Skip to content

Configuration reference

This page lists every configuration key accepted by genai-otel-bridge, with types, defaults, and a brief description. The generated source of truth with inline comments is deploy/helm/values.yaml. The Go struct definitions and validation rules are in internal/config/config.go.

For a narrative walk-through, see Configuration.

emit

Key Type Default Description
emit.telemetry.otlp.endpoint string (required) Grafana Cloud OTLP gateway base URL (no trailing /v1/metrics). Must be https:// or loopback.
emit.telemetry.otlp.instance_id string ${GC_INSTANCE_ID} Grafana Cloud instance ID (Basic auth username).
emit.telemetry.otlp.token string ${GC_OTLP_TOKEN} Grafana Cloud access-policy token (Basic auth password).
emit.telemetry.otlp.allow_insecure bool false Opt out of the https-only gate for an in-cluster cleartext OTLP receiver. Requires empty credentials and a private/DNS target.
emit.self.otlp.* — (optional) Same structure as emit.telemetry.otlp. When unset, self-observability signals use the product endpoint.
emit.self.metric_interval duration 60s Self-obs PeriodicReader export period. Must be ≥ 60s (1 DPM constraint).

identity

Key Type Default Description
identity.service_namespace string genai-otel-bridge OTLP resource attribute service.namespace. Appears as a Loki stream label.
identity.deployment_environment string ${ENV} OTLP resource attribute deployment.environment.name (e.g. dev, prod).

ha

Key Type Default Description
ha.coordinator string lease lease — Kubernetes Lease leader election. none — single-replica / dev.
ha.checkpoint string configmap configmap — watermarks in a Kubernetes ConfigMap (required with coordinator=lease). file — local file (dev only; unsafe with coordinator=lease).

queue

Key Type Default Description
queue.max_batches int 256 Per-loop in-memory queue depth (batches). Blocks on full (backpressure).
queue.max_batch_bytes int 1048576 Per-batch size cap in bytes (~1 MiB). Over-cap batches are split proactively; a 413 from the gateway triggers a reactive split.
queue.emit_workers int 1 Must be 1. Per-loop single-flight emit so the watermark advances monotonically.

governance

Key Type Default Description
governance.per_metric_cardinality_budget int 10000 Max distinct label-sets per metric name. Over-budget series are dropped and counted as genai_otel_bridge_guard_dropped_total.
governance.max_dpm int 1 Hard cap on data points per minute per series (both planes). Drives the product-plane LWW coalesce and clamps the self-obs PeriodicReader interval.
governance.max_catchup_per_tick int 1 Max windows drained per cadence period when a loop is backlogged. 1 = no catch-up acceleration (the default).
governance.max_stream_label_keys int 15 Max OTLP resource attributes a single logs loop may contribute as Loki stream labels. The Grafana Cloud Loki default ceiling is 15; the bridge fails fast at startup if a loop would exceed it.
governance.allow_label_keys []string [] Extra content-free attribute keys added to the label allow-list on top of each source's declared keys. Content-floor keys (message bodies, PII) are rejected at startup.

log

Key Type Default Description
log.format string logfmt logfmt or json. Applies to the bridge's own operational stdout logs (not OTLP).
log.level string info debug, info, warn, or error.

selfobs

Key Type Default Description
selfobs.profiling.enabled bool false Enable continuous profiling of the bridge's own runtime.
selfobs.profiling.mode string pull pull — expose net/http/pprof on a dedicated listener. push — push to Grafana Cloud Profiles via the pyroscope-go agent.
selfobs.profiling.pull.addr string :6060 pprof listener address (pull mode).
selfobs.profiling.push.endpoint string (required when push) Grafana Cloud Profiles ingest URL. Must be https://.
selfobs.profiling.push.instance_id string (required when push) Grafana Cloud instance ID for profiles.
selfobs.profiling.push.token string (required when push) Grafana Cloud access-policy token for profiles (requires profiles:write scope).
selfobs.tracing.enabled bool false Enable opt-in self-APM tracing of the bridge's own poll/emit pipeline. Spans are exported to the same endpoint as self-metrics.

sources

Each entry in the sources list has:

Key Type Default Description
type string portkey Source type. Currently: portkey, langsmith.
enabled bool true Enable this source.
base_url string https://api.portkey.ai/v1 Source API base URL. Must be https:// unless http.allow_private=true.
source_instance string portkey-${ENV} Stable per-environment identifier. Part of the CheckpointKey; changing it resets the watermark. Must not contain /.
auth.header string x-portkey-api-key HTTP request header name for the API key.
auth.value string ${PORTKEY_API_KEY} API key value (use a ${ENV_VAR} or file: reference).
rate_limit.rps float 1 Sustained outbound request rate (requests/second).
rate_limit.burst int 3 Token bucket burst size.
http.user_agent string — Override User-Agent (required for some endpoints, e.g. LangSmith behind a WAF).
http.allow_private bool false Allow non-loopback, non-https base URLs (for in-VPC sources).
api_key_use_cases list — Maps human use-case labels to Portkey API key UUIDs. See the Portkey guide.

Per-loop config (sources[].loops.<name>)

Key Type Default Description
enabled bool true Enable this loop.
cadence duration 60s Poll interval (±10% jitter). Must be ≥ 10s.
window duration 50m Time range queried per collect. Must be ≤ 55m (Portkey granularity clamp).
bucket_settle duration 10m Age at which a bucket is considered final.
bootstrap_lookback duration 50m How far back to bootstrap on first run or watermark reset. Must be ≤ max_backfill.
max_backfill duration 90m Maximum backfill depth. On Grafana Cloud, the Mimir out-of-order accept window is 2h; 90m leaves margin.
metric_prefix string portkey_api Prefix applied to all metrics emitted by this loop.
graphs []string [requests, cost, tokens, latency, errors] Which Portkey analytics graphs to collect.
settings map[string]string — Source-specific knobs. Each source package documents its own keys. See the Portkey guide and LangSmith guide.

Window constraint

The window must satisfy window ≥ cadence × 1.2 + bucket_settle to ensure no time is left uncovered between jittered ticks. The config validator enforces this and reports the constraint in the error message.

Validation rules

The config validator (internal/config/config.go) enforces these rules at startup:

  • emit.telemetry.otlp.endpoint is required.
  • queue.emit_workers must be exactly 1.
  • ha.checkpoint=file with ha.coordinator=lease is rejected.
  • source_instance must not contain /.
  • source.base_url must be https:// unless http.allow_private=true.
  • auth.header and auth.value are required for every enabled source.
  • cadence must be ≥ 10s.
  • window must be ≤ 55m and satisfy the cadence×1.2+bucket_settle lower bound.
  • bootstrap_lookback must be ≤ max_backfill.
  • governance.max_dpm, governance.per_metric_cardinality_budget, governance.max_catchup_per_tick, governance.max_stream_label_keys must be ≥ 0.
  • log.format must be logfmt, json, or empty (defaults to logfmt).
  • log.level must be debug, info, warn, error, or empty (defaults to info).