensers to reduce the amount of events they need to consider when deciding which action to take. To use a condenser, agents can call the `condensed_history` method on the current `State` being considered and use the results instead of the full history.
Example usage::
condenser = Condenser.from_config(condenser_config)
events = condenser.condensed_history(state)
"""
def __init__(self):
self._metadata_batch: dict[str, Any] = {}
def add_metadata(self, key: str, value: Any) -> None:
"""Add information to the current metadata batch.
Any key/value pairs added to the metadata batch will be recorded in the `State` at the end of the current condensation.
Args:
key: The key to store the metadata under.
value: The metadata to store.
"""
self._metadata_batch[key] = value
def write_metadata(self, state: State) -> None:
"""Write the current batch of metadata to the `State`.
Resets the current metadata batch: any metadata added after this call will be stored in a new batch and written to the `State` at the end of the next condensation.
"""
if CONDENSER_METADATA_KEY not in state.extra_data:
state.extra_data[CONDENSER_METADATA_KEY] = []
if self._metadata_batch:
state.extra_data[CONDENSER_METADATA_KEY].append(self._metadata_batch)
# Since the batch has been written, clear it for the next condensation
self._metadata_batch = {}
@contextmanager
def metadata_batch(self, state: State):
"""Context manager to ensure batched metadata is always written to the `State`."""
try:
yield
finally:
self.write_metadata(state)
@abstractmethod
def condense(self, events: list[Event]) -> list[Event]:
"""Condense a sequence of events into a potentially smaller list.
New condenser strategies should override this method to implement their own condensation logic. Call `self.add_metadata` in the implementation to record any relevant per-condensation diagnostic information.
Args:
events: A list of events representing the entire history of the agent.
Returns:
list[Event]: An event sequence representing a condensed history of the agent.
"""
def condensed_history(self, state: State) -> list[Event]:
"""Condense the state's history."""
with self.metadata_batch(state):
return self.condense(state.history)
@classmethod
def from_config(cls, config: CondenserConfig) -> Condenser:
"""Create a condenser from a configuration object.
Args:
config: Configuration for the condenser.
Returns:
Condenser: A condenser instance.
Raises:
ValueError: If the condenser type is not recognized.
"""
match config:
case NoOpCondenserConfig():
return NoOpCondenser()
case ObservationMaskingCondenserConfig():
return ObservationMaskingCondenser(
**config.model_dump(exclude=['type'])
)
case RecentEventsCondenserConfig():
return RecentEventsCondenser(**config.model_dump(exclude=['type']))
case LLMSummarizingCondenserConfig(llm_config=llm_config):
return LLMSummarizingCondenser(llm=LLM(config=llm_config))
case AmortizedForgettingCondenserConfig():
return AmortizedForgettingCondenser(
**config.model_dump(exclude=['type'])
)
case LLMAttentionCondenserConfig(llm_config=llm_config):
return LLMAttentionCondenser(
llm=LLM(config=llm_config),
**config.model_dump(exclude=['type', 'llm_config']),
)
case _:
raise ValueError(f'Unknown condenser config: {config}')
class RollingCondenser(Condenser, ABC):
"""Base class for a specialized condenser strategy that applies condensation to a rolling history.
The rolling history is computed by appending new events to the most recent condensation. For example, the sequence of calls::
assert state.history == [event1, event2, event3]
condensation = condenser.condensed_history(state)
# ...new events are added to the state...
assert state.history == [event1, event2, event3, event4, event5]
condenser.condensed_history(state)
will result in second call to `condensed_history` passing `condensation + [event4, event5]` to the `condense` method.
"""
def __init__(self) -> None:
self._condensation: list[Event] = []
self._last_history_length: int = 0
super().__init__()
@override
def condensed_history(self, state: State) -> list[Event]:
new_events = state.history[self._last_history_length :]
with self.metadata_batch(state):
results = self.condense(self._condensation + new_events)
self._condensation = results
self._last_history_length = len(state.history)
return results
class NoOpCondenser(Condenser):
"""A condenser that does nothing to the event sequence."""
def condense(self, events: list[Event]) -> list[Event]:
"""Returns the list of events unchanged."""
return events
class ObservationMaskingCondenser(Condenser):
"""A condenser that masks the values of observations outside of a recent attention window."""
def __init__(self, attention_window: int = 5):
self.attention_window = attention_window
super().__init__()
def condense(self, events: list[Event]) -> list[Event]:
"""Replace the content of observations outside of the attention window with a placeholder."""
results: list[Event] = []
for i, event in enumerate(events):
if (
isinstance(event, Observation)
and i < len(events) - self.attention_window
):
results.append(AgentCondensationObservation('<MASKED>'))
else:
results.append(event)
return results
class RecentEventsCondenser(Condenser):
"""A condenser that only keeps a certain number of the most recent events."""
def __init__(self, keep_first: int = 0, max_events: int = 10):
self.keep_first = keep_first
self.max_events = max_events
super().__init__()
def condense(self, events: list[Event]) -> list[Event]:
"""Keep only the most recent events (up to `max_events`)."""
head = events[: self.keep_first]
tail_length = max(0, self.max_events - len(head))
tail = events[-tail_length:]
return head + tail
class LLMSummarizingCondenser(Condenser):
"""A condenser that relies on a language model to summarize the event sequence as a single event."""
def __init__(self, llm: LLM):
self.llm = llm
super().__init__()
def condense(self, events: list[Event]) -> list[Event]:
"""Applies an LLM to summarize the list of events.
Raises:
Exception: If the LLM is unable to summarize the event sequence.
"""
try:
# Convert events to a format suitable for summarization
events_text = '\n'.join(f'{e.timestamp}: {e.message}' for e in events)
summarize_prompt = f'Please summarize these events:\n{events_text}'
resp = self.llm.completion(
messages=[{'content': summarize_prompt, 'role': 'user'}]
)
summary_response = resp.choices[0].message.content
# Create a new summary event with the condensed content
summary_event = AgentCondensationObservation(summary_response)
# Add metrics to state
self.add_metadata('response', resp.model_dump())
self.add_metadata('metrics', self.llm.metrics.get())
return [summary_event]
except Exception as e:
logger.error(f'Error condensing events: {str(e)}')
raise e
class AmortizedForgettingCondenser(RollingCondenser):
"""A condenser that maintains a condensed history and forgets old events when it grows too large."""
def __init__(self, max_size: int = 100, keep_first: int = 0):
"""Initialize the condenser.
Args:
max_size: Maximum size of history before forgetting.
keep_first: Number of initial events to always keep.
Raises:
ValueError: If keep_first is greater than max_size, keep_first is negative, or max_size is non-positive.
"""
if keep_first >= max_size // 2:
raise ValueError(
f'keep_first ({keep_first}) must be less than half of max_size ({max_size})'
)
if keep_first < 0:
raise ValueError(f'keep_first ({keep_first}) cannot be negative')
if max_size < 1:
raise ValueError(f'max_size ({keep_first}) cannot be non-positive')
self.max_size = max_size
self.keep_first = keep_first
super().__init__()
def condense(self, events: list[Event]) -> list[Event]:
"""Apply the amortized forgetting strategy to the given list of events."""
if len(events) <= self.max_size:
return events
target_size = self.max_size // 2
head = events[: self.keep_first]
events_from_tail = target_size - len(head)
tail = events[-events_from_tail:]
return head + tail
class ImportantEventSelection(BaseModel):
"""Utility class for the `LLMAttentionCondenser` that forces the LLM to return a list of integers."""
ids: list[int]
class LLMAttentionCondenser(RollingCondenser):
"""Rolling condenser strategy that uses an LLM to select the most important events when condensing the history."""
def __init__(self, llm: LLM, max_size: int = 100, keep_first: int = 0):
if keep_first >= max_size // 2:
raise ValueError(
f'keep_first ({keep_first}) must be less than half of max_size ({max_size})'
)
if keep_first < 0:
raise ValueError(f'keep_first ({keep_first}) cannot be negative')
if max_size < 1:
raise ValueError(f'max_size ({keep_first}) cannot be non-positive')
self.max_size = max_size
self.keep_first = keep_first
self.llm = llm
# This condenser relies on the `response_schema` feature, which is not supported by all LLMs
if not supports_response_schema(
model=self.llm.config.model,
custom_llm_provider=self.llm.config.custom_llm_provider,
):
raise ValueError(
"The LLM model must support the 'response_schema' parameter to use the LLMAttentionCondenser."
)
super().__init__()
def condense(self, events: list[Event]) -> list[Event]:
"""If the history is too long, use an LLM to select the most important events."""
if len(events) <= self.max_size:
return events
target_size = self.max_size // 2
head = events[: self.keep_first]
events_from_tail = target_size - len(head)
message: str = """You will be given a list of actions, observations, and thoughts from a coding agent.
Each item in the list has an identifier. Please sort the identifiers in order of how important the
contents of the item are for the next step of the coding agent's task, from most important to least
important."""
response = self.llm.completion(
messages=[
{'content': message, 'role': 'user'},
*[
{
'content': f'<ID>{e.id}</ID>\n<CONTENT>{e.message}</CONTENT>',
'role': 'user',
}
for e in events
],
],
response_format={
'type': 'json_schema',
'json_schema': {
'name': 'ImportantEventSelection',
'schema': ImportantEventSelection.model_json_schema(),
},
},
)
response_ids = ImportantEventSelection.model_validate_json(
response.choices[0].message.content
).ids
self.add_metadata('all_event_ids', [event.id for event in events])
self.add_metadata('response_ids', response_ids)
self.add_metadata('metrics', self.llm.metrics.get())
# Filter out any IDs from the head and trim the results down
head_ids = [event.id for event in head]
response_ids = [
response_id for response_id in response_ids if response_id not in head_ids
][:events_from_tail]
# If the response IDs aren't _long_ enough, iterate backwards through the events and add any unfound IDs to the list.
for event in reversed(events):
if len(response_ids) >= events_from_tail:
break
if event.id not in response_ids:
response_ids.append(event.id)
# Grab the events associated with the response IDs
tail = [event for event in events if event.id in response_ids]
return head + tail
|