Function run

optillm/plugins/compact_plugin.py:155–197

(
    system_prompt: str,
    initial_query: str,
    client,
    model: str,
    request_config: Optional[dict] = None,
)

def run(
    system_prompt: str,
    initial_query: str,
    client,
    model: str,
    request_config: Optional[dict] = None,
) -> Tuple[str, int]:
    context_window = _get_context_window(client, model, request_config)
    threshold = _get_config(request_config, "compact_threshold", "COMPACT_THRESHOLD", DEFAULT_THRESHOLD)
    keep_recent = _get_config(request_config, "compact_keep_recent", "COMPACT_KEEP_RECENT", DEFAULT_KEEP_RECENT)

    token_count = estimate_tokens(initial_query)
    budget = int(context_window * threshold)

    if token_count < budget:
        logger.debug(f"Compact: passthrough ({token_count} tokens < {budget} budget)")
        return initial_query, 0

    turns = parse_tagged_conversation(initial_query)
    if len(turns) <= keep_recent:
        logger.debug(f"Compact: too few turns to compress ({len(turns)} <= {keep_recent})")
        return initial_query, 0

    split_idx = len(turns) - keep_recent
    older_turns = turns[:split_idx]
    recent_turns = turns[split_idx:]

    logger.info(f"Compact: compressing {len(older_turns)} older turns, keeping {len(recent_turns)} recent")

    summary, tokens_used = compress_with_llm(older_turns, system_prompt, client, model)

    if summary is None:
        logger.warning("Compact: compression failed, returning original query")
        return initial_query, 0

    compressed_turns = [("user", f"[Conversation summary]:\n{summary}")]
    compressed_turns.extend(recent_turns)

    result = reconstruct_tagged(compressed_turns)
    new_token_count = estimate_tokens(result)
    logger.info(f"Compact: {token_count} -> {new_token_count} tokens (used {tokens_used} for compression)")

    return result, tokens_used