Skip to main content
Version: Next

Debug Log Redaction

At the debug log level, the external processor logs every request and response it processes. Those bodies carry prompts, conversation history and other caller data, and debug logs are usually readable by more people and retained longer than the requests themselves. enableRedaction replaces that data with placeholders, so debug logs can be used for troubleshooting without persisting what callers sent.

Enabling redaction​

Set both values in the Helm chart:

extProc:
logLevel: debug
enableRedaction: true

enableRedaction only takes effect when logLevel is debug:

logLevelenableRedactionRequest and response bodies in the debug log
not debuganyNot logged
debugfalseLogged as received
debugtrueLogged with sensitive data replaced by placeholders

Credential headers set by the gateway for upstream authentication (authorization, x-api-key, api-key, x-amz-security-token) are redacted whenever debug logging is on, regardless of enableRedaction.

Placeholder format​

Each redacted value is replaced with:

[REDACTED LENGTH=n HASH=h]

n is the length of the original value in bytes and h is a prefix of its SHA-256 hash. The hash is deterministic: the same value always produces the same placeholder. This lets you tell whether two requests sent the same prompt or came from the same end user, or confirm that a known value was received, without the value itself appearing in the log.

What is redacted​

Request bodies are parsed and each endpoint redacts its fields according to what the value is, not who sets it:

  • Redacted: caller content. Messages, prompts, instructions, documents, images, audio and tool call arguments and results.
  • Redacted: end-user identifiers and free-form caller annotations, such as user, safety_identifier and metadata values. These are usually set by the application rather than typed by the end user, but their values identify a person, and the gateway cannot tell whether the application hashed them.
  • Kept: model and format configuration, such as the model name, sampling parameters, tool definitions and response format schemas. These describe how to call the model rather than who is calling or what they said, and keeping them makes the logs useful for debugging.

The external processor logs the response body as a single placeholder. Translators additionally log the parsed response with message content, refusals, reasoning, tool arguments, audio, citation URLs and titles, and grounding metadata redacted; metadata such as the response id, model, usage and tool names is kept.

Error responses generated by the gateway itself, such as a rejected malformed request, are logged as is.

Fields that an endpoint does not explicitly redact are logged as received. Treat redacted debug logs as sensitive, and do not enable debug logging in production for longer than you need it.

Relationship to trace redaction​

enableRedaction only affects the external processor's debug logs. Content recorded in OpenTelemetry traces is controlled separately by the OPENINFERENCE_HIDE_* environment variables described in GenAI Tracing. Enabling one has no effect on the other.