Module server

Expand description

HTTP + WebSocket server that accepts audio and streams transcripts.

Single port serves both REST API (health, transcribe, SSE) and WebSocket.

Modules§

http: HTTP handlers for REST API endpoints.
metrics: Minimal Prometheus text-exposition registry (replaces metrics-exporter-prometheus).
rate_limit: Per-IP token-bucket rate limiter.

OriginPolicy: Origin policy for CORS + cross-origin deny middleware.
RuntimeLimits: Runtime limits surfaced to per-request handlers. Separate from ServerConfig because it needs to travel through http::AppState to the WebSocket handler.
ServerConfig: Server runtime configuration. run_with_config is the canonical entry point; run / run_with_shutdown remain as thin wrappers for callers that only need the pre-0.6 positional parameters.

FrameOutcome: Outcome returned by per-frame handlers. Keeps handle_ws_inner a thin orchestration loop instead of a 250-line one-big-match.

handle_binary_frame: Handle a single PCM16 audio frame: resample if needed, run inference in a spawn_blocking guarded by catch_unwind, and emit partial/final/error payloads. Always returns the triplet to triplet_opt (or recovers a fresh state after an inference panic) so the connection can keep serving.
handle_ws
run: Start the HTTP + WebSocket STT server on the given host and port.
run_with_config: Start server with a full ServerConfig and optional programmatic shutdown signal. This is the canonical entry point — the other run_* helpers construct a default ServerConfig and dispatch here.
run_with_shutdown: Start server with an optional programmatic shutdown signal.
session_deadline_instant: Clamp shutdown_drain_secs to a minimum of 1 so the drain window is never zero-length (which would immediately kill in-flight tasks). fires in practice.
ws_handler
ws_handler_legacy: Deprecated WebSocket endpoint at /ws. Identical behaviour to /v1/ws but emits a warn-level log on every upgrade and adds RFC 8594 Deprecation plus Link: </v1/ws>; rel="successor-version" headers on the upgrade response so client libraries can surface the migration warning to users before v1.0 drops the alias.
ws_shutdown_response