from datetime import datetime from langchain_core.messages import BaseMessage, HumanMessage, AIMessage, SystemMessage from app.utils.token_counter import count_messages_tokens from app.config import settings class ConversationBuffer: """Sliding window conversation buffer with token-based trimming.""" def __init__(self): self.messages: list[BaseMessage] = [] self.last_accessed: datetime = datetime.utcnow() def add_user_message(self, content: str) -> None: self.messages.append(HumanMessage(content=content)) self._trim() self.last_accessed = datetime.utcnow() def add_ai_message(self, content: str) -> None: self.messages.append(AIMessage(content=content)) self._trim() self.last_accessed = datetime.utcnow() def get_messages(self, max_tokens: int | None = None) -> list[BaseMessage]: """Return messages, token-trimmed to max_tokens (default from settings).""" self.last_accessed = datetime.utcnow() limit = max_tokens or settings.max_context_tokens return self._token_trim(self.messages.copy(), limit) def _trim(self) -> None: """Apply sliding window: keep last MAX_CONVERSATION_MESSAGES.""" max_msgs = settings.max_conversation_messages if len(self.messages) > max_msgs: self.messages = self.messages[-max_msgs:] def _token_trim(self, messages: list[BaseMessage], max_tokens: int) -> list[BaseMessage]: """ Trim oldest messages (excluding last user message) until total tokens <= max_tokens. Always preserves the most recent user message. """ if not messages: return messages total = count_messages_tokens(messages) if total <= max_tokens: return messages # Always keep the last message (current user query) last = messages[-1] trimable = messages[:-1] while trimable and count_messages_tokens(trimable) + count_messages_tokens([last]) > max_tokens: trimable.pop(0) return trimable + [last] def clear(self) -> None: self.messages = [] def __len__(self) -> int: return len(self.messages)