AI prompt and response logging

Opt-in capture of AI prompts and responses with PII redaction, sampling, and retention. Capture is OFF by default (privacy-first); when on, a redaction pass scrubs PII and secrets before storage, sampling controls volume, and retention ages records out.

When to use this

• You need an audit trail of what was sent to and returned by AI providers -- for incident review, compliance, or quality analysis.
• PII or secrets may appear in prompts and responses; capture scrubs them before anything is stored.

Configuration

ai:
  logging:
    enabled: true
    sample_rate: 0.01          # capture 1% of requests
    retention_secs: 604800     # 7 days
    redaction:
      patterns:                # custom patterns beyond the built-ins
        - "ACME-\\d{6}"
      replacement: "[REDACTED]"

Field	Default	Description
enabled	false	Master switch. Off by default (privacy-first).
sample_rate	1.0	Fraction of requests to capture (0.0 to 1.0). Sampling is deterministic by request id -- the same request id always captures or skips.
retention_secs	604800 (7 days)	Records older than this are deleted by the analytics maintenance tick.
redaction.patterns	(empty)	Additional regex patterns to scrub beyond the built-in PII patterns.
redaction.replacement	[REDACTED]	String that replaces redacted content.

Per-consumer toggle

A consumer can override the global setting:

consumers:
  - name: acme
    credentials:
      - type: api_key
        key: ${ACME_KEY}
    ai_logging: false          # disable even if global is on

ai_logging: false disables capture for that consumer even when the global ai.logging.enabled is true. ai_logging: true enables it even when the global is off. Omit the field to inherit the global setting.

PII redaction

Built-in patterns (always active when logging is on):

• Email addresses
• Phone numbers (US and international)
• API keys (OpenAI sk-, Anthropic, GitHub, Slack, AWS AKIA, Bearer tokens)
• Credit card numbers

Custom patterns from config are applied alongside the built-ins. The redaction pass runs on all string values in the serialized prompt and response JSON before storage -- no PII reaches the log store.

Querying logs

Prompt logs are queryable via the admin API:

POST /analytics/prompt-logs
{
  "from_ms": 1693526400000,
  "to_ms": 1693612800000,
  "consumer": "acme",
  "limit": 100
}

The response is one row per captured request with: request id, consumer, route, provider, model, version, redacted prompt JSON, redacted response JSON, and a stream flag.

Streaming

For streaming responses, the prompt is captured and redacted in full. The response is marked as {"streamed": true} -- the full streamed content is not reassembled for logging (that would require buffering the entire stream, contradicting the zero-buffer design).

Runnable demo

Run prompt logging against a live gateway: demos/07-ai-gateway/ in the repository (test script: test-08-prompt-logging.sh, which queries captured rows via the mTLS admin API). The category README covers prerequisites and teardown.

See also

• AI gateway