How to Build the Omni-Channel Agents (Mini-Course)

Overview

This mini-course walks through building an omni-channel AI agent that works across WhatsApp, SMS, and live chat. It focuses on:

• Handling multi-message user inputs
• Running an agent workflow
• Doing RAG (knowledge base retrieval)
• Returning multiple assistant messages in a “human-like” way

---

Demo / Expected Behavior

• A phone number (e.g., OpenPhone) is connected to an omni-channel agent.
• When the user sends separate messages that form one idea (e.g., “Hi…”, “I like dogs”, “Do you sell cars?”), the system:
  1. Groups utterances into one chunk
  2. Sends the chunk to the agent
  3. Receives a response and sends it back as multiple messages (like humans do)

The bot also:

• Keeps conversation context (“remembers the whole conversation”)
• Can issue follow-ups/reminders if the user doesn’t respond

---

Agent Types (Conceptual Structure)

Single-prompt agents

• One user question → context (guidelines + knowledge base) → single response
• Easier to implement and can work well for many use cases

Multi-step agents

• Deeper conversational flow that mimics human behavior using strict rules/guidelines and multi-turn orchestration
• This video builds a multi-step omni-channel system

---

Backend Architecture: Multi-Step Omni-Channel Flow

The speaker describes a standard multi-step pipeline (platform-agnostic) with these stages:

1. Utterance processing

   • Capture multiple inbound messages
   • Group them into a single chunk before sending to the agent backend

2. Agent reasoning (LLM / agent platform)

   • The “agent step” typically takes one input and returns one output
   • The system may need post-processing to simulate multi-message outputs

3. Humanizing / response chunking

   • The agent cannot natively output multiple chat messages
   • Solution:
     • An additional prompt generates the reply
     • It also determines how many messages to split it into
   • JSON is produced with message_1, message_2, etc., then parsed and sent as separate messages

4. Interface / conversation loop

   • A UI layer supports conversation management and “handoff” to a human
   • Supports stopping the bot and letting a human take over

---

Tooling Used in the Build

• Voiceflow (the “brain/engine”)

  • Used to build conversational logic, connections, and RAG knowledge retrieval
  • Mentions Voiceflow’s “agent step” and “functions/paths” concept

• n8n

  • Used for automation “tools” on top of Voiceflow when needed (not always required)

• One Prompt (white-labeled automation suite)

  • Core solution for omni-channel orchestration, including:
    • multi-channel “receive/process/wait/reply” workflows
    • humanizing and utterance grouping
  • Includes “production-ready” omni-channel templates intended to plug into Voiceflow

---

RAG Knowledge Base Setup (Knowledge Ingestion)

Building the KB (key ideas)

• For websites

  • Use a scraping tool (example mentioned: Firecrawl) to extract and convert content into markdown/structured text
  • Emphasizes preserving link accuracy for complex sites

• For documents (PDF/CSV)

  • Use a document parser (example mentioned: LlamaParse / “L pars”) with advanced parsing settings:
    • tables
    • external links
    • complex structure

Important note

• KB structure quality matters, because retrieval affects downstream answering accuracy and link correctness.

---

The “7 Parts” Highlighted During Development

1. RAG knowledge base
2. B Persona
3. Intent classification
4. Sub-agent
5. Engagement prompts
6. Connections (API/tool calls inside the agent)
7. Utterance processing + humanizing
   • Implemented via One Prompt workflows / automations

---

Development Walkthrough Highlights (Examples)

Building the KB

• Create a knowledge base and import scraped content (example uses apple.com)
• Ensures links and structured markdown are preserved for better retrieval

Persona setup

• Store a “persona” variable (e.g., friendly Apple assistant) in markdown-formatted rules
• Persona influences:
  • brevity
  • tone
  • example-style phrasing

Intent classification + routing

• Uses Voiceflow “agent step” / “function/path” behavior:
  • the agent decides which function/path to run (e.g., billing vs booking)
• Example:
  • Billing path triggers an Airtable lookup
    • requires an account_number variable with a specified input format
  • General flow calls the RAG KB for non-billing queries

Connections example (Airtable via n8n)

• After intent classification, query Airtable by account number
• Retrieved fields (e.g., payment_status) are passed into the engagement prompt context

RAG integration in the flow

• For general questions:
  • run KB search to produce KB chunks
  • feed KB chunks into the engagement prompt so responses are grounded in retrieved content
• If the KB can’t find relevant info:
  • the workflow supports a “not found” pathway to avoid errors

---

Omni-Channel Deployment: Live Chat + WhatsApp + SMS

Live chat widget

• Create an embedded chat widget on a site (e.g., via Framer embedding the provided script)
• Configure Voiceflow API keys in One Prompt’s omni-channel workflows
• Replace triggers/blocks:
  • remove SMS triggers
  • use live chat trigger instead
• Configure multi-message input simulation:
  • live chat is treated similarly to SMS, including follow-ups after delays

SMS and WhatsApp integration

• SMS
  • use OpenPhone-style inbound triggers connected to the omni workflow
• WhatsApp
  • replace live chat blocks with WhatsApp send/trigger blocks
  • templates may be needed depending on whether you’re outside/inside a WhatsApp messaging window (author references a 24-hour window)

Key operational point:

• Each inbound message triggers a new workflow instance
• the system waits/group messages for a short time before sending the combined chunk to the agent

---

How “Utterance Grouping” and “Multi-Message Output” Work

Incoming messages

• Each inbound message is stored in custom fields like dm1, dm2, etc.
• After a configurable waiting period:
  • create a master message combining them
• Send the master message to Voiceflow via API

Outgoing messages

• Voiceflow returns one “generated response,” but the system requests the LLM to split it into multiple chunks
• Output is JSON containing multiple message parts (message_1, message_2, …)
• JavaScript parsing maps chunks to variables, then sends them as separate messages (creating multi-bubble replies)

---

Human Handoff & Interface Features

The UI/dashboard supports:

• Managing conversations across channels
• Assigning conversations to humans/agents
• Pausing bot replies
• Human handoff:
  • if a “human help” condition is triggered, it transfers the conversation and stops the bot

---

Main Speakers / Sources

• Speaker: The instructor (Eugene, referred to as “Eugene cin” in the subtitles/demo)
• Product sources referenced:
  • Voiceflow
  • One Prompt (white-labeled suite)
  • optional: n8n, Airtable, OpenPhone, Make.com, Firecrawl, LlamaParse

---