params

How long the customer can be silent, in milliseconds, after they finish speaking before the sidecar evaluates the conversation. Lower values make the sidecar react faster. If the agent speaks while the customer’s turn is still pending, the sidecar evaluates immediately without waiting. Range: 50-5000.

The minimum time, in milliseconds, between evaluations — a throttle that limits how often the sidecar runs on a busy call. Range: 0-60000.

The maximum number of tool calls the sidecar will chain within a single evaluation before it must produce its advice. Range: 1-20.

The token budget for the sidecar’s running conversation history. When the history grows past this, the oldest messages are dropped. Range: 1000-200000.

Whether actions returned by your tools (such as transferring or hanging up the call) take effect on the call, or are only reported as callbacks.

Whether to generate a closing summary of the sidecar’s session when the call ends. The result is included in the final callback.

Whether to generate an end-of-call summary of the conversation itself, distinct from final_summary (which summarizes the sidecar’s session).

A custom prompt for the end-of-call conversation summary.

The model used for the end-of-call conversation summary, distinct from model (the sidecar’s own model). Suggested values: gpt-4o-mini, gpt-4.1-mini, gpt-4.1-nano.

Whether to emit a callback for each utterance the speech recognizer produces.

Whether each utterance callback includes full speech-recognition detail, such as word timings and alternatives. This increases the callback size, so leave it off unless you need it.

The speech recognition engine to use. Possible values: deepgram, google.

How long, in milliseconds, the recognizer waits before finalizing speech. Defaults to the speech engine’s own default.

The amount of silence, in milliseconds, used to detect the end of speech. Defaults to the speech engine’s own default.

How sensitively the recognizer detects speech. Defaults to the speech engine’s own default.

A bias prompt passed to the speech recognizer to improve accuracy on expected terms, such as product or company names. This is distinct from the operator prompt.

Speech-engine debug verbosity. Range: 0-100.

Whether to enable verbose logging for the sidecar.