dao-ai 0.0.35__py3-none-any.whl → 0.1.0__py3-none-any.whl

dao_ai/middleware/summarization.py

@@ -0,0 +1,192 @@
+"""
+Summarization middleware for DAO AI agents.
+
+This module provides a LoggingSummarizationMiddleware that extends LangChain's
+built-in SummarizationMiddleware with logging capabilities, and provides
+helper utilities for creating summarization middleware from DAO AI configuration.
+
+The middleware automatically:
+- Summarizes older messages using a separate LLM call when thresholds are exceeded
+- Replaces them with a summary message in State (permanently)
+- Keeps recent messages intact for context
+- Logs when summarization is triggered and completed
+
+Example:
+    from dao_ai.middleware import create_summarization_middleware
+    from dao_ai.config import ChatHistoryModel, LLMModel
+
+    chat_history = ChatHistoryModel(
+        model=LLMModel(name="gpt-4o-mini"),
+        max_tokens=256,
+        max_tokens_before_summary=4000,
+    )
+
+    middleware = create_summarization_middleware(chat_history)
+"""
+
+from typing import Any, Tuple
+
+from langchain.agents.middleware import SummarizationMiddleware
+from langchain_core.language_models import LanguageModelLike
+from langchain_core.messages import BaseMessage
+from langgraph.runtime import Runtime
+from loguru import logger
+
+from dao_ai.config import ChatHistoryModel
+
+__all__ = [
+    "SummarizationMiddleware",
+    "LoggingSummarizationMiddleware",
+    "create_summarization_middleware",
+]
+
+
+class LoggingSummarizationMiddleware(SummarizationMiddleware):
+    """
+    SummarizationMiddleware with logging for when summarization occurs.
+
+    This extends LangChain's SummarizationMiddleware to add logging at INFO level
+    when summarization is triggered and completed, providing visibility into
+    when conversation history is being summarized.
+
+    Logs include:
+    - Original message count and approximate token count (before summarization)
+    - New message count and approximate token count (after summarization)
+    - Number of messages that were summarized
+    """
+
+    def _log_summarization(
+        self,
+        original_message_count: int,
+        original_token_count: int,
+        result_messages: list[Any],
+    ) -> None:
+        """Log summarization details with before/after metrics."""
+        # Result messages: [RemoveMessage, summary_message, ...preserved_messages]
+        # New message count excludes RemoveMessage (index 0)
+        new_messages = [
+            msg for msg in result_messages if not self._is_remove_message(msg)
+        ]
+        new_message_count = len(new_messages)
+        new_token_count = self.token_counter(new_messages) if new_messages else 0
+
+        # Calculate how many messages were summarized
+        # preserved = new_messages - 1 (the summary message)
+        preserved_count = max(0, new_message_count - 1)
+        summarized_count = original_message_count - preserved_count
+
+        logger.info(
+            f"Conversation summarized: "
+            f"BEFORE: {original_message_count} messages (~{original_token_count:,} tokens) → "
+            f"AFTER: {new_message_count} messages (~{new_token_count:,} tokens) | "
+            f"{summarized_count} messages condensed into 1 summary"
+        )
+        logger.debug(
+            f"Summarization details: trigger={self.trigger}, keep={self.keep}, "
+            f"preserved_messages={preserved_count}, "
+            f"token_reduction={original_token_count - new_token_count:,}"
+        )
+
+    def _is_remove_message(self, msg: Any) -> bool:
+        """Check if a message is a RemoveMessage."""
+        return type(msg).__name__ == "RemoveMessage"
+
+    def before_model(
+        self, state: dict[str, Any], runtime: Runtime
+    ) -> dict[str, Any] | None:
+        """Process messages before model invocation, logging when summarization occurs."""
+        messages: list[BaseMessage] = state.get("messages", [])
+        original_message_count = len(messages)
+        original_token_count = self.token_counter(messages) if messages else 0
+
+        result = super().before_model(state, runtime)
+
+        if result is not None:
+            result_messages = result.get("messages", [])
+            self._log_summarization(
+                original_message_count,
+                original_token_count,
+                result_messages,
+            )
+
+        return result
+
+    async def abefore_model(
+        self, state: dict[str, Any], runtime: Runtime
+    ) -> dict[str, Any] | None:
+        """Process messages before model invocation (async), logging when summarization occurs."""
+        messages: list[BaseMessage] = state.get("messages", [])
+        original_message_count = len(messages)
+        original_token_count = self.token_counter(messages) if messages else 0
+
+        result = await super().abefore_model(state, runtime)
+
+        if result is not None:
+            result_messages = result.get("messages", [])
+            self._log_summarization(
+                original_message_count,
+                original_token_count,
+                result_messages,
+            )
+
+        return result
+
+
+def create_summarization_middleware(
+    chat_history: ChatHistoryModel,
+) -> LoggingSummarizationMiddleware:
+    """
+    Create a LoggingSummarizationMiddleware from DAO AI ChatHistoryModel configuration.
+
+    This factory function creates a LoggingSummarizationMiddleware instance
+    configured according to the DAO AI ChatHistoryModel settings. The middleware
+    includes logging at INFO level when summarization is triggered.
+
+    Args:
+        chat_history: ChatHistoryModel configuration for summarization
+
+    Returns:
+        LoggingSummarizationMiddleware configured with the specified parameters
+
+    Example:
+        from dao_ai.config import ChatHistoryModel, LLMModel
+
+        chat_history = ChatHistoryModel(
+            model=LLMModel(name="gpt-4o-mini"),
+            max_tokens=256,
+            max_tokens_before_summary=4000,
+        )
+
+        middleware = create_summarization_middleware(chat_history)
+    """
+    logger.debug(
+        f"Creating summarization middleware with max_tokens: {chat_history.max_tokens}, "
+        f"max_tokens_before_summary: {chat_history.max_tokens_before_summary}, "
+        f"max_messages_before_summary: {chat_history.max_messages_before_summary}"
+    )
+
+    # Get the LLM model
+    model: LanguageModelLike = chat_history.model.as_chat_model()
+
+    # Determine trigger condition
+    # LangChain uses ("tokens", value) or ("messages", value) tuples
+    trigger: Tuple[str, int]
+    if chat_history.max_tokens_before_summary:
+        trigger = ("tokens", chat_history.max_tokens_before_summary)
+    elif chat_history.max_messages_before_summary:
+        trigger = ("messages", chat_history.max_messages_before_summary)
+    else:
+        # Default to a reasonable token threshold
+        trigger = ("tokens", chat_history.max_tokens * 10)
+
+    # Determine keep condition - how many recent messages/tokens to preserve
+    # Default to keeping enough for context
+    keep: Tuple[str, int] = ("tokens", chat_history.max_tokens)
+
+    logger.info(f"Summarization middleware configured: trigger={trigger}, keep={keep}")
+
+    return LoggingSummarizationMiddleware(
+        model=model,
+        trigger=trigger,
+        keep=keep,
+    )