Class LlmEventSummarizer

src/google/adk/apps/llm_event_summarizer.py:29–171

An LLM-based event summarizer for sliding window compaction. This class is responsible for summarizing a provided list of events into a single compacted event. It is designed to be used as part of a sliding window compaction process. The actual logic for determining *when* to trigger compa

class LlmEventSummarizer(BaseEventsSummarizer):
  """An LLM-based event summarizer for sliding window compaction.

  This class is responsible for summarizing a provided list of events into a
  single compacted event. It is designed to be used as part of a sliding window
  compaction process.

  The actual logic for determining *when* to trigger compaction and *which*
  events form the sliding window (based on parameters like
  `compaction_invocation_threshold` and `overlap_size` from
  `EventsCompactionConfig`) is handled by an external component, such as an ADK
  "Runner". This compactor focuses solely on generating a summary of the events
  it receives.

  When `maybe_compact_events` is called with a list of events, this class
  formats the events, generates a summary using an LLM, and returns a new
  `Event` containing the summary within an `EventCompaction`.
  """

  _DEFAULT_PROMPT_TEMPLATE = (
      'The following is a conversation history between a user and an AI agent.'
      ' It may or may not start from a compacted history. Please identify and'
      ' reiterate the user request, summarize the context so far, focusing on'
      ' key decisions made and information obtained, as well as any unresolved'
      ' questions or tasks. The summary should be concise and capture the'
      ' essence of the interaction.\n\n{conversation_history}'
  )

  # Tool call args and responses can be large (e.g. search results). Cap how
  # much of each is rendered so compaction does not inflate the very context
  # it exists to shrink.
  _MAX_TOOL_CONTENT_CHARS = 2000

  def __init__(
      self,
      llm: BaseLlm,
      prompt_template: Optional[str] = None,
  ):
    """Initializes the LlmEventSummarizer.

    Args:
        llm: The LLM used for summarization.
        prompt_template: An optional template string for the summarization
          prompt. If not provided, a default template will be used. The template
          should contain a '{conversation_history}' placeholder.
    """
    self._llm = llm
    self._prompt_template = prompt_template or self._DEFAULT_PROMPT_TEMPLATE

  def _format_events_for_prompt(self, events: list[Event]) -> str:
    """Formats events into prompt text, including thoughts and tool calls.

    Thoughts carry the agent's analysis of tool responses, and tool calls and
    responses carry the evidence retrieved so far, so all three are included.
    Thoughts emitted by a compaction event are skipped so a prior summary's
    reasoning does not leak into the next summary.
    """
    formatted_history = []