Sarvam AI

Overview

The Sarvam AI provider combines its OpenAI-compatible chat API with native speech endpoints. Use the sarvam/ model prefix. DeepIntShield sends the configured provider credential using both authentication forms expected by Sarvam, while callers authenticate only with a DeepIntShield Virtual Key.

Supported operations

Operation	Non-streaming	Streaming	DeepIntShield route
List models	✅	—	GET /v1/models
Chat Completions	✅	✅ SSE	POST /v1/chat/completions
Text-to-speech	✅	✅ audio chunks	POST /v1/audio/speech
Speech-to-text	✅	❌	POST /v1/audio/transcriptions

Responses, embeddings, files, batches, image operations, and Realtime are rejected locally as unsupported.

Configure Sarvam AI

1. Create an API subscription key in the Sarvam dashboard.
2. Open Workspace → AI Providers, select Sarvam AI, and add the key.
3. Optionally restrict the key to approved chat and speech models.
4. Use network_config.base_url only for an approved private endpoint or proxy; otherwise the adapter uses https://api.sarvam.ai.

Chat Completions

curl https://app.deepintshield.com/v1/chat/completions \
  -H "Authorization: Bearer $DEEPINTSHIELD_VIRTUAL_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sarvam/sarvam-105b",
    "messages": [{"role": "user", "content": "Answer in Hindi and English."}],
    "reasoning_effort": "medium",
    "stream": true
  }'

The chat adapter supports the parameters advertised by the live model catalog, including temperature, top-p, reasoning effort, token limits, stop sequences, tools, tool choice, and structured response formats. Streaming uses SSE and does not perform model discovery before the request.

Speech

Text-to-speech uses the unified speech route and supports streaming when the selected Sarvam speech model supports it. Transcription is multipart form data:

curl https://app.deepintshield.com/v1/audio/transcriptions \
  -H "Authorization: Bearer $DEEPINTSHIELD_VIRTUAL_KEY" \
  -F "model=sarvam/saaras:v3" \
  -F "file=@meeting.wav" \
  -F "language_code=hi-IN"

Note

Model names, languages, voices, formats, and request limits differ across Sarvam speech products. Validate the selected model in the live catalog and Sarvam’s current API reference.

Discovery, pricing, and errors

GET /v1/models?provider=sarvam fetches Sarvam’s live model catalog and applies the configured key’s allow-list. Provider prices and availability can change; record contracted rates through DeepIntShield pricing overrides.

Sarvam error envelopes are normalized into DeepIntShield errors with provider, model, request type, and safe upstream headers. Retry only normalized transient errors such as throttling or service unavailability.

See Sarvam’s official Chat Completions reference and model catalog.