AI Agents

Create intelligent AI assistants with custom workflows, knowledge bases, voice capabilities, and multi-channel deployment

​

What Are AI Agents?

AI Agents in Revol are intelligent assistants that work with customer interactions across multiple channels — your website widget, phone calls, Telegram, WhatsApp, and Instagram. Each agent has its own personality, knowledge base, conversation workflow, and set of tools.

​

Operating Modes

Agents can operate in three modes depending on your business needs:

Mode	Who Communicates	Agent’s Role
Active	AI Agent	The agent fully handles customer communication — responds to messages, calls tools, follows the workflow, and resolves requests autonomously
Passive	Your Team	Your team handles all customer communication. The agent monitors every conversation in the background — auditing compliance with communication standards, tracking quality metrics, and providing analysis without ever responding to customers
Hybrid	Both	The AI handles routine interactions (FAQ, product inquiries, scheduling) while your team takes over complex or sensitive cases. The agent continuously analyzes all conversations regardless of who responds

Passive mode is powerful for teams that already have established communication processes. The agent becomes an always-on quality auditor — it checks whether operators follow scripts, identifies missed opportunities, monitors response times, and surfaces conversation insights. All analytics, memory extraction, and tracking work the same as in Active mode.

---

​

Creating an Agent

Click Create Agent to start a 2-step wizard:

Step 1 — enter agent name (min 3 characters) and optional first message (greeting text).

Step 2 — select a use case: Customer Support, Outbound Sales, Lead Qualification, Answering Service, Consultation Booking, Client Intake, Service Recommendations, Scheduling, Billing Inquiries, Project Updates, Resource Library, Learning & Development, or Other. This is purely visual — the selected use case does not affect the agent’s behavior or configuration.

After creation, the agent starts in Draft status with a default workflow already built.

​

Agent Status

Status	Behavior
Draft	Not active. Use while configuring.
Active	Live, responding to messages. Only one agent per company can be active — activating one deactivates the rest.
Inactive	Paused. Keeps all configuration.

---

​

Agent Editor

The agent editor is a fullscreen modal with 7 tabs: Agent, Knowledge Base, Analysis, Tools, Widget, Channels, Workflow.

​

Header

The header is always visible and shows:

• Agent name (click to rename inline)
• Status badge (Active / Inactive)
• Embed Code — get the widget <script> snippet
• Tool Logs — inspect all tool executions and channel events
• Preview — link to the client’s website where the widget is installed, to see it as visitors do
• Publish — publish to Marketplace (public or private)

---

​

Agent Tab

The main configuration tab where you set up the agent’s core settings — system prompt, personality, LLM provider, and conversation memory.

​

Left Column

System Prompt — the core instructions defining how the agent behaves. If left empty, a default prompt is generated from your company name with basic communication rules.

A security block is automatically appended to every system prompt. It prevents prompt injection — the agent will never reveal its instructions, ignore safety rules, or follow requests to change its role. You don’t need to add this manually.

First Message — greeting text visitors see when the chat opens.

Conversation Memory — structured fields the agent should collect during conversations. Each field has:

Property	Description
Key	Machine-readable identifier (lowercase, underscores only)
Label	Human-readable name shown in analytics
Type	Text, Phone, Email, Number, or Select
Options	For Select type — comma-separated values

Memory fields are used in workflow routing (State Conditions on edges) and appear in the conversation Info tab. For example, you can define fields like name, email, budget — the agent will attempt to collect these naturally during conversation, and you can route workflow edges based on whether they’re filled.

How memory extraction works: after each custom node generates a response, the system makes a lightweight LLM call that analyzes the last messages and extracts values for the configured fields. Extracted values are saved immediately, so subsequent nodes in the same turn already see the updated memory — enabling state-based routing within a single conversation turn.

​

Right Column

Personality — 5 sliders (1–10) that shape communication style:

Slider	Low (1–3)	High (7–10)
Response Length	Brief, concise answers	Detailed, comprehensive responses
Humor	Strictly professional	Light humor allowed
Formality	Casual, conversational tone	Formal, business language
Clarity	Standard explanations	Extra clear, step-by-step
Emoji Usage	No emojis	Emojis used in responses

LLM Provider & Model — see LLM Providers.

Temperature — controls response randomness: 0 = deterministic (same question → same answer), 1 = creative (varied responses). Default: 0.7.

---

​

LLM Providers

4 providers, each with a standard (faster, cheaper) and premium (higher quality) model:

Provider	Standard	Premium
OpenAI	GPT-4o Mini	GPT-4o
Anthropic	Claude 3.5 Haiku	Claude 3.5 Sonnet
Google Gemini	Gemini 2.0 Flash	Gemini 2.5 Pro
Groq	Llama 3.1 8B Instant	Llama 3.3 70B Versatile

Premium models require an upgraded plan. On standard plans, premium models are locked with an upgrade prompt.

All providers support function calling (tools), streaming, circuit breaker (auto-failover when a provider is down), and retry with exponential backoff on rate limits.

The LLM is set globally on the Agent tab but can be overridden per workflow node — for example, use a fast model for product search and a premium model for the final response formatting.

---

​

Workflow

The Workflow tab is a visual canvas where you design the agent’s conversation logic. Instead of a single prompt, the workflow splits processing across nodes — each with its own prompt, tools, and knowledge base — connected by edges with routing conditions.

​

Default Workflow

Every new agent starts with this pre-built workflow:

[Voice Input*] → [Start] → [Product Agent] ──→ [Formatter] → [Voice Output*]
                           → [Media Agent]   ──↗
                           → [Company Agent] ──↗

Voice nodes are inactive by default — they activate when you enable voice.

The Start node routes the message to three parallel system agents. Each processes the message with its own tools and knowledge. The Formatter combines their results into a single response.

​

Node Types

Node	Type	Description
Start	start	Entry point. Always present, cannot be deleted.
Product Agent	system_product	Searches products, checks availability, shows details. Has access to product tools.
Media Agent	system_media	Retrieves photos, videos, and documents.
Company Agent	system_company	Company info, support questions.
Response Formatter	system_formatter	Combines outputs from parallel nodes into a coherent final response.
Voice Input (STT)	system_stt	Speech-to-Text conversion. Inactive by default.
Voice Output (TTS)	system_tts	Text-to-Speech synthesis. Inactive by default.
Custom	custom	Your own node with custom prompt, tools, KB, and LLM settings.

​

Adding Custom Nodes

Click + Add Node in the canvas toolbar → enter a name → a new Custom node appears on the canvas. You can add as many custom nodes as needed.

Custom nodes are the most powerful part of the workflow. Each custom node is essentially its own mini-agent with:

• Conversation Goal — a system prompt specific to this node (e.g., “Help users choose the right subscription plan based on their team size and budget”)
• Tools — select which tools this node can call (independent from the agent-level tools)
• Knowledge Base — select specific knowledge sources for this node’s RAG context
• LLM Override — use a different model for this node
• Agent Mode — enable multi-turn tool calling (see below)

​

Node Settings Panel

Click any node on the canvas to open its settings panel on the right. The panel has tabs that vary by node type:

General tab (all nodes):

Setting	Description
Name	Display name (disabled for Start node)
Active	Toggle whether this node participates in the workflow
Conversation Goal	System prompt for this node. Available on all nodes except Start and Formatter.

Agent Mode (custom nodes with tools only):

When enabled, the node can make multiple rounds of tool calls before responding. Without Agent Mode, the node makes at most 1 tool call per message.

Setting	Range	Default	Description
Max Rounds	2–10	5	How many tool-calling rounds the LLM can make before it must respond
Timeout	10–60s	30s	Maximum time for the entire agent mode execution

Use Agent Mode when a node needs to chain multiple tools — for example, first search products, then check availability of the best match, then get detailed specs. Without Agent Mode, the node would need 3 separate conversation turns to do this.

LLM Override (all except Start):

Toggle to use a different LLM provider and model for this specific node. When off, the node uses the agent’s global LLM setting. The dropdown shows “Default (agent setting)” when not overridden.

Tools tab (custom and system nodes):

Lists all available tools grouped by category (Products, Support, Documents). Each tool has a toggle switch. Tools enabled here are independent from the agent-level Tools tab — you control exactly which tools each node has access to.

Knowledge Base tab (custom nodes only):

Select specific knowledge sources for this node’s RAG context. Has an “Add document” dropdown with search and type filtering. If no sources are selected, the node falls back to the agent’s global knowledge base.

A hint at the bottom links to the agent’s Knowledge Base tab: “To make more data available for this node, add files to the agent’s Knowledge Base.”

Edges tab (all nodes):

Configure outgoing connections to other nodes. See Edge Conditions below.

Voice Settings tab (STT and TTS nodes only):

Configure speech recognition and synthesis — see Voice section.

Delete Node — available only for custom nodes. System nodes cannot be deleted.

​

Edge Conditions

Edges connect nodes and control message routing. Each edge has a target node, a condition type, and a priority. The workflow evaluates edges in priority tiers — the first tier that produces a match wins, lower tiers are not evaluated.

Click + Add Edge in a node’s Edges tab to create a connection.

Condition	Priority	When It Routes
Keyword	100 (highest)	Message contains any of the specified keywords. Case-insensitive, word-boundary matching.
State Condition	95	All specified memory field conditions are met (AND logic).
Always	90	Always routes — use for unconditional connections.
Intent	50	Based on detected message intent.
Fallback	10 (lowest)	Routes only if no other edge from this node matched.

​

Keyword Condition

Enter comma-separated keywords. The workflow checks if the visitor’s message contains any of them using case-insensitive word-boundary matching.

Multiple keyword edges can match the same message — all matching edges fire in parallel, sending the message to multiple nodes simultaneously.

Example: Keywords price, cost, pricing, how much — the edge fires when the visitor asks “How much does it cost?” or “What’s the pricing?”

​

State Condition

Check values of Conversation Memory fields. You build rules with:

Operator	Meaning	Example
is filled	Field has any value	email is filled → route to “Send offer” node
is empty	Field has no value	name is empty → route to “Ask name” node
equals	Exact match	budget equals enterprise
not equals	Does not match	plan not equals free
contains	Substring match	interests contains premium
greater than	Numeric comparison	budget > 5000
less than	Numeric comparison	team_size < 10

Multiple conditions within one edge use AND logic — all must be true.

State conditions are skipped entirely if memory is empty (no fields collected yet). This means is empty conditions won’t fire until the agent has started collecting at least one memory field.

Example workflow with State Conditions:

[Start] → (name is empty)     → [Ask Name Node]
        → (name is filled,     → [Send Offer Node]
           email is filled)
        → (fallback)           → [General Chat Node]

​

Always Condition

The edge always fires. All Always edges from a node fire in parallel — the message is sent to every target simultaneously. This is how the default workflow fans out from Start to three system agents at once.

​

Fallback Condition

Routes only when no higher-priority edge matched. Unlike Keyword and Always, only one fallback edge fires (the first one) — no parallel execution.

If no edge matches and there’s no fallback, the node’s output may not reach the Formatter and the visitor won’t get a response from that branch.

​

Merge Strategy

When multiple nodes produce results in parallel (like the 3 system agents in the default workflow), the Formatter combines them:

Strategy	Cost	How It Works
Concat	Free	Concatenates all node outputs into context for the Formatter
LLM	Additional LLM call	Uses an LLM to synthesize a single coherent response from all outputs

​

Canvas Controls

Control	Action
Drag empty area	Pan the canvas
Ctrl + Scroll or pinch	Zoom (0.3x–2.0x)
Two-finger scroll	Pan
+ Add Node	Create a new custom node
Template	Apply a pre-built workflow template from the marketplace
Reset	Rebuild the default workflow (confirmation dialog)
Test	Open the test chat panel at the bottom of the canvas

---

​

Tools

Tools are functions the agent can call during conversations. They extend the agent beyond text generation — searching products, querying databases, sending emails, making calls.

​

How Tools Work

1. LLM decides to use a tool

   Based on the visitor’s message and the tool descriptions in its prompt, the LLM generates a tool call with parameters (e.g., get_products({ query: "running shoes", available_only: true })).

2. Tool executes

   The system runs the tool function with the provided parameters and gets a result.

3. Result goes back to LLM

   The tool result is injected back into the conversation. The LLM uses it to formulate a natural response.

4. Multi-round (Agent Mode)

   If the node has Agent Mode enabled, the LLM can make another tool call based on the first result — up to the configured max rounds.

Every tool call is logged with input, output, execution time, and status. View logs via Tool Logs in the agent header.

​

Built-In Tools (9)

Always available, no integration required:

Tool	Category	What It Does
get_products	Products	Search products by name/description. Returns list with prices and availability.
get_product_details	Products	Full details for one product — all parameters, pricing, description.
check_availability	Products	Check if a specific product is in stock.
search_by_parameters	Products	Filter products by attribute values with operators: =, <=, >=, <, >, like.
get_company_info	Support	Company name, description, phone, contacts.
search_documents	Documents	Semantic RAG search — finds relevant passages across the knowledge base.
get_photos	Documents	Retrieve photos by query or product ID.
get_videos	Documents	Retrieve videos by query or product ID.
get_documents	Documents	Retrieve PDF/Word/Excel files, filterable by format.

​

Integration Tools (60+)

Connect external services in Integrations to unlock tools:

VoIP — Twilio, Binotel, Ringostat

Make outbound calls, send SMS, retrieve call history.

Telegram

Send messages, send files, edit/delete messages, get chat history and info.

WhatsApp

Send messages, send media, send template messages, get profile, mark as read.

Facebook & Instagram

Send messages, send media, send buttons (Facebook), get profile. Facebook Ads: get campaigns, ad sets, insights, pause campaigns.

Gmail

Read inbox, send emails, reply, create drafts, search, get attachments.

Google Calendar

List/create/update/delete events, check availability, find free time slots.

Google Docs

Read/create/append/export/search documents.

Google Sheets

Read/write ranges, append rows, update cells, search rows, create spreadsheets.

Google Drive

List/read/create/update/delete files, create folders, share files, search.

Google Meet

Create meeting links.

Google Ads

Get campaigns, get keywords, pause campaigns.

Webhooks

Send custom JSON payloads to any URL, test webhook connections.

​

Tools Tab vs Node-Level Tools

There are two places to manage tools:

1. Agent → Tools tab — shows all integration tools grouped by provider. Toggle tools on/off at the agent level. Tools here become available to all workflow nodes.

2. Workflow → Node Settings → Tools tab — toggle tools per node. A node can only use tools that are enabled at the agent level. This lets you restrict which nodes have access to which tools.

Example: Enable Gmail tools at the agent level, but only give the “Send Follow-up” custom node access to send_email — other nodes won’t trigger emails.

---

​

Knowledge Base (RAG)

The Knowledge Base tab connects data sources that the agent uses to answer questions. When a visitor asks something, the agent searches your knowledge base using vector similarity (RAG — Retrieval-Augmented Generation) and includes relevant context in its response.

​

Source Types

Source	What Gets Indexed
Products	Name, description, additional prompt, price
Documents	File name + extracted text content (PDF, Word, Excel, TXT)
Photos	Photo name + description
Videos	Video name + description/content
Text	Name + free-text content
Links	URL name + fetched page content
Company	Company name, promo text, description, phone

​

How RAG Works

1. Add sources

   In the Knowledge Base tab, click “Add document” and select sources from your company’s data — products, files, text snippets, links.

2. Training

   The system chunks each source into segments (default max 2000 chars per chunk, 20% overlap at sentence boundaries), generates vector embeddings using OpenAI text-embedding-ada-002 (1536 dimensions), and stores them in PostgreSQL with pgvector.

3. Visitor sends a message

   The visitor’s message is embedded into the same vector space. The system finds the most similar chunks using cosine similarity.

4. Context injection

   Top matching chunks (default: up to 5, minimum similarity 0.6) are injected into the agent’s prompt as knowledge base context.

5. Response

   The LLM sees the visitor’s question alongside relevant knowledge and generates an informed answer.

Training runs automatically when you add or change sources. Progress is shown in the Knowledge Base tab (percentage, items processed).

​

Agent-Level vs Node-Level Knowledge

• Agent-level KB (Knowledge Base tab) — sources available to all workflow nodes
• Node-level KB (Workflow → Node → Knowledge Base tab) — restrict RAG to specific sources for this node only

Example: Your “Technical Support” custom node only sees product manuals and FAQ documents, while the “Sales” node sees product catalogs with pricing.

Specific, well-structured content performs better than long general documents. Product descriptions with clear parameters give the agent precise answers. Short FAQ entries work better than lengthy manuals.

---

​

Voice

Voice settings are configured in the Widget tab’s Voice sub-tab or directly in the workflow’s STT/TTS nodes.

​

Speech-to-Text (STT)

Setting	Options
Provider	OpenAI Whisper, Google Speech (coming soon)
Language	Ukrainian, English
Greeting	Text + pre-synthesized audio played when voice starts
Farewell	Text + pre-synthesized audio played when voice ends

​

Text-to-Speech (TTS)

Setting	Options
Provider	OpenAI TTS (ElevenLabs and Google Cloud coming soon)
Voice	Alloy (neutral), Echo (warm), Fable (expressive), Onyx (deep), Nova (friendly, default), Shimmer (gentle)
Model	tts-1 (standard), tts-1-hd (HD quality)
Speed	0.5x – 2.0x

​

Voice Pipeline

1. Audio received

   Visitor’s audio sent to server (max 10MB, formats: webm, ogg, wav, mp3).

2. STT transcription

   Audio transcribed to text. Transcript streamed back in real-time via SSE.

3. Workflow processing

   Transcript processed through the same workflow as text — RAG, tools, node routing.

4. TTS synthesis

   Response split into sentences, each synthesized to audio. Chunks stream back as generated.

5. Playback

   Visitor hears the response as streaming audio while seeing the text.

Voice requires a plan with voice capabilities. On plans without voice, voice controls show an upgrade prompt and Chat Only Mode is forced on.

---

​

Channels

The Channels tab controls where your agent receives messages.

Channel	Requires	Description
Widget	Nothing (built-in)	Chat widget on your website via tracker script
Phone	VoIP integration	Inbound/outbound calls via Twilio, Binotel, or Ringostat
Telegram	Telegram integration	Telegram bot conversations
WhatsApp	WhatsApp integration	WhatsApp Business API
Instagram	Instagram integration	Instagram DM

​

Connecting a Channel

Widget — click Enable. The widget responds on your website immediately.

Other channels:

1. Connect the integration in Integrations
2. In the Channels tab, select a resource from the dropdown (e.g., a phone number, a bot, a page)
3. Click Enable — a unique webhook URL is generated

Each channel card shows connection status, resource name, total conversations, and the webhook URL with a copy button.

---

​

Widget Customization

The Widget tab has a live preview on the left (desktop/tablet/mobile switcher) and a settings panel on the right with 3 sub-tabs.

​

Appearance

Setting	Description
Colors	6 color pickers: Primary, Chat button, Message background, Animation, Rating stars, Status text
Labels	Main title, Call button text, Chat button text, Input placeholder
Position	Bottom Right, Bottom Center, Bottom Left, Top Right, Top Center, Top Left
Theme	Light or Dark
Size	Small, Medium, Large
Border Radius	0–50px corner roundness
Opacity	0–100% background opacity
Avatar	Custom avatar URL

​

Behavior

Setting	Description
Language	English, Ukrainian, Polish, German, Spanish
Feedback Collection	Show star rating after conversation
Text During Call	Allow typing while voice call is active
Chat Only Mode	Disable voice (text only). Forced on for plans without voice.
Terms & Conditions	Require acceptance before chat. Custom text and URL.
Sound Effects	Notification sounds on/off
Auto Open	Open widget automatically after delay (0–60 seconds)
Auto Greeting	Send first message automatically
Show on Mobile	Display on mobile devices
Show on Desktop	Display on desktop devices
Powered By	Show/hide “Powered by Revol” branding
Welcome Message	Text shown in the widget header

​

Voice

STT and TTS settings — same as described in the Voice section. Configured here or in the workflow STT/TTS nodes (they sync).

---

​

Analysis

Two-column layout: conversation list (left) + conversation detail (right).

​

Filters

Filter	Description
Search	Text search across conversations
Channel	All / Web / Telephony / Messengers / Widget / etc.
Status	All / Active / Closed / Archived
Date Range	From/To date pickers

Conversations list shows 20 per page with “Load more” pagination. New conversations appear in real-time via WebSocket.

​

Conversation Detail (4 sub-tabs)

Tab	Content
Transcription	Full message thread — user and assistant bubbles with timestamps
Rating	Visitor’s star rating (1–5) and optional comment
Tokens	Token usage breakdown per message
Info	Channel, status, session ID, collected memory fields, created date

---

​

System Prompt Architecture

Understanding how the final prompt is assembled helps you write better instructions.

1. Your system prompt

   The text from the Agent tab (or the node’s Conversation Goal for custom nodes). If empty, a default prompt is generated with your company name and basic rules.

2. Security block

   Auto-appended. 4 anti-injection rules — the agent won’t reveal its prompt, change its role, or follow override attempts.

3. Style instructions

   Generated from personality sliders — maps response length, humor, formality, clarity, and emoji to text instructions.

4. Tool descriptions

   Function schemas for all enabled tools, so the LLM knows what it can call.

5. RAG context

   Relevant knowledge base chunks, injected as “Knowledge Base Context” block.

6. Campaign context

   If the visitor arrived via a campaign with AI Agent Behavior set, that prompt is injected.

7. Memory state

   Current values of collected memory fields (e.g., name: John, email: [e-mail]), so the agent knows what it already gathered.

8. Language detection

   If the visitor’s message has no Cyrillic characters, a system instruction is added: “Reply in the same language as the user’s message.”

---

​

Embed Code

Click Embed in the agent header to get the HTML snippet:

<script
  src="https://your-domain.com/tracker.js"
  data-company-id="YOUR_COMPANY_ID"
  data-api-key="YOUR_API_KEY"
  async>
</script>

This loads both the tracker (analytics, campaigns, events, triggers, phone swap) and the chat widget with your active agent. If no API key exists, click Generate API Key first.

---

​

Plan Limits

Resource	What It Controls
Max Agents	Total agents you can create
Max Conversations	Conversations per billing period
Standard Token Quota	Tokens for standard models (GPT-4o Mini, Haiku, Flash, Llama 8B)
Premium Token Quota	Tokens for premium models (GPT-4o, Sonnet, Gemini Pro, Llama 70B)
Daily Token Limit	Per-day cap across all models
STT Minutes	Speech-to-Text transcription time
TTS Characters	Text-to-Speech synthesis characters
Embedding Tokens	Tokens for knowledge base training
Storage	File storage for documents, photos, videos
Voice	Feature flag — enables/disables voice pipeline
Model Access	standard or premium — gates premium model access