Function compact_history

web/pgadmin/llm/compaction.py:222–352

Compact conversation history to fit within a token budget. Strategy: 1. Always keep the first message (provides original context) 2. Always keep the last `recent_window` messages 3. Among remaining messages, classify by importance and drop lowest-value messages first 4. K

(
    messages: list[Message],
    max_tokens: int = DEFAULT_MAX_TOKENS,
    recent_window: int = DEFAULT_RECENT_WINDOW,
    provider: str = 'openai'
)

def compact_history(
    messages: list[Message],
    max_tokens: int = DEFAULT_MAX_TOKENS,
    recent_window: int = DEFAULT_RECENT_WINDOW,
    provider: str = 'openai'
) -> list[Message]:
    """Compact conversation history to fit within a token budget.

    Strategy:
    1. Always keep the first message (provides original context)
    2. Always keep the last `recent_window` messages
    3. Among remaining messages, classify by importance and drop
       lowest-value messages first
    4. Keep tool_call/tool_result pairs together

    Args:
        messages: Full conversation history.
        max_tokens: Maximum token budget for the history.
        recent_window: Number of recent messages to always preserve.
        provider: LLM provider name for token estimation.

    Returns:
        Compacted list of messages that fits within the token budget.
    """
    if not messages:
        return messages

    # Check if we're already within budget
    current_tokens = estimate_history_tokens(messages, provider)
    if current_tokens <= max_tokens:
        return messages

    total = len(messages)

    # Determine protected indices
    protected = set()

    # Always protect the first message
    protected.add(0)

    # Always protect the recent window
    recent_start = max(1, total - recent_window)
    for i in range(recent_start, total):
        protected.add(i)

    # If protected messages alone exceed the budget, shrink the
    # recent window until we have room for compaction candidates.
    while recent_window > 0:
        protected_tokens = sum(
            estimate_message_tokens(messages[i], provider)
            for i in protected
        )
        if protected_tokens <= max_tokens:
            break
        recent_window -= 1
        recent_start = max(1, total - recent_window)
        protected = {0} | set(range(recent_start, total))