zylon-ai/private-gpt · error · ValueError

Condensed history exceeds maximum length after applying cond

Error message

Condensed history exceeds maximum length after applying condensation strategy.

What it means

Post-condition check in the TLDR processor: after build_condensed_history assembled system messages + condensed conversation history, the final token count still exceeds max_length. The chosen strategy produced a summary/trimmed history that, together with the always-kept system messages, is too large for the budget.

Source

Thrown at private_gpt/components/chat/processors/chat_history/memory/tldr_processor.py:159

        condensation_timeout,
        **kwargs,
    )

    # 4. Build final history with TLDR
    condensed_history, condense_blocks = await asyncio.to_thread(
        build_condensed_history,
        system_messages=system_messages,
        conversation_history=conversation_history,
    )

    current_token_count = await estimate_token_count(
        condensed_history,
        tokenizer_fn=tokenizer_fn,
        message_to_input=message_to_input,
        **kwargs,
    )
    if current_token_count > max_length:
        raise ValueError(
            "Condensed history exceeds maximum length after applying condensation strategy."
        )

    logger.debug(
        "Applied condensation strategy '%s' to conversation history. "
        "Current token count: %d, Available max length: %d, Max length: %d",
        strategy_type,
        current_token_count,
        remaining_max_length,
        max_length,
    )
    yield CondenseResponse(
        is_condensed=True,
        chat_history=condensed_history,
        condense_blocks=condense_blocks,
    )

View on GitHub (pinned to 4a030776a3)

Solutions

1. Increase max_length so remaining budget after system messages fits a typical summary
2. Lower the number of recent messages retained by the strategy
3. Cap the summary output length in the strategy's LLM prompt
4. Shrink the system prompt to free budget (see also error 41)

Example fix

# before
condensed = build_condensed_history(system_messages, strategy_result)
# may exceed max_length

# after
condensed = build_condensed_history(system_messages, strategy_result)
assert await estimate_token_count(condensed, ...) <= max_length

remaining = max_length - system_token_count
if remaining < min_summary_budget:
    raise ConfigError(f"only {remaining} tokens left for condensed history")

try:
    async for resp in condense(...):
        ...
except ValueError as e:
    if "exceeds maximum length after applying" in str(e):
        # fall back to hard truncation of the condensed history
        condensed = condensed[:keep_last_n]
    else:
        raise

When it happens

Trigger: Running the tldr condense flow where system_token_count < max_length (passes error 41) but system + condensed history > max_length; the remaining_max_length given to the strategy is too small for the summary it generates.

Common situations: Large system prompt leaving a tiny remaining budget; an LLM-generated TLDR summary longer than the space left; aggressive keep-last-N strategy with many retained turns.

Related errors

• Condensed chat history exceeds maximum length after applying
• System messages alone exceed the maximum length. Cannot cond
• No user messages found in the chat history.
• Maximum number of iterations for condensing exceeded.
• The last user message exceeds the maximum length allowed.

AI-assisted analysis of zylon-ai/private-gpt@4a030776a3 (2026-08-15). Data as JSON: /api/errors/451f41e892f321aa. Report an issue: GitHub.