Limits

What Curltale AI does not do yet.

• One model per installation, one active connector at a time.
• Row-level security (rows filtered per caller) is not implemented; access is decided per column.
• Agents receive at most AGENT_MAX_ROWS rows per query; results are not paged (they are told the total and to aggregate or filter).
• Rate limits are kept in memory: one app process. Several app instances need a shared store.
• Sign-in is username and password; single sign-on and multi-factor authentication are not built in yet.
• Translation of the model's SQL to other databases is best effort; constructs that do not translate are reported by the Model check.
• The demo database is generic: it follows the model's keys, links and allowed values, not your business rules.
• Library: collections are isolated by grants, not by separate databases; one embedding model and dimension per installation (changing it means re-embedding); tokens get read grants only; scanned pages are read by a model, up to OCR_MAX_PAGES per document; the browser service does not fetch pages behind a login; collection export has no matching import yet; changing the storage driver does not move files already stored.
• State-changing browser requests rely on the session cookie's SameSite setting; there is no separate CSRF token.