Limitations and roadmap#

Limitations#

  • One call per guard. Inline guards are not combined into one Enkrypt call; a saved guardrail is the way to run several detectors in one call.

  • Rate limits. Enkrypt’s default plan allows 10 calls a minute. Guards wait rather than fail, which adds latency; see Rate limits. Waiting relies on Enkrypt’s rate-limit headers, which are not documented by Enkrypt.

  • Personal data has its own limits: detection, altered placeholders, encoded data, structured values and concurrent runs; see Limits.

  • Tool arguments can be checked and restored, but not redacted.

  • No grounding checks. Enkrypt’s hallucination and relevancy detectors are not exposed yet.

Not yet built#

  • restoring placeholders inside structured outputs;

  • combining inline guards into a single Enkrypt call;

  • grounding and relevancy guards;

  • choosing how conversations are identified for placeholders, for example from deps;

  • a callback for findings, in addition to logging.