Workflow-Editor

# Instruction: Generate a Revol Workflow Template JSON

## Context

I use Revol — a platform for building AI sales agents. Each agent
has a visual workflow editor with nodes and edges that define
conversation logic.

I need you to generate a workflow template as a JSON object that
I can import into my Revol agent. The JSON follows a strict schema
used by Revol's template system.

## JSON Schema

The output must be a single JSON object with these top-level keys:

{
  "nodes_data": [...],
  "edges_data": [...],
  "tools_data": {...},
  "memory_fields_data": [...]   // optional
}

### nodes_data (array of objects)

Each node object has these fields:

| Field | Type | Required | Description |
|-------|------|----------|-------------|
| index | integer | yes | Unique index (0-based), used for edge references |
| node_type | string | yes | One of: start, custom, system_stt, system_tts, system_product, system_media, system_company, system_formatter |
| name | string | yes | Display name on the canvas |
| position_x | integer | yes | Canvas X coordinate |
| position_y | integer | yes | Canvas Y coordinate |
| is_active | boolean | yes | Whether this node is active |
| conversation_goal | string | no | System prompt for this node (custom nodes only) |
| config | object | no | Node-specific configuration (see below) |
| llm_provider_override | string | no | Override LLM provider: openai, anthropic, gemini, groq |
| llm_model_override | string | no | Override model name |
| temperature_override | number | no | Override temperature (0.0-2.0) |

Node types and their purpose:
- start — entry point, exactly one per workflow, no conversation_goal
- custom — LLM-powered node with its own prompt, tools, and knowledge base
- system_product — searches products, checks availability, shows details
- system_media — retrieves photos, videos, documents
- system_company — company info and support questions
- system_formatter — combines outputs from parallel nodes into final response
- system_stt — speech-to-text (voice input), usually inactive by default
- system_tts — text-to-speech (voice output), usually inactive by default

Config for system_stt:
{
  "provider": "openai",
  "language": "en",
  "greeting_text": "Hi! How can I help you?",
  "greeting_audio_url": null,
  "farewell_text": "Thanks! Goodbye!",
  "farewell_audio_url": null
}

Config for system_tts:
{
  "provider": "openai",
  "voice": "nova",
  "model": "tts-1",
  "speed": 1.0
}

Config for custom nodes (optional):
{
  "save_to_memory": ["field_key1", "field_key2"],
  "max_tool_rounds": 3,
  "tool_timeout_seconds": 30
}

### edges_data (array of objects)

Each edge object connects two nodes:

| Field | Type | Required | Description |
|-------|------|----------|-------------|
| source_index | integer | yes | Index of the source node |
| target_index | integer | yes | Index of the target node |
| condition_type | string | yes | One of: always, keyword, intent, state, fallback |
| condition_value | object | no | Condition parameters (null for always/fallback) |
| priority | integer | no | Evaluation priority (default: 90 for always) |

Condition types and priorities:
- keyword (priority: 100) — matches words in the message
  condition_value: { "keywords": ["word1", "word2", ...] }
- state (priority: 95) — checks memory field values
  condition_value: { "conditions": [{"field": "name", "operator": "filled|empty|=|!=|contains|>|<", "value": ""}] }
- always (priority: 90) — unconditional, enables parallel fan-out
  condition_value: null
- intent (priority: 50) — LLM classifies the message intent
  condition_value: { "intents": ["intent_name"] }
- fallback (priority: 10) — routes only when nothing else matches
  condition_value: null

### tools_data (object, keys are node indexes as strings)

Maps node index to an array of tool names:

{
  "2": ["search_documents", "get_company_info"],
  "3": ["get_products", "get_product_details"]
}

Available built-in tools (9):
- get_products — search products by name/description
- get_product_details — full details for one product
- check_availability — check if product is in stock
- search_by_parameters — filter products by attributes
- get_company_info — company name, description, contacts
- search_documents — semantic RAG search in knowledge base
- get_photos — retrieve photos
- get_videos — retrieve videos
- get_documents — retrieve PDF/Word/Excel files

### memory_fields_data (optional array)

Define structured fields the agent should collect:

[
  {"key": "client_name", "label": "Client Name", "type": "text"},
  {"key": "client_phone", "label": "Phone", "type": "phone"},
  {"key": "topic", "label": "Topic", "type": "select", "options": ["sales", "support"]}
]

Field types: text, phone, email, number, select (with options array).

## Workflow Patterns

### Pattern 1: Simple (single expert)
Start → Custom Node → Formatter
Best for: FAQ bots, simple support, single-topic agents.

### Pattern 2: Parallel experts (always routing)
Start → Expert A → Formatter
      → Expert B → Formatter
      → Expert C → Formatter
Best for: agents that need to search products + documents + company info simultaneously.

### Pattern 3: Smart router (intent/keyword routing)
Start → (intent: sales) → Sales Expert → Formatter
      → (intent: support) → Support Expert → Formatter
      → (fallback) → General Expert → Formatter
Best for: multi-topic agents where different questions need different expertise.

### Pattern 4: Lead collection (state routing)
Start → (name empty) → Collect Name → Formatter
      → (name filled, phone empty) → Collect Phone → Formatter
      → (all filled) → Final Response → Formatter
Best for: lead qualification, appointment booking, form-filling flows.

### Pattern 5: Voice agent
STT → Start → Expert → Formatter → TTS
Same as any pattern above, but with STT/TTS nodes active.

## Layout Guidelines

- Start node at position (0, 0)
- STT node to the left: (-200, 0)
- Expert nodes to the right: (200, -100), (200, 0), (200, 100)
- Formatter further right: (400, 0)
- TTS node furthest right: (600, 0)
- Vertical spacing between parallel nodes: 100-200px

## Rules

1. Every workflow MUST have exactly one "start" node
2. Every workflow MUST have exactly one "system_formatter" node
3. STT and TTS nodes should be included but set to is_active: false
   unless voice is specifically requested
4. Custom nodes should have clear, focused conversation_goal text
5. Each custom node should have at least search_documents tool
6. Use intent routing instead of keyword for multilingual agents
7. The fallback edge ensures no message goes unanswered
8. Node indexes must be sequential starting from 0
9. All edge source_index and target_index must reference valid node indexes
10. conversation_goal should be written in the language the agent will use

## Output

Return ONLY the JSON object — no markdown, no explanation, no
code fences. The JSON must be valid and parseable.

## Input

My business: [DESCRIBE YOUR BUSINESS]
Agent role: [e.g., sales consultant, support agent, lead collector]
Languages: [e.g., English, Ukrainian, multilingual]
Voice enabled: [yes/no]
Key topics/departments: [e.g., products, billing, technical support]
Memory fields to collect: [e.g., name, phone, email, or none]
Special requirements: [any additional requirements]

Erste Schritte

KI-Agenten

Gespräche

Abrechnung

Workflow-Editor

Überblick

Knotentypen

Start-Knoten

LLM-Knoten

Bedingungs-Knoten

Tool-Aufruf-Knoten

Sprachanruf-Knoten

End-Knoten

Knoten verbinden

Best Practices

Workflow mit Claude Code generieren

​Überblick

​Knotentypen

​Start-Knoten

​LLM-Knoten

​Bedingungs-Knoten

​Tool-Aufruf-Knoten

​Sprachanruf-Knoten

​End-Knoten

​Knoten verbinden

​Best Practices

​Workflow mit Claude Code generieren

Überblick

Knotentypen

Start-Knoten

LLM-Knoten

Bedingungs-Knoten

Tool-Aufruf-Knoten

Sprachanruf-Knoten

End-Knoten

Knoten verbinden

Best Practices

Workflow mit Claude Code generieren