Skip to content

Audit log

Audit log

GatewAI can emit a structured slog record for every LLM proxy request. Disabled by default to avoid unexpected log volume.

Configuration

audit_log:
  enabled: true
  prompt: false   # include raw request body — opt-in only

audit_log is a top-level config key (not per-service).

Field	Default	Description
`enabled`	`false`	Emit one log record per LLM request
`prompt`	`false`	Include the raw request body as a `prompt` field. Opt-in only — the body may contain PII.

Log fields

Each record is a single slog.InfoContext call with the message "llm request" and the following key-value pairs:

Field	Type	Description
`service_type`	string	Service type (e.g. `llm`)
`model`	string	Client-facing model alias
`backend_model`	string	Model name forwarded to the backend
`provider`	string	Provider: `openai`, `anthropic`, `ollama`, `passthrough`
`consumer`	string	Consumer name from `consumer_header`; empty if not configured
`user_type`	string	User type from `user_type_header`; empty if not configured
`status`	int	HTTP status code returned to the client
`duration_ms`	int64	End-to-end request duration in milliseconds
`backend_url`	string	Backend URL that served the request
`stream`	bool	`true` if the request used SSE streaming
`prompt_tokens`	int	Prompt token count (absent for streaming responses)
`completion_tokens`	int	Completion token count (absent for streaming responses)
`prompt`	string	Raw request body (only present when `audit_log.prompt: true`)

Example log entry

{
  "time": "2026-03-05T10:04:32.123Z",
  "level": "INFO",
  "msg": "llm request",
  "service_type": "llm",
  "model": "chat-smart",
  "backend_model": "meta-llama/Meta-Llama-3-8B-Instruct",
  "provider": "passthrough",
  "consumer": "alice",
  "user_type": "user",
  "status": 200,
  "duration_ms": 842,
  "backend_url": "http://vllm-primary.default.svc.cluster.local:8000",
  "stream": false,
  "prompt_tokens": 312,
  "completion_tokens": 128
}

PII warning