Sarvam AI¶

TraceVerde provides full instrumentation for Sarvam AI, India's sovereign AI platform for multi-modal Indian language processing.

Installation¶

pip install genai-otel-instrument[sarvamai]
pip install sarvamai

Quick Start¶

import genai_otel
genai_otel.instrument(service_name="sarvam-app")

from sarvamai import SarvamAI

client = SarvamAI(api_subscription_key="your_key")

# Chat completion with sarvam-m
response = client.chat.completions.create(
    model="sarvam-m",
    messages=[{"role": "user", "content": "What is AI?"}],
)
print(response.choices[0].message.content)
# Traces, tokens, and costs captured automatically

Supported Models and APIs¶

Chat (sarvam-m)¶

Sarvam's multilingual chat model with support for Indian languages.

response = client.chat.completions.create(
    model="sarvam-m",
    messages=[
        {"role": "system", "content": "You are a helpful assistant who speaks Hindi."},
        {"role": "user", "content": "Tell me about India's space program"},
    ],
)

Translation (Mayura v1)¶

Translation across 22+ Indian languages with speaker gender adaptation.

translation = client.text.translate(
    input="Hello, how are you?",
    source_language_code="en-IN",
    target_language_code="hi-IN",
    model="mayura:v1",
    mode="classic-formal",
    speaker_gender="female",
)
print(translation.translated_text)

Features:

22+ Indian languages
mayura:v1 (latest) and sarvam-translate:v1 (legacy)
Translation modes: classic-formal, classic-colloquial
Speaker gender adaptation
Numerals format control (international vs native)

See Mayura translation example.

Text-to-Speech (Bulbul v3)¶

High-quality TTS with 48 speakers across 11 Indian languages.

tts_response = client.text.speech(
    input="Namaste, aap kaise hain?",
    target_language_code="hi-IN",
    model="bulbul:v2",
    speaker="meera",
    pace=1.0,
    pitch=0,
    loudness=1.5,
)
# tts_response.audios contains base64-encoded WAV audio

Bulbul v3 features:

48 speakers (male and female voices)
11 Indian languages
Pace control (0.5 to 2.0)
Temperature control for expressiveness
24000 Hz sample rate
Max 2500 characters per request

See Bulbul v3 TTS example.

Speech-to-Text (Saarika)¶

Automatic speech recognition for Indian languages.

with open("audio.wav", "rb") as f:
    transcript = client.speech.text(
        file=f,
        model="saarika:v2",
        language_code="hi-IN",
    )
print(transcript.transcript)

Language Detection¶

Identify the language of input text.

detection = client.text.detect_language(
    input="Namaste, kaise ho?",
)
print(detection.language_code)  # "hi-IN"

Multilingual Pipeline Example¶

A full production-style pipeline combining multiple Sarvam APIs:

Customer Query (any Indian language)
  +-- Language Detection
  +-- Translate to English
  +-- sarvam-m Chat (generate response)
  +-- Translate response back to customer's language
  +-- Bulbul v3 TTS (voice response)

All steps are automatically traced with cost breakdown per API call.

See multilingual pipeline example.

Arya Agent Orchestration¶

Simulates a Sarvam Arya-style agentic workflow with:

Voice/text input in any Indian language
Multi-step reasoning pipeline
Tool usage (translate, detect, search, synthesize)
Immutable conversation state
Localized voice+text output

See Arya agent example.

Cost Tracking¶

TraceVerde includes pricing for all Sarvam AI models:

sarvam-m: Chat model with per-token pricing
Saarika/Saaras: STT models with per-second pricing
Bulbul: TTS with per-character pricing
Mayura/Sarvam Translate: Translation with per-token pricing
Vision: Image understanding models

12+ models tracked in the pricing database.

Environment Setup¶

export SARVAM_API_KEY=your_sarvam_api_key
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318
export OTEL_SERVICE_NAME=sarvam-app

All Examples¶

Example	Description
Basic usage	Chat, translate, TTS, STT, language detection
Arya agent	Agent orchestration pipeline
Bulbul v3 TTS	Text-to-speech with 48 speakers
Mayura translate	Translation with modes and gender adaptation
Multilingual pipeline	Full customer support pipeline