Skip to main content
Use Platform Settings for organization-level telemetry. Traces, call-event diagnostics, and workflow webhooks are different integrations with different payloads and delivery behavior.

​
Connect Langfuse

  1. Open Platform Settings → Telemetry.
  2. Enter Host, Public Key, Secret Key, and Project ID for your Langfuse project. The project ID identifies the project in trace links.
  3. Leave Make traces publicly viewable off unless you intend anyone holding a trace URL to view its contents, including conversation and tool data.
  4. Choose Save. After an eligible controlled call, inspect its trace in the intended Langfuse project and correlate it with the Athena run.
  5. Use Remove if the organization should stop using those credentials. Removal is separate from deletion of previously exported traces at the destination.
A successful settings save is not a trace-delivery test. If a trace is missing, check project identity, host, provider access, and the run’s telemetry support. Do not enable public traces merely to troubleshoot authentication.

​
Prepare a BigQuery call-events table

Use an existing table whose schema matches the columns below. The service identity needs permission to read the table schema and insert rows. Use nullable, non-repeated columns; additional columns must not be required unless Athena can populate them. For example, replace this illustrative table identifier with your own project, dataset, and table before running it in your BigQuery environment:

​
Connect the event destination

  1. Open Platform Settings → Call events and choose BigQuery as the destination.
  2. Set Table to project.dataset.table.
  3. Choose Service account, then enter its Service account email and Private key in the secure form. If Deployment identity is offered and provisioned for your account, it uses the hosted deployment’s identity instead.
  4. Choose Test connection and resolve schema or permission errors. The connection check does not prove a completed call exported its events.
  5. Enable Export call events and save. After a controlled call, query the destination for its run ID and inspect the expected diagnostic events.
Saved private keys are masked; leave the masked value unchanged to retain the current key. Diagnostics can include stage latencies, silence events, and call outcomes. Avoid treating a call with no applicable event as evidence that every exporter stage failed.

​
Delivery and changes

Call-event delivery makes up to three attempts in a bounded window. Events can be discarded after delivery failure, a destination change, or queue exhaustion. This is not an unlimited durable queue and does not provide exactly-once delivery. Destination retries use best-effort deduplication; do not build an exactly-once assumption around the exported rows. Disable export or remove the destination to stop future use, then check the saved state. Previously exported records remain at the destination. Workflow-webhook retry/dead-letter rules are separate; see Webhooks.