> ## Documentation Index
> Fetch the complete documentation index at: https://docs.slng.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Reliability
> Fallbacks, per-turn timeouts, and failure audio that keep a call graceful when a model step fails.
Any of Listen, Think, and Speak can fail on a single turn: a provider returns an
error, or a response arrives too late for a real-time call. The agent's `models`
object carries three controls that handle this, and they work the same way for
all three steps. The API sets them together in the `models` block; the dashboard
exposes them on each section (Listen, Think, Speak), with failure audio under
Agent behavior.
## Prerequisites
Before you begin:
* An agent already exists. If not, see
[Create an agent](/guides/get-started/create-a-project/create-agent).
* These controls live in the `models` object. A `PATCH` replaces it as a unit
rather than merging fields, so send the whole block; see
[Update an agent](/guides/get-started/create-a-project/create-agent#update-an-agent).
## Fallbacks
`models.fallbacks` lists the models to try, in order, when the primary model for
a step fails. Each step has its own list: `stt`, `llm`, and `tts`. A list can mix
`slng/` and BYOK routes, so a hosted model can fall back to your own provider or
the other way around.
Open a section (Listen, Think, or Speak) and find **Fallback routing**. Click
**Add** to append a backup model. The rows are an ordered list, tried top to
bottom, and each row has controls to reorder or remove it. On Speak, each row
also picks a voice for the fallback model.
Send the whole `models` object; a `PATCH` replaces it as a unit.
```json highlight={7-11} theme={null}
{
"models": {
"stt": "deepgram/nova:3",
"llm": "bedrock-mantle/nvidia.nemotron-super-3-120b:latest",
"tts": "slng/fish/tts:s2.1-pro",
"tts_voice": "16cabdb7f8d240569aff36c9e480d783",
"fallbacks": {
"stt": ["soniox/speech-ai:rt-v5"],
"llm": ["groq/openai/gpt-oss-120b:latest"],
"tts": []
}
}
}
```
For how a route resolves to a hosted or BYOK model, see
[Choose a model route](/guides/integrate/pipecat#choose-a-model-route).
## Per-turn timeouts
Each step has a timeout that bounds how long the agent waits before it gives up
on the primary model and moves to a fallback:
* `models.stt_final_timeout_s`: the wait for a final transcript.
* `models.llm_first_token_timeout_s`: the wait for the model's first token.
* `models.tts_first_audio_timeout_s`: the wait for the first audio chunk.
A `null` value uses the provider default. Set a number to fail faster on a slow
step.
Each section's **Advanced settings** card has a timeout field ("STT timeout
(seconds)" on Listen, "LLM timeout (seconds)" on Think, "TTS timeout
(seconds)" on Speak). Leave it empty to use the provider default.
```json highlight={7} theme={null}
{
"models": {
"stt": "deepgram/nova:3",
"llm": "bedrock-mantle/nvidia.nemotron-super-3-120b:latest",
"tts": "slng/fish/tts:s2.1-pro",
"tts_voice": "16cabdb7f8d240569aff36c9e480d783",
"llm_first_token_timeout_s": 5
}
}
```
## Failure audio
`models.failure_audio_enabled` is `true` by default. When a turn cannot be
completed, even after fallbacks, the agent plays a short clip instead of going
silent, so the caller hears that something went wrong. In the dashboard this
toggle lives under **Agent behavior**, not the Listen, Think, or Speak sections.
## Next steps
* [Listen](/guides/agents/configure/listen), [Think](/guides/agents/configure/think),
and [Speak](/guides/agents/configure/speak): the model and voice for each step.