Sarvam AI for the Laravel AI SDK

A Sarvam AI provider for the Laravel AI SDK. It adds chat, text-to-speech and speech-to-text in Indian languages, so you can use the AI SDK's standard agent(), Audio and Transcription APIs with provider: 'sarvam'.

Installation

composer require dwivedianuj9118/laravel-ai-sarvam

Add a provider to your application's config/ai.php and set SARVAM_API_KEY in your .env file:

'providers' => [
    // ...

    'sarvam' => [
        'driver' => 'sarvam',
        'key' => env('SARVAM_API_KEY'),
        'language' => 'hi-IN', // Default language for text-to-speech...
    ],
],

Usage

Chat

use function Laravel\Ai\agent;

$response = agent()->prompt('आज का पंचांग संक्षेप में बताइए।', provider: 'sarvam');

echo $response->text;

Agents, tools, streaming and structured output work the same way as with other providers, because Sarvam's chat API follows the OpenAI chat completions format.

Text-to-speech

use Laravel\Ai\Audio;

$audio = Audio::of('शुभ प्रभात! आपका दिन मंगलमय हो।')->generate(provider: 'sarvam');

Storage::put('greeting.wav', base64_decode($audio->audio));

female() and male() choose a default speaker for the model (priya / shubh for bulbul:v3, anushka / abhilash for bulbul:v2). Pass any Sarvam speaker with voice():

Audio::of('नमस्ते')->voice('kavya')->generate(provider: 'sarvam');

Use provider options for another language or any other Sarvam field:

Audio::of('வணக்கம்')
    ->withProviderOptions([
        'language_code' => 'ta-IN',
        'pace' => 1.1,
        'output_audio_codec' => 'mp3',
    ])
    ->generate(provider: 'sarvam');

Speech-to-text

use Laravel\Ai\Transcription;

$transcript = Transcription::fromPath(storage_path('call.mp3'))
    ->language('hi') // Optional. Normalized to "hi-IN"; Sarvam detects the language when omitted.
    ->generate(provider: 'sarvam');

echo $transcript->text;

Calling diarize() requests Sarvam's timestamps and returns them as segments. Sarvam's REST endpoint does not identify speakers, so each segment's speaker is empty.

Sarvam's REST speech-to-text endpoint is meant for audio under 30 seconds. Use Sarvam's batch API for longer recordings.

Configuration

'sarvam' => [
    'driver' => 'sarvam',
    'key' => env('SARVAM_API_KEY'),
    'url' => env('SARVAM_URL'), // Optional API root, defaults to https://api.sarvam.ai
    'language' => 'hi-IN',
    'reasoning_effort' => 'low', // Optional: low, medium, high, max, or null to disable reasoning
    'max_tokens' => 4096, // Optional default token limit for chat
    'models' => [
        'text' => ['default' => 'sarvam-105b', 'cheapest' => 'sarvam-30b'],
        'audio' => ['default' => 'bulbul:v3'],
        'transcription' => ['default' => 'saaras:v4'],
    ],
    'voices' => [
        'bulbul:v3' => ['default-male' => 'shubh', 'default-female' => 'priya'],
    ],
],

Sarvam updates its models often. Check Sarvam's model list for the latest names.

Reasoning and empty replies

sarvam-105b reasons before it answers, and the reasoning counts towards max_tokens (Sarvam's default is 2048). If a long reasoning step uses the whole limit, the reply text is empty. Set a lower reasoning_effort, disable reasoning with null, or raise max_tokens in the provider config. A token limit set on the agent itself takes precedence over max_tokens.

Testing

composer test

License

MIT. See LICENSE.md.