Advanced agent settings (AI model, speech, keypad)

• Change one setting at a time. After each change, click Save Settings and try the AI agent with Speak with Agent or Chat with Agent. See Trying your AI agent yourself.
• For a bigger change, such as a new AI model, try it on a copy of the agent first. See Copying or deleting an AI agent.
• Write down the current value before you change it. AlloMia doesn’t keep a history of settings: to undo a change, put the old value back and click Save Settings.
• Calls that start after you save use the new settings.
• If the agent answers as part of a team, the team’s Team Overrides can replace its AI model and speech settings. See Using several AI agents together (Teams).

1. In the menu on the left, click Voice Agents, then click your AI agent in the My Agents list.
2. On the Voice Agent tab, find Model Configuration.
3. Under LLM, choose OpenAI or Google.
4. Under Model, choose a model: GPT-4.1-mini, GPT-5, GPT-5-mini, GPT-5-nano, GPT-5.4-mini, GPT-5.4-nano or GPT-5.1 for OpenAI, or Gemma 4 31B for Google.
5. Click Save Settings.

New AI agents start with Gemma 4 31B. Models differ in how fast they answer and how closely they follow instructions. After a change, ask the questions your patients ask most, and check that the agent still follows its instructions and still uses its actions, such as passing calls to your staff.

Under Response Configuration, on the same tab:

• Max Tokens limits the length of each answer. New agents start at 150, and the hint on the page suggests 120 to 180. If answers stop in the middle of a sentence, raise it a little.
• Temperature is a slider from 0 to 1. Higher values give more varied but less focused answers. New agents start at 0.5.

Temperature has no effect with GPT-5 models. GPT-5, GPT-5-mini, GPT-5-nano, GPT-5.4-mini, GPT-5.4-nano and GPT-5.1 ignore it. Only GPT-4.1-mini and Gemma 4 31B use it.

1. On the Voice Agent tab, click the gear button next to Primary Language. The Speech Recognition Settings window opens, with a Basic tab and an Advanced tab.
2. Change the settings you need (see below).
3. Close the window, then click Save Settings.

• Provider, Model and Language: leave them as they are. Choosing the Primary Language sets them.
• Important Keywords: words the agent often mishears, such as your clinic’s name, doctors’ names or street names. Separate them with spaces: each word counts on its own. This box is for English agents only. It doesn’t show for French agents and doesn’t work in multilingual mode.
• Noise Reduction: removes background noise from the caller’s voice. Try it if many patients call from noisy places.
• Endpointing: how long, in milliseconds, the agent waits after the caller stops talking before it takes its turn. It works only when Model is Nova-2 or Nova-3, not with Flux-General.
• Minimum Global Endpointing, on the Advanced tab once you turn on Enable Advanced Endpointing: the shortest pause, from 10 to 500 milliseconds, before the agent answers. It starts at 300 and works for every language. Raise it if the agent cuts off patients who pause, for example while reading out a number. Lower it if the agent is slow to answer.
• French agents show Phrase Detection, Pause Detection and Maximum Phrase Length instead of Endpointing. Pause Detection is how long a silence must last to end what the caller is saying. It doesn’t apply when Detection Method is Content-based.

On the agent’s Tools & Actions tab, Keypad Input lets the agent read digits the caller presses, such as a file number. The steps to turn it on are in Turning actions on or off for your AI agent. What happens during a call:

• AlloMia collects the digits, then hands them to the agent as one message, such as [DTMF Input: 1234].
• It sends them when the caller has stopped pressing keys for the Timeout (0 to 10 seconds, 2 by default), or as soon as the caller presses a Delimiter key you turned on (# or *). The delimiter itself isn’t sent.
• With a Timeout of 0, the digits are sent only when the caller presses a delimiter, so turn one on.
• The agent’s instructions must say when to ask for the digits and what to do with them.
• Test it with a real phone call: Speak with Agent can’t test it.

Enable Dial Keypad, on the Call Control tab, is a different setting. It doesn’t do anything today.

These settings are on screen, but changing them has no effect on calls for now. Leave them as they are:

• Speech Quality Filter, and the Predefined Rules and Custom Rules under Advanced Endpointing, in Speech Recognition Settings.
• Enable Dial Keypad, Punctuation Boundaries and Voicemail Detection (Outbound), on the Call Control tab.
• Voicemail Message, on the Conversation tab.
• Volume, in the voice settings window that opens from the gear button next to Voice Settings.

• Answers stop in the middle of a sentence: raise Max Tokens a little.
• A Temperature change makes no difference: with a GPT-5 model, Temperature is ignored. Change the instructions instead, or try another model.
• The agent cuts off patients who pause: raise Minimum Global Endpointing a little.
• The agent mishears names: add them to Important Keywords (English agents).
• Keypad digits are ignored: check that Keypad Input is on and saved, that the instructions ask for the digits, and test from a real phone.
• The agent works less well after a change: put the old value back and click Save Settings.