Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.tavus.io/llms.txt

Use this file to discover all available pages before exploring further.

Personas are the ‘character’ or ‘AI agent personality’ and contain all of the settings and configuration for that character or agent. For example, you can create a persona for ‘Tim the sales agent’ or ‘Rob the interviewer’. Personas combine identity, contextual knowledge, and CVI pipeline configuration to create a real-time conversational agent with a distinct behavior, voice, and response style. They are the main place you configure CVI behavior before you start a Conversation. At a glance
  • Persona — The agent’s identity (name, voice, behavior) plus pipeline mode, default replica, layers, documents (RAG), objectives, and guardrails.
  • Relationship to CVI — Personas hold the settings that drive a real-time CVI session; see the CVI overview for the full stack (WebRTC, layers, Phoenix, and latency characteristics on the default path).
  • Layers (order of guides below) — Perception → STT → Conversational Flow → LLM → TTS; each has its own configuration page.

​
Persona Customization Options

Each persona includes configurable fields. Here’s what you can customize:
  • Persona Name: Display name shown when the replica joins a call.
  • System Prompt: Instructions sent to the language model to shape the replica’s tone, personality, and behavior.
  • Pipeline Mode: Controls which CVI pipeline layers are active and how input/output flows through the system.
  • Default Replica: Sets the digital human associated with the persona.
  • Layers: Perception, STT, conversational flow, LLM, and TTS—each processes part of the interaction and can be tuned independently (see Layers below).
  • Documents: A set of documents that the persona has access to via Retrieval Augmented Generation.
  • Objectives: The goal-oriented instructions your persona will adhere to throughout the conversation.
  • Guardrails: Conversational boundaries that can be used to strictly enforce desired behavior.

​
Objectives & Guardrails

Provide your persona with robust workflow management tools, curated to your use case

Objectives

The sequence of goals your persona will work to achieve throughout the conversation—for example gathering a piece of information from the user.

Guardrails

Conversational boundaries that can be used to strictly enforce desired behavior.

​
Layers

Explore our in-depth guides to customize each layer to fit your specific use case:

Perception Layer

Defines how the persona interprets visual input like facial expressions and gestures.

STT Layer

Transcribes user speech into text using the configured speech-to-text engine.

Conversational Flow Layer

Controls turn-taking, interruption handling, and active listening behavior for natural conversations.

LLM Layer

Generates persona responses using a language model. Supports Tavus-hosted or custom LLMs.

TTS Layer

Converts text responses into speech using Tavus or supported third-party TTS engines.

​
Pipeline Mode

Tavus provides several pipeline modes, each with preconfigured layers tailored to specific use cases:
The default and recommended end-to-end configuration optimized for real-time conversation. All CVI layers are active and customizable.
  • Low utterance-to-utterance latency with Tavus defaults (see the CVI overview)
  • Best for natural humanlike interactions

CVI quickstart

​
Custom LLM / Bring Your Own Logic

Use this mode to integrate a custom LLM or a specialized backend for interpreting transcripts and generating responses.
  • Adds latency due to external processing
  • Does not require an actual LLM—any endpoint that returns a compatible chat completion format can be used

Integrate your own custom LLM or logic