File size: 6,724 Bytes
3370983 |
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 |
"""
History Manager for conversation memory and compaction.
Handles persistent conversation state and implements "Compactive Summarization"
to prevent token overflow while preserving critical conversation history.
"""
import time
import random
import uuid
from datetime import datetime
from typing import List
from langchain_openai import ChatOpenAI
from langchain_core.messages import BaseMessage, SystemMessage, HumanMessage, AIMessage
from src.backend.prompts import get_prompt
class HistoryManager:
"""
Manages persistent conversation state and implements compaction logic.
Responsibilities:
1. Compaction: Summarizing old messages to save tokens.
2. Persistence: Safely updating the low-level checkpoint storage.
"""
def __init__(self, memory_saver):
self.memory = memory_saver
def _messages_to_text(self, messages: List[BaseMessage]) -> str:
"""Convert messages to a plain text transcript."""
text_parts = []
for msg in messages:
role = msg.__class__.__name__
content = msg.content
if isinstance(content, str):
text_parts.append(f"{role}: {content}")
else:
text_parts.append(f"{role}: {str(content)}")
return "\n\n".join(text_parts)
def _is_tool_message(self, msg: BaseMessage) -> bool:
"""Check if a message is a ToolMessage or Tool output."""
msg_type = getattr(msg, "type", None)
role = getattr(msg, "role", None)
return msg_type == "tool" or role == "tool" or msg.__class__.__name__ == "ToolMessage"
def compact_messages(self, messages: List[BaseMessage], compaction_ratio: float = 0.5) -> List[BaseMessage]:
"""
Apply "Compactive Summarization" to the conversation history.
Technique:
- Splits history into Old and Recent based on compaction_ratio.
- Summarizes Old messages into a single narrative block using an LLM.
- Preserves the System Prompt and Recent messages verbatim.
Args:
messages: Full list of conversation messages.
compaction_ratio: Fraction of messages to compact (0.0 to 1.0).
- 0.5 (Default): Summarizes 50% (Oldest half).
- 0.8: Aggressive. Summarizes 80% (Keeps only very recent messages).
- 0.2: Gentle. Summarizes only the oldest 20%.
Returns:
List[BaseMessage]: optimized list with summary replacing old history.
"""
if len(messages) < 2:
return messages
system_msg = None
conversation_msgs = messages
# Preserve system prompt
if isinstance(messages[0], SystemMessage):
system_msg = messages[0]
conversation_msgs = messages[1:]
if len(conversation_msgs) < 2:
return messages
# Calculate split point based on ratio
split_idx = int(len(conversation_msgs) * compaction_ratio)
# Ensure we compact at least something if ratio > 0, but keep at least one recent message
split_idx = max(1, min(split_idx, len(conversation_msgs) - 1))
first_half = conversation_msgs[:split_idx]
second_half = conversation_msgs[split_idx:]
# Ensure second_half does not start with orphaned tool message
while second_half and self._is_tool_message(second_half[0]):
if first_half:
second_half.insert(0, first_half.pop())
else:
second_half.pop(0)
# Generate summary
compactor_prompt = get_prompt(template_name="Compactor", latest_version=True)
conversation_text = self._messages_to_text(first_half)
llm = ChatOpenAI(model="gpt-4o-mini", temperature=0, max_tokens=1000)
messages_for_llm = [
SystemMessage(content=compactor_prompt),
HumanMessage(content=f"Conversation history to summarize:\n\n{conversation_text}")
]
response = llm.invoke(messages_for_llm)
summary_text = response.content
print(f"\n{'='*80}\n📝 COMPACTION MESSAGE:\n{summary_text}\n{'='*80}\n", flush=True)
summary_message = AIMessage(content=f"[COMPACTED SUMMARY OF EARLIER CONVERSATION]\n\n{summary_text}")
result = []
if system_msg:
result.append(system_msg)
result.append(summary_message)
result.extend(second_half)
return result
def replace_thread_history(self, thread_id: str, new_messages: List[BaseMessage]) -> bool:
"""
Atomically overwrite the message history in the checkpoint storage.
This bypasses the standard append-only reducer to force a history rewrite.
Crucial for finalizing the compaction process.
"""
config = {"configurable": {"thread_id": thread_id}}
current_checkpoint = self.memory.get_tuple(config)
if not current_checkpoint or not current_checkpoint.checkpoint:
return False
checkpoint_config = {
"configurable": {**current_checkpoint.config.get("configurable", {})}
}
checkpoint_config["configurable"].setdefault("thread_id", thread_id)
checkpoint_config["configurable"].setdefault("checkpoint_ns", "")
current_versions = current_checkpoint.checkpoint.get('channel_versions', {})
new_msg_version = f"{str(int(time.time())).zfill(32)}.0.{random.random()}"
new_versions = current_versions.copy()
new_versions['messages'] = new_msg_version
new_checkpoint = {
'v': current_checkpoint.checkpoint.get('v', 1) + 1,
'ts': datetime.utcnow().isoformat(),
'id': str(uuid.uuid4()),
'channel_versions': new_versions,
'versions_seen': current_checkpoint.checkpoint.get('versions_seen', {}),
'updated_channels': ['messages'],
'channel_values': {'messages': new_messages}
}
existing_metadata = current_checkpoint.metadata or {}
new_metadata = {
**existing_metadata,
"source": "compaction",
"compacted_at": datetime.utcnow().isoformat(),
}
if "step" not in new_metadata:
new_metadata["step"] = existing_metadata.get("step", 0)
self.memory.put(
config=checkpoint_config,
checkpoint=new_checkpoint,
metadata=new_metadata,
new_versions={'messages': new_msg_version}
)
return True
|