Length Limits

InputLengthGuard and OutputLengthGuard block an LLM interaction when the character count exceeds a configured ceiling. Both return BLOCK over the limit and ALLOW otherwise; they never transform content.

InputLengthGuard sums len(message.content) across all messages in the input history (user, system, assistant, tool). OutputLengthGuard measures the assistant reply on event.output_message; if there is no output message, it allows.

Usage

max_chars must be a positive integer; non-positive values raise ValueError at construction. Each decision's meta carries total_chars and max_chars for logging.

from railtracks.prebuilt.guardrails import InputLengthGuard, OutputLengthGuard


input_length = InputLengthGuard(max_chars=4000)

result = input_length.decide("a" * 5000)
# result.action == rt.guardrails.GuardrailAction.BLOCK
# result.meta == {"total_chars": 5000, "max_chars": 4000}

output_length = OutputLengthGuard(max_chars=2000)

result = output_length.decide("ok")
# result.action == rt.guardrails.GuardrailAction.ALLOW

Scope

Counting is character-based and dependency-free. Word- or token-based counting (e.g. via tiktoken) is out of scope today and may arrive in later releases. Non-string content is treated as zero-length.