buildStepBody($provider, $model, $instructions, $messages, $tools, $schema, $options, $stepContext); $response = $this->withErrorHandling( $provider->name(), fn () => $this->client($provider, $timeout)->post("models/{$model}:generateContent", $body), ); $data = $response->json(); $this->validateTextResponse($data); return $this->parseTextResponse($data, $provider, $model, filled($schema))->withRawResponse($response); } /** * {@inheritdoc} */ public function generateStreamStep( string $invocationId, TextProvider $provider, string $model, ?string $instructions, array $messages, array $tools, ?array $schema, ?TextGenerationOptions $options, ?int $timeout, StepContext $stepContext, ): Generator { $body = $this->buildStepBody($provider, $model, $instructions, $messages, $tools, $schema, $options, $stepContext); $response = $this->withErrorHandling( $provider->name(), fn () => $this->client($provider, $timeout) ->withOptions(['stream' => true]) ->post("models/{$model}:streamGenerateContent?alt=sse", $body), ); return yield from $this->processTextStream($invocationId, $provider, $model, $response->getBody()); } /** * Generate an image. * * @param array $attachments * @param '3:2'|'2:3'|'1:1'|null $size * @param 'low'|'medium'|'high'|null $quality */ public function generateImage( ImageProvider $provider, string $model, string $prompt, array $attachments = [], ?string $size = null, ?string $quality = null, ?int $timeout = null, ): ImageResponse { $parts = [['text' => $prompt]]; if (filled($attachments)) { $parts = array_merge($parts, $this->mapAttachments(collect($attachments))); } $imageOptions = $provider->defaultImageOptions($size, $quality); $body = [ 'contents' => [['role' => 'user', 'parts' => $parts]], 'generationConfig' => array_filter([ 'responseModalities' => ['IMAGE', 'TEXT'], 'imageConfig' => array_filter([ 'imageSize' => $imageOptions['image_size'] ?? null, 'aspectRatio' => $imageOptions['aspect_ratio'] ?? null, ]), ]), ]; $response = $this->withErrorHandling( $provider->name(), fn () => $this->client($provider, $timeout ?? 120)->post("models/{$model}:generateContent", $body), ); $data = $response->json(); $images = (new Collection($data['candidates'][0]['content']['parts'] ?? [])) ->filter(fn ($part): bool => isset($part['inlineData'])) ->values() ->map(fn ($part): GeneratedImage => new GeneratedImage( $part['inlineData']['data'], $part['inlineData']['mimeType'], )); return new ImageResponse( $images, $this->extractUsage($data), new Meta($provider->name(), $model), ); } /** * {@inheritdoc} */ public function generateEmbeddings( EmbeddingProvider $provider, string $model, array $inputs, int $dimensions, int $timeout = 30, array $providerOptions = [], ): EmbeddingsResponse { $model = $this->normalizeEmbeddingModel($model); $requests = array_map(fn (mixed $input): array => array_merge(Arr::except($providerOptions, 'output_dimensionality'), [ 'model' => "models/{$model}", 'content' => ['parts' => [$this->mapEmbeddingInput($input)]], 'outputDimensionality' => $dimensions, ]), $inputs); $response = $this->withErrorHandling( $provider->name(), fn () => $this->client($provider, $timeout)->post("models/{$model}:batchEmbedContents", [ 'requests' => $requests, ]), ); $data = $response->json(); return new EmbeddingsResponse( (new Collection($data['embeddings'] ?? []))->pluck('values')->all(), $data['usageMetadata']['promptTokenCount'] ?? 0, new Meta($provider->name(), $model), ); } /** * Generate audio from the given text. * * @throws RuntimeException if Gemini returns no audio data or invalid base64 audio. */ public function generateAudio( AudioProvider $provider, string $model, string $text, string $voice, ?string $instructions = null, int $timeout = 30, ): AudioResponse { $response = $this->withErrorHandling( $provider->name(), fn () => $this->client($provider, $timeout)->post("models/{$model}:generateContent", [ 'contents' => [[ 'role' => 'user', 'parts' => [[ 'text' => $instructions !== null && trim($instructions) !== '' ? trim($instructions)."\n\n".$text : $text, ]], ]], 'generationConfig' => [ 'responseModalities' => ['AUDIO'], 'speechConfig' => [ 'voiceConfig' => [ 'prebuiltVoiceConfig' => [ 'voiceName' => match ($voice) { 'default-female' => 'Kore', 'default-male' => 'Puck', default => $voice, }, ], ], ], ], ]), ); $data = $response->json(); $encodedAudio = $data['candidates'][0]['content']['parts'][0]['inlineData']['data'] ?? null; if (! is_string($encodedAudio) || $encodedAudio === '') { throw new RuntimeException('No audio data received from Gemini API.'); } $pcm = base64_decode($encodedAudio, true); if ($pcm === false) { throw new RuntimeException('Gemini returned invalid audio data.'); } return new AudioResponse( base64_encode($this->pcmToWav($pcm)), new Meta($provider->name(), $model), 'audio/wav', ); } /** * Generate text from the given audio. * * @param array $providerOptions */ public function generateTranscription( TranscriptionProvider $provider, string $model, TranscribableAudio $audio, ?string $language = null, bool $diarize = false, int $timeout = 30, array $providerOptions = [], ): TranscriptionResponse { $inlineData = ['inlineData' => [ 'mimeType' => $audio->mimeType() ?? 'audio/mp3', 'data' => base64_encode($audio->content()), ]]; if ($diarize) { $prompt = $language !== null ? "Transcribe this audio with timestamps in {$language}. Return the full transcript and a list of segments. Use MM:SS or HH:MM:SS timestamps, with optional fractional seconds, for start_time and end_time." : 'Transcribe this audio with timestamps. Return the full transcript and a list of segments. Use MM:SS or HH:MM:SS timestamps, with optional fractional seconds, for start_time and end_time.'; $response = $this->withErrorHandling( $provider->name(), fn () => $this->client($provider, $timeout)->post("models/{$model}:generateContent", array_merge($providerOptions, [ 'contents' => [[ 'parts' => [['text' => $prompt], $inlineData], ]], 'generationConfig' => [ 'responseMimeType' => 'application/json', 'responseSchema' => [ 'type' => 'OBJECT', 'properties' => [ 'transcript' => ['type' => 'STRING'], 'segments' => [ 'type' => 'ARRAY', 'items' => [ 'type' => 'OBJECT', 'properties' => [ 'text' => ['type' => 'STRING'], 'start_time' => ['type' => 'STRING'], 'end_time' => ['type' => 'STRING'], ], 'required' => ['text', 'start_time', 'end_time'], ], ], ], 'required' => ['transcript', 'segments'], ], ], ])), ); $data = json_decode($response->json('candidates.0.content.parts.0.text') ?? '{}', true); $text = $data['transcript'] ?? ''; $segments = (new Collection($data['segments'] ?? []))->map(fn ($seg): TranscriptionSegment => new TranscriptionSegment( $seg['text'], '', $this->timestampToSeconds($seg['start_time'] ?? '0:00'), $this->timestampToSeconds($seg['end_time'] ?? '0:00'), )); } else { $prompt = $language !== null ? "Transcribe this audio. Output only the transcription in {$language}." : 'Transcribe this audio. Output only the transcription text.'; $response = $this->withErrorHandling( $provider->name(), fn () => $this->client($provider, $timeout)->post("models/{$model}:generateContent", array_merge($providerOptions, [ 'contents' => [[ 'parts' => [['text' => $prompt], $inlineData], ]], ])), ); $text = $response->json('candidates.0.content.parts.0.text') ?? ''; $segments = new Collection; } $usageMeta = $response->json('usageMetadata') ?? []; return new TranscriptionResponse( trim((string) $text), $segments, new Usage( promptTokens: $usageMeta['promptTokenCount'] ?? 0, completionTokens: $usageMeta['candidatesTokenCount'] ?? 0, reasoningTokens: $usageMeta['thoughtsTokenCount'] ?? 0, ), new Meta($provider->name(), $model), ); } /** * Convert a timestamp string to seconds. */ protected function timestampToSeconds(string $timestamp): float { $timestamp = str_replace(',', '.', trim($timestamp)); if (preg_match('/^\d+(?:\.\d+)?$/', $timestamp) === 1) { return (float) $timestamp; } if (preg_match('/^\d+(?::\d+){1,2}(?:\.\d+)?$/', $timestamp) !== 1) { return 0.0; } $parts = array_reverse(explode(':', $timestamp)); return (float) $parts[0] + ((float) ($parts[1] ?? 0)) * 60 + ((float) ($parts[2] ?? 0)) * 3600; } }