ReMe/reme_ai/summary/task/simple_comparative_summary_op.py

111 lines
4.5 KiB
Python

"""Simple comparative summary operation for task memory generation.
This module provides a simplified operation to extract task memories by
comparing trajectories with different scores for the same task.
"""
import json
from typing import List, Dict
from flowllm.core.context import C
from flowllm.core.enumeration import Role
from flowllm.core.op import BaseAsyncOp
from flowllm.core.schema import Message as FlowMessage
from loguru import logger
from reme_ai.schema import Message, Trajectory
from reme_ai.schema.memory import BaseMemory, TaskMemory
from reme_ai.utils.op_utils import merge_messages_content
@C.register_op()
class SimpleComparativeSummaryOp(BaseAsyncOp):
"""Extract task memories by comparing trajectories with different scores.
This operation compares the highest and lowest scoring trajectories for
each task to extract comparative insights and best practices.
"""
file_path: str = __file__
async def compare_summary_trajectory(self, trajectory_a: Trajectory, trajectory_b: Trajectory) -> List[BaseMemory]:
"""Compare two trajectories and extract comparative task memories.
Args:
trajectory_a: First trajectory to compare (typically higher scoring)
trajectory_b: Second trajectory to compare (typically lower scoring)
Returns:
List of extracted task memories from the comparison
"""
summary_prompt = self.prompt_format(
prompt_name="summary_prompt",
execution_process_a=merge_messages_content(trajectory_a.messages),
execution_process_b=merge_messages_content(trajectory_b.messages),
summary_example=self.get_prompt("summary_example"),
)
def parse_content(message: Message):
content = message.content
task_memory_list = []
try:
content = content.split("```")[1].strip()
if content.startswith("json"):
content = content.strip("json")
for tm_dict in json.loads(content):
when_to_use = tm_dict.get("when_to_use", "").strip()
task_memory_content = tm_dict.get("experience", "").strip()
if when_to_use and task_memory_content:
task_memory_list.append(
TaskMemory(
workspace_id=self.context.get("workspace_id", ""),
when_to_use=when_to_use,
content=task_memory_content,
author=getattr(self.llm, "model_name", "system"),
),
)
return task_memory_list
except Exception as e:
logger.exception(f"parse content failed!\n{content}")
raise e
return await self.llm.achat(
messages=[FlowMessage(role=Role.USER, content=summary_prompt)],
callback_fn=parse_content,
)
async def async_execute(self):
"""Execute the comparative summary operation.
Groups trajectories by task_id, compares the highest and lowest scoring
trajectories for each task, and extracts task memories from the comparison.
"""
trajectories: list = self.context.get("trajectories", [])
trajectories: List[Trajectory] = [Trajectory(**x) if isinstance(x, dict) else x for x in trajectories]
task_id_dict: Dict[str, List[Trajectory]] = {}
for trajectory in trajectories:
if trajectory.task_id not in task_id_dict:
task_id_dict[trajectory.task_id] = []
task_id_dict[trajectory.task_id].append(trajectory)
memory_list = []
for _, task_trajectories in task_id_dict.items():
task_trajectories: List[Trajectory] = sorted(task_trajectories, key=lambda x: x.score, reverse=True)
if len(task_trajectories) < 2:
continue
if task_trajectories[0].score > task_trajectories[-1].score:
task_memories = await self.compare_summary_trajectory(
trajectory_a=task_trajectories[0],
trajectory_b=task_trajectories[-1],
)
memory_list.extend(task_memories)
self.context.response.answer = json.dumps([x.model_dump() for x in memory_list])
self.context.response.metadata["memory_list"] = memory_list
for tm in memory_list:
logger.info(f"add task memory when_to_use={tm.when_to_use}\ncontent={tm.content}")