Generative Agents

Generative assistants are AI agents based on Large Language Models (LLMs) that process natural language messages, query knowledge bases, and execute actions via triggers. Unlike deterministic flows, assistants dynamically decide how to respond.

Assistant vs. Chatbot

Prompt Configuration

`Prompt`: Main Instructions

The main instruction field for the assistant. Defines personality, goal, constraints, and behavior.

• Maximum: 4,000 characters
• Minimum: 10 characters

You are Sofia, a sales assistant for Company X.
Your goal is to qualify leads and schedule product demos.

PERSONALITY:
- Professional yet relaxed tone
- Empathetic and solution-oriented
- Always use the customer's name when available

CONSTRAINTS:
- Never mention competitors
- Don't discuss pricing before qualifying the lead
- If you don't know how to respond, offer to connect with a human

QUALIFICATION PROCESS:
1. Understand the problem/need
2. Identify company size
3. Ask about available budget
4. Offer a free demo

`CustomPrompt`: Advanced Template

When IsCustomPrompt is enabled, the CustomPrompt replaces how the knowledge context is injected. Use the ${{conhecimento}} variable to control where RAG data appears:

You are an assistant specialized in technical support.
Answer ONLY based on the information below:

=== KNOWLEDGE BASE ===
${{conhecimento}}
=== END OF BASE ===

Current time reference: ${{horario}}

If the answer is not in the knowledge base above,
respond: "I couldn't find that information in the documentation.
I can connect you with our technical team."

`RestrictedByContext`

When enabled, the assistant automatically adds to the prompt the instruction:

"When the subject goes beyond the provided context, inform the user that you cannot answer."

Ideal for specialized assistants that should stay within a specific domain.

Language

Forces the language of the assistant's responses:

Knowledge Base (RAG)

The Retrieval-Augmented Generation system allows the assistant to respond based on specific documents from your company.

How it works

User message
      ↓
Vector similarity search
      ↓
Most relevant snippets retrieved
      ↓
Inserted into LLM context
      ↓
LLM responds based on snippets

Adding documents

Go to Assistant → Knowledge Base → Add File. Accepted formats:

• PDF, DOCX, TXT
• Web page URLs
• Structured data (JSON, CSV)

`${{conhecimento}}` in CustomPrompt

The ${{conhecimento}} variable automatically receives the most relevant snippets retrieved by RAG. Use it in CustomPrompt to position them exactly where you want in the prompt.

If CustomPrompt is not used, the snippets are automatically inserted after the main prompt.

Audio Transcription

The assistant automatically transcribes voice messages sent by the user before processing them. The transcription is returned in the transcriptions field of the API response.

Copilot

The Copilot is a second agent that operates in internal mode, used to assist human agents or automation systems.

Configuration

Copilot Mode vs. Normal Mode

Copilot is activated by sending "copilot": true in the session update request body.