AI Models
Achilles is not tied to one AI vendor. Each agent can use a different model, and you can change it at any time without rebuilding the agent. That keeps your investment safe as better and cheaper models appear, and it lets you match the model to the job and to your data policies.
Supported providers
| Provider | What it offers | Choose it when |
|---|---|---|
| Anthropic | Strong reasoning and careful instruction following | Complex or regulated conversations where quality matters most |
| OpenRouter | One account that reaches many hosted models | You want to compare models or keep options open |
| Ollama (cloud or self-hosted) | Open models you can run yourself | Data must stay in your own environment |
| Sarvam | Models specialized for Indian languages | Hindi and regional-language customers |
| Automatic | Uses whichever configured provider is ready, in priority order | You want resilience without picking |
A test mode with canned answers is also available for demos and automated checks. It needs no account and makes no external calls.
How to choose
| If your priority is | Lean toward |
|---|---|
| Best answer quality | A leading commercial model |
| Lowest cost at high volume | A smaller hosted model, plus a cheaper model for simple tasks |
| Data residency | A self-hosted model inside your environment |
| Indian languages | A language specialist, tested against real customer phrasing |
Whatever you pick, run your test cases in the Playground before launch and again after any change. Models differ in tone and tool use, and testing is the way to see it.
Per-agent settings
Set these in the agent's model section in the console.
| Setting | Meaning | Default |
|---|---|---|
| Provider | Which vendor to use, or automatic | Automatic |
| Model | A model name, or provider and model together for explicit routing | Automatic |
| Maximum output length | Cap on how long a reply can be | 1024 tokens |
| Temperature | 0 to 1. Lower is more consistent, higher is more varied | 0.3 |
| Fast model | A cheaper model for routine work | None |
For finance and insurance, keep the temperature low. Consistent answers are easier to test and to defend.
A cheaper model for simple tasks
Not every step needs your best model. The optional fast model handles work where the facts are already known:
- Phrasing the result of a simple lookup, such as an order status.
- Condensing a long conversation so it stays affordable.
- Summarizing a conversation for a human who takes over.
The main model still does the reasoning and decides which abilities to use. In practice this is the most effective single lever for cost, and it does not change what customers are allowed to see.
Data-handling notes
- What the model sees. The conversation, the agent's instructions, and the results your systems return for that request, after masking rules are applied. It does not get direct access to your database. See Personal data.
- Provider terms. When you use a hosted provider, conversation content is sent to that provider under your agreement with them. Review their retention and training terms against your policy before launch.
- Keeping data private. If policy requires it, use a self-hosted model so nothing leaves your environment. See Self-hosting.
- Keys. Provider keys are configured by your administrator on the API and are never shown in the console or stored in an agent.
- Spend visibility. The console overview estimates cost from token use and shows each provider's readiness, so a missing key is visible before customers notice.
Limits to know
- There are no per-tenant spending caps yet. Watch usage in the console overview.
- Failover to another provider happens per request, not in the middle of one.
- Free model tiers usually have hourly and daily limits. Move to a paid plan for production.
Technical settings
| Item | Value |
|---|---|
| Provider options | auto, anthropic, openrouter, ollama, ollama-local, sarvam |
| Explicit routing | provider:model in the model field |
| Definition fields | provider, model, max_tokens, temperature, fast_model |
| Server settings | A provider key per vendor, plus a priority list of models for automatic routing |
| Readiness check | Runtime statistics endpoint in the API reference |
Field details are in the Agent definition reference.