( config: Partial<RealtimeSessionConfig>, )
| 14 | * never sent; the adapter logs when it drops the option. |
| 15 | */ |
| 16 | export function buildSessionUpdate( |
| 17 | config: Partial<RealtimeSessionConfig>, |
| 18 | ): Record<string, unknown> { |
| 19 | // Always enable input audio transcription so user speech is transcribed |
| 20 | const audioInput: Record<string, unknown> = { |
| 21 | transcription: { model: 'whisper-1' }, |
| 22 | } |
| 23 | |
| 24 | if (config.vadMode) { |
| 25 | if (config.vadMode === 'semantic') { |
| 26 | audioInput.turn_detection = { |
| 27 | type: 'semantic_vad', |
| 28 | eagerness: config.semanticEagerness ?? 'medium', |
| 29 | } |
| 30 | } else if (config.vadMode === 'server') { |
| 31 | audioInput.turn_detection = { |
| 32 | type: 'server_vad', |
| 33 | threshold: config.vadConfig?.threshold ?? 0.5, |
| 34 | prefix_padding_ms: config.vadConfig?.prefixPaddingMs ?? 300, |
| 35 | silence_duration_ms: config.vadConfig?.silenceDurationMs ?? 500, |
| 36 | } |
| 37 | } else { |
| 38 | audioInput.turn_detection = null |
| 39 | } |
| 40 | } |
| 41 | |
| 42 | const audio: Record<string, unknown> = { input: audioInput } |
| 43 | |
| 44 | if (config.voice) { |
| 45 | audio.output = { voice: config.voice } |
| 46 | } |
| 47 | |
| 48 | const sessionUpdate: Record<string, unknown> = { |
| 49 | type: 'realtime', |
| 50 | audio, |
| 51 | } |
| 52 | |
| 53 | if (config.instructions) { |
| 54 | sessionUpdate.instructions = config.instructions |
| 55 | } |
| 56 | |
| 57 | if (config.tools !== undefined) { |
| 58 | sessionUpdate.tools = config.tools.map((t) => ({ |
| 59 | type: 'function', |
| 60 | name: t.name, |
| 61 | description: t.description, |
| 62 | parameters: t.inputSchema ?? { type: 'object', properties: {} }, |
| 63 | })) |
| 64 | sessionUpdate.tool_choice = 'auto' |
| 65 | } |
| 66 | |
| 67 | if (config.outputModalities) { |
| 68 | // GA only supports a single output modality: ['audio'] or ['text'] |
| 69 | // (Beta accepted ['audio', 'text']). Audio replies still stream text |
| 70 | // via `response.output_audio_transcript.*` events, so collapsing |
| 71 | // ['audio', 'text'] to ['audio'] preserves the visible behavior. |
| 72 | sessionUpdate.output_modalities = config.outputModalities.includes('audio') |
| 73 | ? ['audio'] |
no outgoing calls
no test coverage detected