Use Sarvam AI as Your Model¶

Add the built-in Sarvam AI provider, then chat with an agent in Tamil or another Indian language. Sarvam models are built for Indian languages. You paste your key and pick a model. AgentZ fills in the URL and the authentication.

The announcement banner for Sarvam AI in AgentZ.

What You Need¶

A Ready workspace. See Create a workspace.
A Sarvam AI API key from sarvam.ai.

Note

Sarvam AI handles text only in AgentZ. To read images or scanned PDF pages, add a second model as the Attachment model. See Inference providers.

Add the Sarvam AI Provider¶

In the workspace sidebar, open Workspace settings → Inference → Inference providers. The page Inference providers opens.
Select Add inference provider. The sheet Add inference provider opens.
Enter a Display name, for example Sarvam AI.
Open the Provider list and choose Sarvam AI. The label OpenAI-compatible appears next to the name. The fields Base URL and Authentication fill themselves in.
Paste your key in the API key field under Credentials.
In the Models field, choose a Sarvam model.
Select Add provider.

The new provider appears in the table on the Inference providers page. Its Status shows Accepted, then Ready.

The Add inference provider sheet with Display name Sarvam AI, Provider Sarvam AI marked OpenAI-compatible, Base URL https://api.sarvam.ai/v1, Authentication set to API key, a hidden API key and the model Sarvam-105B — The sheet with Sarvam AI chosen. The API key shows as dots.

Create a Sandbox That Uses the Model¶

Follow Create a sandbox. On the Models step, open your Sarvam AI provider and switch on your model. The first model you switch on becomes the Default model.

The message Sandbox created appears.

Create an Agent¶

Follow Create an agent. In Sandbox, choose the sandbox from the last step.

The agent shows Progressing, then Idle. Idle means the agent is ready.

Chat in Tamil or Another Indian Language¶

Select New chat. If you have more than one agent, open Choose agent and select yours.
Click the prompt box and type a short greeting in Tamil:
```
வணக்கம்
```
Select Send message.

Your message appears in the chat. The agent writes its reply below it.

See the Rural Health Triage Demo¶

The AgentZ team built a demo on Sarvam AI. The workflow rural-health-triage takes clinic records and sorts patient complaints into urgent, routine or pharmacy-only. Some test prompts, written in mixed languages, try to force a high priority. The workflow sends those cases to human review.

You build a workflow like this by asking your agent in the chat. See Workflows.

The Workflows page with the agent billy and the workflow rural-health-triage selected, and a graph with the steps triage-classify, safety-override-detect, route-human-review, route-normal, output-structured and publish-report — The graph of the workflow rural-health-triage. Edge labels include override and clean.

The same agent built a dashboard named PHC Triage Metrics for the cases.

The Dashboards page with the agent billy and the dashboard PHC Triage Metrics, showing a Triage History chart, a Recent Cases table with complaints in Tamil script and a Triage Count gauge with the value 4 — The dashboard PHC Triage Metrics. The Complaint column holds Tamil text.

If It Does Not Work¶

The message Model catalog unavailable appears. Type the model ID in the Models field. The sheet above lists sarvam-105b for Sarvam-105B.
The Status of the provider shows Degraded. Hover over the badge. The tooltip shows the failing message.

Next Step¶

Dashboards