Content Length

The Content Length guardrail enforces a minimum and / or maximum on the character count of the concatenated message content. Cheapest possible defence against:

• One-character inputs that aren’t worth a provider call (min)
• Multi-megabyte inputs that would blow your token budget (max)
• Buffer overflow attempts and DoS-by-megaprompt

Configuration

{
  "enabled": true,
  "min": 5,
  "max": 50000
}

Either or both can be set. Both are optional — omitting one means no bound in that direction.

The mode is implicitly block — there’s no useful “mask” semantics for length.

Behaviour

Input length	min: 5, max: 50000	Result
0 chars	violates min	422
1–4 chars	violates min	422
5–50 000 chars	OK	passes
50 001+ chars	violates max	422

422 response:

{
  "ok": false,
  "error": "Request blocked: content length 80012 exceeds maximum 50000."
}

Multi-byte handling

Length is measured with mb_strlen() — character count, not byte count. So "日本語" is 3, not 9. Match what your usage_hard_limit_tokens budget expects.

Combining with budgets

This guardrail is a fast length cap that runs before guardrails / provider call. The budget enforcer has a per-request token cap (usage_hard_limit_tokens) that estimates tokens at ~4 chars/token.

In practice they cover similar ground but at different levels:

• Content Length — character cap, runs in the guardrail step. Prevents degenerate inputs from even entering the pipeline.
• Token cap — token-estimate cap, runs in the budget step. Closer to actual provider cost.

You usually want both: a generous content-length cap (e.g. 100 000 chars) to reject truly absurd inputs, and a tighter token cap (e.g. 8 000 tokens) tuned to your endpoint’s actual context window.

Performance

mb_strlen() of a few hundred KB completes in microseconds. Free compared to anything else.

Recommended defaults

For most endpoints:

{
  "enabled": true,
  "min": 1,
  "max": 50000
}

For high-context endpoints (long-document summarisation, etc.):

{
  "enabled": true,
  "min": 100,
  "max": 500000
}

For short-form classifiers (single-sentence input):

{
  "enabled": true,
  "min": 1,
  "max": 1000
}

When to enable

✅ Always. The cost is microseconds and the protection is meaningful (DoS, runaway tokens, degenerate inputs).

❌ Never disable globally unless you’ve convinced yourself another layer is catching the same thing.

---

Next: Rate Limits.

---

© Akyros Labs LLC. All rights reserved.