time_to_first_token

Measures time-to-first-token for streaming AI responses.

Configuration Notes

Activates only for successful text/event-stream responses. On the first non-empty body chunk, records ctx.request_start.elapsed() as a Prometheus histogram and deactivates for the remainder of the response.

The histogram’s model label is read from metadata set by an upstream format filter (openai_responses_format, anthropic_messages_format, or anthropic_messages_to_chat_completions). If no format filter runs before this filter, all TTFT samples are labeled unknown.

Example

filter: time_to_first_token