Docs · limits and errors

Limits and errors

Daily limits

Free and without a key, with a daily limit so one caller cannot use up everyone's share. Each caller gets 300 calls a day across /api/v1/find, /api/v1/search, /api/v1/resolve, /api/answer and searches on this site; the count resets at 00:00 UTC. A repeated question is often served from our cache and may not count. Past the limit the reply is 429 with Retry-After (seconds until the reset). Every metered reply from /api/v1/find and /api/v1/resolve carries X-RateLimit-Limit and X-RateLimit-Remaining; replies from /api/v1/search and /api/answer are kept at the edge for five minutes, so they carry them only on a refusal. Passports, the public key and stats are not limited. Need more? Write to support@unfragile.ai.

Through Claude's connectors: one shared free allowance of 3,000 questions a day for all Claude users together; a key is not subject to it.

Errors

If the answer layer fails, answer.error says so; if the search service is down the call returns 503, never a guess. Every page is screened before the writer reads it: a page whose text carries instructions aimed at an AI reader ("ignore previous instructions", "don't tell the user", hidden characters) is withheld, never cited, and listed in answer.screened and on the receipt.

Call settings

Call settings. lane picks the work done: fast (the ranked results, no written answer), standard (the default) or best (a written answer, always fresh, never from memory). client_model (the model that will read the answer) and session_id (your own id for one task) are recorded on the signed receipt. A value we can't use is a 400 that names the parameter and what it accepts, before any charge or metering; an unknown parameter is answered, with a warning in call.warnings that it was ignored; a bounded value (k above 25) is used bounded, with a warning. Every response carries call: the lane, the two ids and the warnings.

Ground rules

  • Free without a key, up to 300 questions a day. With a key, $4.00 per 1,000 questions.
  • Deterministic retrieval. Same query, same results; every ranking reproducible. The only model is the one-sentence answer layer, and it cannot change a label or promote a demo or failed result.
  • Honest tiers. static-verified = contract machine-checked; execution-verified = we ran it in our sandbox and it worked; "not yet verified" is never "failed".