Rates and ceilings
The API reference lists every endpoint with its request and response schema, rendered from the OpenAPI document; this page explains how the calls behave.
Rate limits
Every limit below is per workspace. A workspace on a trial gets half of each rate, because that traffic is unvetted; every other plan gets the full number. Over a rate limit the call returns 429 with Retry-After; nothing is queued, so retry after the window rather than immediately. Reads and writes are counted separately, so polling a result can never use up the budget for submitting one.
| Calls | Limit | On a trial |
|---|---|---|
QA writes: POST /qa-adapter/v1/check, /check-segments, /runs, the finding decisions, share links and deletes | 60 per minute. | 30 |
Runs: POST /qa-adapter/v1/runs, which also counts as a QA write | 30 per minute. | 15 |
Compare: POST /qa-adapter/v1/compare | 20 per minute. | 10 |
QA reads: polling a check, a change report, /reports and its downloads, /profiles, /usage | 300 per minute. | 150 |
LQA writes: POST /lqa/v1/... and DELETE /lqa/v1/templates/{id} | 30 per minute. | 15 |
LQA reads: GET /lqa/v1/... | 60 per minute. | 30 |
Eval runs: POST /eval/v1/runs | 20 per minute. | 10 |
Eval reads: GET /eval/v1/..., plus POST /eval/v1/source/upload and POST /eval/v1/calibrate | 60 per minute. | 30 |
CSV exports: GET /lqa/v1/export.csv and GET /eval/v1/export.csv, each also counted as an LQA or Eval read | 10 per minute. | 5 |
| Eval runs in generate mode, which buys MT and LLM output. | 2 at once, 20 a day. | the same |
The AI ceiling
A QA check's AI review reads at most 20,000 segments in one run. While a workspace is on a trial the ceiling is 3,000. A larger file is still checked in full by the deterministic rules; the AI pass stops at the ceiling and the report says so. A one million word job is around seventy thousand segments, so on a trial its AI review covers the first three thousand and the rules cover all of it.
There is also a daily AI budget, and on a trial it is 75,000 segments. It counts what each run's AI review could read, which is the run's size or the ceiling above, whichever is smaller. One enormous job therefore costs its ceiling rather than its size. Other plans have no daily budget unless one is set for them. When the day's budget is gone the AI review is skipped for the rest of the day and the report says so. The run still happens and the deterministic checks are untouched. The budget resets at midnight in the workspace's time zone.
The deterministic rules have no ceiling and no budget, on any plan.
These are the shipped numbers. If an integration needs more, ask us rather than working around the limit. They are configurable per deployment. A client that retries into a 429 loop is slower than one that waits.
Retention and polling
The job files uploaded to make a report are removed on their own within 24 hours; the QA result and the finalized LQA scorecard stay stored under your workspace until someone deletes them or the workspace's retention setting removes them (Retention). A workspace can agree a longer keep for its uploaded files with us; we set it per workspace. Keep polling to a few seconds between calls and prefer webhooks to learn that work is done. An LQA upload is capped at 64 MB, and a large one can be sampled with sample_words.