***

title: LLM Settings
description: 'Configure models, parameters, and providers.'
-----------------------------------------------------------

The `OpenAIClient` is the unified interface for calling LLMs. It works with OpenAI by default, but supports any OpenAI-compatible endpoint by changing the `base_url`.

## Basic Configuration

```python
import os
from smallestai.atoms.agent.clients.openai import OpenAIClient

self.llm = OpenAIClient(
    model="gpt-4o-mini",
    temperature=0.7,
    api_key=os.getenv("OPENAI_API_KEY")
)
```

## Parameters

| Parameter     | Type  | Default              | Description                                                           |
| ------------- | ----- | -------------------- | --------------------------------------------------------------------- |
| `model`       | str   | —                    | The model identifier (e.g., `gpt-4o-mini`, `llama-3.1-70b-versatile`) |
| `api_key`     | str   | `OPENAI_API_KEY` env | Your provider's API key                                               |
| `base_url`    | str   | OpenAI               | Custom endpoint URL for other providers                               |
| `temperature` | float | `0.7`                | Controls randomness. Lower = consistent, higher = creative            |
| `max_tokens`  | int   | `1024`               | Maximum tokens in the response                                        |

### Temperature

Controls how "creative" vs "predictable" the model behaves:

* **0.0–0.3**: Consistent, factual. Best for support, FAQ bots.
* **0.4–0.6**: Balanced. Good for general conversation.
* **0.7–1.0**: Creative, varied. Better for sales, engagement.

```python
# Consistent support agent
self.llm = OpenAIClient(model="gpt-4o-mini", temperature=0.3)

# Engaging sales agent
self.llm = OpenAIClient(model="gpt-4o-mini", temperature=0.8)
```

## Using Other Providers

Any provider with an OpenAI-compatible API works by setting `base_url`. This includes Groq, Together.ai, Fireworks, Anyscale, OpenRouter, and Azure OpenAI.

```python
import os
from smallestai.atoms.agent.clients.openai import OpenAIClient

# Example: Using Groq
self.llm = OpenAIClient(
    model="llama-3.1-70b-versatile",
    base_url="https://api.groq.com/openai/v1",
    api_key=os.getenv("GROQ_API_KEY")
)
```

Just swap the `base_url` and `api_key`—your agent code stays the same.

## Streaming

Voice agents **must** use streaming. Without it, users wait for the entire response before hearing anything.

```python
async def generate_response(self):
    response = await self.llm.chat(
        messages=self.context.messages,
        stream=True
    )
    
    async for chunk in response:
        if chunk.content:
            yield chunk.content
```

## Tool Calling

To enable function calling, pass tool schemas:

```python
response = await self.llm.chat(
    messages=self.context.messages,
    stream=True,
    tools=self.tool_registry.get_schemas()
)
```

## Error Handling

```python
from openai import RateLimitError, APIError

async def generate_response(self):
    try:
        response = await self.llm.chat(
            messages=self.context.messages,
            stream=True
        )
        async for chunk in response:
            if chunk.content:
                yield chunk.content
                
    except RateLimitError:
        yield "I'm experiencing high demand. Please try again."
        
    except APIError as e:
        logger.error(f"LLM error: {e}")
        yield "I'm having trouble. Let me try again."
```

***

## Voice Configuration

Agents also need a voice for text-to-speech. **Waves** is our recommended TTS engine—ultra-low latency, optimized for real-time telephony.

### Basic Voice Setup

```python
from smallestai.atoms.models import CreateAgentRequest

agent = CreateAgentRequest(
    name="SupportAgent",
    synthesizer={
        "provider": "smallest",  # Use Waves
        "model": "waves_lightning_v2",
        "voice_id": "zorin",     # Male, professional
        "speed": 1.0
    },
    # ... other config
)
```

### Waves Voice IDs

| Voice ID | Description                      |
| -------- | -------------------------------- |
| `zorin`  | Male, professional (recommended) |
| `emily`  | Female, warm                     |
| `raj`    | Male, Indian English             |
| `aria`   | Female, neutral                  |

<Tip>
  For the complete Waves voice library with audio samples:
  → [Waves Voice Models](https://waves-docs.smallest.ai/content/getting-started/models)
</Tip>

### Voice Cloning

Create custom voices from audio samples:
→ [Waves Voice Cloning Guide](https://waves-docs.smallest.ai/content/voice-cloning/types-of-clone)

### Third-Party Providers

OpenAI and ElevenLabs voices are also supported. Set `provider` to `openai` or `elevenlabs` and use their respective voice IDs.

***

## Tips

<AccordionGroup>
  <Accordion title="Keep max_tokens low for voice">
    Set `max_tokens` to 100-200. Shorter responses mean faster audio playback. Guide conciseness in your prompt too.
  </Accordion>

  <Accordion title="Warm up connections on start">
    The first LLM call has connection overhead. Send a tiny request in `start()` to warm up before the user speaks.
  </Accordion>

  <Accordion title="Use fallbacks for reliability">
    Configure a backup provider (e.g., Groq primary, OpenAI fallback) to handle rate limits or outages gracefully.
  </Accordion>
</AccordionGroup>