> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bidsmith.pro/llms.txt
> Use this file to discover all available pages before exploring further.

# Advanced usage

> Production patterns for retries, streaming, async workloads, and request controls.

## Configure client behavior

```python theme={null}
from aris.client import Aris

client = Aris(
    api_key="sk-aris-your-key",
    timeout=45,
    headers={"X-Environment": "staging"},
)
```

## Retry with exponential backoff

```python theme={null}
import time
from aris.client import Aris
from aris.errors import ArisNodeError

client = Aris(api_key="sk-aris-your-key")

for attempt in range(4):
    try:
        result = client.generate(prompt="Generate release notes from this changelog")
        print(result["output"])
        break
    except ArisNodeError:
        if attempt == 3:
            raise
        time.sleep(2 ** attempt)
```

## Async fan-out requests

```python theme={null}
import asyncio
from aris.client import AsyncAris

client = AsyncAris(api_key="sk-aris-your-key")

async def run(prompt: str):
    return await client.generate(prompt=prompt, max_tokens=120)

async def main():
    prompts = [
        "Summarize incident 1021",
        "Summarize incident 1022",
        "Summarize incident 1023",
    ]
    results = await asyncio.gather(*(run(p) for p in prompts))
    for item in results:
        print(item["output"])

asyncio.run(main())
```

## Runtime guidance

<Tabs>
  <Tab title="Latency sensitive">
    Keep `max_tokens` small, use deterministic prompts, and cap retries to avoid tail latency.
  </Tab>

  <Tab title="Throughput focused">
    Use `AsyncAris`, batch independent calls, and apply concurrency limits at your app layer.
  </Tab>

  <Tab title="Cost constrained">
    Lower output token limits, avoid unnecessary context, and monitor credit burn per endpoint.
  </Tab>
</Tabs>

<Info>
  Pair this guide with [/errors](/errors) to standardize retryable vs non-retryable failures.
</Info>