Configuring AI Nodes

How to set up and configure AI Prompt and AI Agent nodes

This guide covers how to configure AI Prompt and AI Agent nodes for different use cases.

AI Prompt Configuration

Basic Configuration

Parameter	Type	Required	Description
model	enum	Yes	Model tier (low, medium, high)
messages	array	Yes	Messages array with role and content
return_type	enum	Yes	Output type (string, integer, boolean, etc.)

Messages Array

The messages array defines what the AI processes:

json

[
  {
    "role": "system",
    "content": "You are a meeting analyst. Extract key information accurately."
  },
  {
    "role": "user",
    "content": "{{ json.callRecording.transcriptSummary }}"
  }
]

Message Roles

System Message

Sets the AI's persona and behavior
Provides instructions for the task
Should be clear and specific

liquid

You are a professional meeting summarizer.
- Be concise (2-3 sentences max)
- Focus on decisions and action items
- Use bullet points for lists

User Message

Contains the data to process
Uses Liquid templates for dynamic content
Should include all relevant context

liquid

Summarize this meeting:

Title: {{ json.meeting.title }}
Date: {{ json.meeting.startTime | date: "%B %d, %Y" }}
Attendees: {{ json.meeting.attendees | map: "name" | join: ", " }}

Summary:
{{ json.callRecording.transcriptSummary }}

Assistant Message (Optional)

Provides example outputs (few-shot learning)
Helps guide the expected format

liquid

Example output:
The team discussed Q4 planning. Key decisions included...

Return Types

Type	Description	Output Access
`string`	Text response	`{{ json.value }}`
`integer`	Whole number	`{{ json.value }}`
`boolean`	True/false	`{{ json.value }}`
`float`	Decimal number	`{{ json.value }}`
`string_list`	Array of strings	`{% for item in json.value %}`
`integer_list`	Array of integers	`{% for item in json.value %}`
`boolean_list`	Array of booleans	`{% for item in json.value %}`
`float_list`	Array of decimals	`{% for item in json.value %}`

AI Agent Configuration

Basic Configuration

Parameter	Type	Required	Description
model	enum	Yes	Model tier (low, medium, high)
messages	array	Yes	Messages array with role and content
tools	array	No	Tools the agent can use
return_type	enum	Yes	Output type

Available Tools

Tool	Purpose
`list`	List files in the deal room
`search`	Search deal room content
`open`	Read deal room documents
`web_search`	Search the web for information

Tool Configuration

json

{
  "tools": ["search", "open", "web_search"]
}

List Tool

Lists available files in the deal room
Useful for discovering what's available

Search Tool

Searches across deal room content
Returns relevant excerpts

Open Tool

Reads full content of specific documents
Use after searching to get details

Web Search Tool

Searches the public web
Useful for company research, industry context

Agent Instructions

When configuring an AI Agent, be explicit about tool usage:

liquid

You are a research assistant with access to deal room documents and web search.

Task: Research the prospect company and prepare a briefing.

Available tools:
- Use 'search' to find relevant documents in the deal room
- Use 'open' to read specific documents
- Use 'web_search' to find recent news and company information

Steps:
1. Search the deal room for any existing information about the company
2. Search the web for recent news and updates
3. Compile your findings into a structured briefing

Execution Modes

Per-Item Mode

Processes each input item independently:

text

[Select Many: attendees] ──▶ [AI: Generate personalized message]

Each attendee gets a separate AI call with their specific data.

Batch Mode

Processes all items together:

text

[Load items] ──▶ [AI: Summarize all items together]

All items are available in a single AI call.

When to use which:

Mode	Use When
Per-item	Each item needs individual processing
Batch	Items need to be analyzed together
Per-item	Generating individual outputs
Batch	Creating aggregate summaries

Common Configurations

Meeting Summary

yaml

model: medium
return_type: string
messages:
  - role: system
    content: |
      Summarize the meeting in 2-3 sentences.
      Focus on decisions made and next steps.
  - role: user
    content: |
      Meeting: {{ json.meeting.title }}
      {{ json.callRecording.transcriptSummary }}

Action Item Extraction

yaml

model: medium
return_type: string_list
messages:
  - role: system
    content: |
      Extract action items from this meeting.
      Return as a list of clear, actionable tasks.
      Each item should start with an action verb.
  - role: user
    content: |
      {{ json.callRecording.transcriptSummary }}

Urgency Classification

yaml

model: low
return_type: boolean
messages:
  - role: system
    content: |
      Determine if this meeting requires urgent follow-up.
      Return true if urgent, false otherwise.

      Urgent indicators:
      - Customer expressed concerns
      - Deal at risk mentioned
      - Immediate action required
  - role: user
    content: |
      {{ json.callRecording.transcriptSummary }}

Sentiment Score

yaml

model: low
return_type: integer
messages:
  - role: system
    content: |
      Rate the overall sentiment of this meeting.
      Return a score from 1-10.
      1 = Very negative, 10 = Very positive
  - role: user
    content: |
      {{ json.callRecording.transcriptSummary }}

Research with Agent

yaml

model: high
return_type: string
tools: [search, open, web_search]
messages:
  - role: system
    content: |
      You are a sales research assistant.
      Research the prospect company and create a briefing.

      Include:
      - Company overview
      - Recent news
      - Potential talking points
      - Any concerns or opportunities
  - role: user
    content: |
      Prospect: {{ json.company.name }}
      Website: {{ json.company.website }}

Error Handling

Timeout Handling

AI nodes have a 120-second timeout. To avoid timeouts:

Use transcriptSummary instead of full transcripts
Keep prompts concise
Use lower model tiers for simple tasks
Split complex tasks into multiple nodes

Parse Errors

If the AI output doesn't match the expected type:

Error is sent to the error output
Review prompt instructions
Consider adding output format examples
Use simpler return types when possible

Configuring Error Paths

Always connect error outputs for AI nodes:

text

[AI Prompt]
    ├── Success ──▶ [Use output]
    └── Error ────▶ [Fallback or notify]

Performance Tips

1. Minimize Input Size

❌ Slow:

liquid

{{ json.callRecording.transcript }}

✅ Fast:

liquid

{{ json.callRecording.transcriptSummary }}

2. Use Appropriate Model Tier

Task	Recommended Tier
Yes/no classification	Low
Simple extraction	Low
Standard summarization	Medium
Content generation	Medium
Complex analysis	High
Multi-step reasoning	High

3. Be Specific in Instructions

❌ Vague:

text

Summarize this meeting.

✅ Specific:

text

Summarize this meeting in 2-3 sentences.
Focus on: 1) Key decisions, 2) Action items, 3) Next steps.
Do not include pleasantries or small talk.

4. Provide Output Examples

liquid

System: Extract action items from this meeting.

Format each item as:
"[Owner] will [action] by [date]"

Example:
"John will send the proposal by Friday"
"Sarah will schedule the follow-up call this week"