Skip to content

AI Model ​

Asha is setting up a support agent for Acme Diagnostics. It will answer web chat in English and take phone calls in Hindi and Tamil. She wants chat replies to be cheap and fast, and she wants the phone agent to sound like a warm, female Indian voice. On Perfox she sets both on one card: the AI Model sub-node attached to her AI Agent.

This page covers everything on that card: which model runs each job, the two ways a voice agent can be built, how to pick a voice and how the agent refers to itself, what callers hear while they wait, and the rules for how a voice agent treats callers. If you haven't placed an AI Agent yet, start with AI Agent Node.

What the AI Model sub-node is ​

Every AI Agent needs exactly one AI Model sub-node, connected to the agent's AI Model port (marked with a red asterisk because it is required). The canvas won't let you connect a second one, or a different kind of node, to that port. The AI models are included and managed by Perfox. You never enter a model key here. What you choose is which model does which job, plus a few settings for each.

The panel is split into one section per job:

SectionWhat it controls
Chat agent modelThe model that writes replies on web chat, WhatsApp, SMS and email.
Voice agent modelHow the agent listens and speaks on web voice and phone calls, including the voice itself. It also holds While the caller waits (what a caller hears during a lookup) and Caller rules (keypad input, misheard turns, quiet callers, an AI disclosure line, and following the caller's language).
How the agent refers to itselfThe agent's grammatical gender. This section only appears when no chosen voice already implies a gender.
Document extraction modelThe model used when the agent reads a document it's been given.
Image / vision modelThe model used to understand images a customer sends.

Every model picker starts on Platform default. Leave it there and Perfox uses its recommended model for that job, and your agent picks up improvements to that default automatically. Pick a specific model only when you have a reason to.

To open it: go to Build → Agents, open your agent, and click the violet AI Model card (lightbulb icon) under the AI Agent. The card's subtitle summarises your choices, for example Gemini 3.1 Flash Lite · voice: live · t=0: the chat model, the voice mode, and the creativity setting. The panel has Config, Settings and Notes tabs.

The AI Model panel: Chat agent model, Voice agent model (Live mode) with a voice picker, the "While the caller waits" cue, and the start of Caller rules.

Chat agent model ​

SettingWhat it doesDefault
ModelThe model that writes text replies. Each option shows a short quality note, a speed note and a cost band ($ Cheapest, $$ Balanced, $$$ Premium) so you can trade quality against cost.Platform default
CreativityHow varied replies are: 0 (focused and consistent) to 2 (varied), in steps of 0.1.0.0
Max reply lengthUpper limit on the length of each reply, from 64 to 8192 in steps of 64.2048
ThinkingHow much the model reasons before it answers. Only shown for models that support it, and only lists the levels that model accepts (for example Minimal / Low / Medium / High).Platform default

Each model only shows the settings it actually supports. For example, a model that has no reasoning control shows no Thinking row. Changing a model re-checks that job's settings against the new model, so a setting the new model doesn't accept is cleared instead of being silently ignored.

Each picker lists every model that can do that job, whoever makes it, so choosing a Sarvam voice doesn't limit which chat models you can pick. The list comes from Perfox's model catalogue and what's enabled for your workspace, so it changes over time as models are added or retired. A recent addition is Gemini 3.8 Flash, the strongest Flash model for long, multi-step agentic work. Its Thinking row offers Low / Medium / High. The platform default model hasn't changed.

A model that is no longer offered fails loudly

If a chat model you picked earlier is no longer offered, that agent's replies fail with an error that names the model. Perfox never swaps in a different model behind your back. Open the AI Model card and pick a model from the current list.

Voice agent model ​

The Mode setting decides how the voice agent is built:

ModeHow it worksWhen to use it
Live — one speech-to-speech model (default)One model hears the caller and speaks the reply directly.The lowest latency and the most natural turn-taking. A good default.
Pipeline — speech-to-text → LLM → text-to-speechThree separate parts: one transcribes the caller, a voice brain writes the reply, and a text-to-speech voice speaks it.When you want a specialist voice or transcription, for example Indian-language speech from Sarvam, or a premium ElevenLabs voice.

In Live mode you choose:

  • Model: the speech-to-speech model, such as Gemini Live or OpenAI Realtime.
  • Voice: a voice from that model's own catalogue. The list only offers voices the chosen model can speak, and each gendered voice is labelled, for example Aoede (female) or Charon (male).

In Pipeline mode you choose each part separately:

  • Speech-to-text: for example Sarvam Saaras (v3 or v4) or ElevenLabs Scribe. Under it sits an optional Advanced — brand words block (see Brand words).
  • Voice brain: the model that writes the spoken reply, with its own Creativity and Max reply length.
  • Text-to-speech: for example Sarvam Bulbul (v3 or v4-flash) or ElevenLabs, plus the Voice to speak with.

The speech-to-text and text-to-speech lists only show models Perfox can run live on a call.

The Voice agent model section in Pipeline mode: Speech-to-text, Voice brain, and Text-to-speech, each with its own model and settings.

Language comes from the Personality

The voice agent's language isn't set here. The Personality sub-node's Language, Region and Language behavior decide what language the agent listens and speaks in. With Auto-detect, a phone call can switch language mid-call when the caller does.

Brand words ​

Speech-to-text sometimes mishears product and brand names. On a Pipeline agent, open Advanced — brand words under Speech-to-text. When the block is collapsed, its label shows how many words you've added.

  • Use brand words: On or Off. Turning it off keeps your list but transcribes speech exactly as heard.
  • Brand words: a comma-separated list, for example Fascino, Aerox, RayZR. Up to 50 words, 64 characters each.

One list does two jobs. Speech-to-text models that support it (Saaras v4) listen out for these words. On every model, a near match is corrected to your spelling before the agent reads it. Under the list, From your knowledge base: offers names found in your knowledge base as chips. Nothing is added until you click a chip.

Voice and gender ​

The voice you pick also tells Perfox how the agent should refer to itself. That matters in languages where verbs and adjectives change with the speaker's gender, such as Hindi, Marathi and Punjabi. Under the voice picker you'll see one of these notes:

  • "Agent refers to itself as female, from this voice." Gender is set and matches the voice. It also applies on WhatsApp, email and web chat, where there's no voice.
  • "No gendered wording is set. This voice is female — use female." Nothing is stored yet. Click use female (or use male) to set it.
  • "Agent refers to itself as male, which does not match this female voice." The two disagree. Pick the voice again to bring them back in line.

Picking a voice from the list sets the agent's gender to match it. Simply opening the panel never changes anything.

When the chosen voice doesn't imply a gender, for example a text-only agent or an OpenAI Realtime voice, a separate How the agent refers to itself section appears instead, with a Gender menu: Not set — no gendered wording, Female, Male, or Non-binary.

If you don't pick a voice, the agent uses the model's default voice. When the agent's gender is set, the default voice follows that gender, so moving an agent from Live to Pipeline won't change how it sounds.

Sarvam Bulbul v4-flash offers 224 voices, grouped by language in the list. Some have no gender label because Sarvam doesn't state one. If you switch the text-to-speech model between Bulbul v3 and v4, the agent keeps the same speaker.

While the caller waits ​

AI looking up controls what a phone or web-voice caller hears while the agent runs a tool or looks something up. The default is Off (silence). The other choices are Periodic tone, one of three hold-music tracks (Music — Calm, Music — Bright or Music — Soft chimes), and Spoken.

Spoken works only when the voice Mode is Pipeline. Before a slow lookup, the agent says one short line such as "One moment, let me check" in the caller's language, instead of a tone. In Live mode it appears greyed out as Spoken — Pipeline voice only.

The three sounds about people (Connecting to a person, On hold by an operator and While the team is being rung) are set on the agent's Operator node, under Routing → Waiting sounds. If you set one of them on this card before they moved, that value still applies until the Operator node sets its own.

Caller rules ​

At the bottom of the Voice agent model section, Caller rules decide how the voice agent handles callers. They apply to every voice call on this agent: phone, web voice and SIP. Each rule starts with the default shown.

The Caller rules block: Keypad input with an optional key menu, Not understood, Quiet caller, End quiet calls, AI disclosure and Caller language.

RuleWhat you setDefaultWhat it does
Keypad inputLet callers press keys, plus an optional Key menuOnPhone and SIP calls only (web voice has no keypad). A key the caller presses reaches the agent as their answer ("the caller pressed 2"). A number typed on the keypad is delivered when they press # or stop typing. In the Key menu, map a key (0–9, *, #) to what pressing it means, for example 1 → "I want to check my order". The agent treats it as if the caller said it. Keys are ignored while a person has the call, during hold music, and during a voicemail.
Not understoodHand off after 1–5 failed turns in a rowOn, 2Pipeline voice only. A failed turn is one where nothing the caller said was recognised, or the agent produced no answer. The agent then hands the caller to a person through this agent's Operator node. With no Operator node it apologises and ends the call.
Quiet callerCheck in after 12–120 seconds of silenceOn, 15 secondsEvery voice call. When the caller says nothing, the agent asks once whether they're still there, in the language of the call. Silence isn't counted while the caller waits for or talks to a person, hears hold music, leaves a voicemail, or while the agent is looking something up.
End quiet callsafter 1–3 unanswered check-ins (shown when Quiet caller is on)2After the last check-in goes unanswered, the agent says a short goodbye and ends the call.
AI disclosureSay a separate disclosure line before the greeting, plus Disclosure text (up to 300 characters)OffYour greeting is always spoken exactly as written. When this is on and the greeting doesn't already say the caller is talking to an AI, this line is spoken first. Leave the text blank to use a default line in the persona's language.
Caller languageFollow the caller's language after 1–5 sentencesOn, 2Pipeline voice. On Live voice the model follows the caller by itself. Uses the language the speech-to-text hears. Hindi and English mixed together (Hinglish) is kept as one mix instead of switching back and forth. A persona whose language is Fixed is never switched.

Document extraction and image models ​

  • Document extraction model: used when the agent reads a document it was given (see Document actions). It also has a Service tier setting. Standard (default) runs immediately. flex is cheaper but best-effort, so an extraction can take minutes when demand is high.
  • Image / vision model: used to understand images a customer sends.

Both start on Platform default. These pickers only list models that can actually read files: Gemini models for both, plus Google Cloud speech-to-text for documents. If an agent still holds another pick, publishing it is refused with a message naming the models that can read the file.

Options ​

FieldWhat it doesOptionsDefaultShown when
ModelsOne section per job. Each model shows only the settings it actually supports.

The generated schema for this node is a single dynamic "Models" block. See the Field reference table below and the sections above for the actual per-job settings.

Worked example: Acme Diagnostics' bilingual agent ​

Setup. Asha's agent answers web chat and inbound phone calls. Her Personality is set to Hindi, region India, with Auto-detect language behaviour.

Action. On the AI Model panel she leaves the Chat agent model on Platform default with Creativity 0.2. Under Voice agent model she switches Mode to Pipeline, keeps Sarvam for speech-to-text and text-to-speech, and picks the female voice Ritu. The note under the voice changes to "Agent refers to itself as female, from this voice." She opens Advanced — brand words and adds Acme, ThyroCheck. She sets AI looking up to Spoken, and under Caller rules she adds a key menu entry 1 → "I want my report status".

Result. Priya calls in and speaks Hindi. The agent transcribes her with an Indian-language speech model, spells "ThyroCheck" correctly, answers, and speaks back in Ritu's voice, using feminine first-person grammar. When Priya presses 1, the agent treats it as "I want my report status". When she later messages on WhatsApp, the agent uses the same feminine grammar even though there's no voice.

What just happened. One card set the chat model, the voice pipeline, the voice, the agent's gender and how it treats callers. The gender followed the voice automatically and carried across every channel.

Field reference ​

Section · fieldValuesDefault
Chat · ModelModels enabled for your workspacePlatform default
Chat · Creativity0–2, step 0.10.0
Chat · Max reply length64–8192, step 642048
Chat · ThinkingLevels the chosen model supportsPlatform default
Voice · ModeLive · PipelineLive
Voice (Live) · Model, VoiceSpeech-to-speech models; that model's voicesPlatform default; model's default voice
Voice (Pipeline) · Speech-to-text, Voice brain, Text-to-speech, VoiceModels for each part; the text-to-speech model's voicesPlatform default
Voice (Pipeline) · Brand wordsOn/Off; up to 50 words, 64 characters eachOn, empty
While the caller waits · AI looking upOff (silence) · Periodic tone · Music — Calm / Bright / Soft chimes · Spoken (Pipeline only)Off
Caller rules · Keypad inputOn/Off; key menu (0–9, *, #)On
Caller rules · Not understoodOn/Off; 1–5 failed turnsOn, 2
Caller rules · Quiet caller / End quiet callsOn/Off; 12–120 seconds; 1–3 check-insOn, 15 seconds, 2
Caller rules · AI disclosureOn/Off; text up to 300 charactersOff
Caller rules · Caller languageOn/Off; 1–5 sentencesOn, 2
How the agent refers to itself · GenderNot set · Female · Male · Non-binaryNot set (or taken from the voice)
Document extraction · Model, Service tierModels that can read files; Standard · flexPlatform default; Standard
Image / vision · ModelModels that can read imagesPlatform default

The panel also has Settings (node label and description) and Notes tabs, like every node.

See also ​

  • AI Agent Node: the agent this card belongs to.
  • Personality: the agent's identity, system prompt, greeting, and language.
  • Voice Providers: more on Live and Pipeline voice.
  • Operator: your team on this agent — ringing, the hand-off to a person, and the waiting sounds while a caller is connected, on hold or waiting for the team.