- Select BigQuery as the destination.
- Enter the table as
project.dataset.table. - Enter the email and PEM private key of a Google service account.
- Select Test connection to check authentication and table compatibility. This reads table metadata without inserting events; successful exports also require write permission.
- Enable Export call events and select Save.
ts and event can also be
required. The adapter does not create or migrate tables.
You receive one row per diagnostic event.
ts is when the event happened, and
detail contains JSON, stored as a native JSON value or serialized text according
to the column type. Events cover speech, accepted user turns,
mute state, tool execution, silence warnings, stage latencies, pipeline errors
and the final call summary. The transcription event records character counts
and empty/finalized flags; it does not contain transcript text or audio.
Delivery behavior
Events are sent after a call finishes, with their original timestamps.
Capture and export are separate from the real-time feedback logs.
Bananaflow stores the destination settings for your organization; event
payloads are not stored by Bananaflow.
Exports use bounded retries for retryable row and HTTP errors. BigQuery’s
insertion IDs provide best-effort deduplication; duplicates can still occur.
Capture and pending export queues are bounded, and unsent events can be lost
on a process crash, queue overflow or exhausted retries. There is no durable
replay queue in this version.
Table changes apply to new calls. An active call’s batch is discarded if its
destination is changed or disabled before export, so it cannot accidentally
land in a different table. Credentials are read again before each delivery
attempt to support rotation. Requests already sent to BigQuery cannot be
recalled when you disable export.