LLM Node

The LLM node calls a language model with a configured prompt, optionally using record fields via template syntax, and appends the model's response to each record. It can also call tools, but does not run a multi-step reasoning loop — for that, use the Agent node.

When to Use

Use LLM when you need to generate text (summaries, emails, descriptions, classifications) or structured data (JSON extraction, scoring) for each record in your dataset.

Configuration

Credentials

Leave credentialId blank to use the platform-managed provider key (subject to your plan; frontier models may require a paid plan). Set it to a specific credential ID to use your own API key.

Example

Generate a personalized email subject line per lead:

provider: anthropic
model: claude-sonnet-4-6
systemPrompt: "You write concise, engaging B2B email subject lines."
userPromptTemplate: "Write a subject line for an email to {{first_name}} at {{company_name}}, a {{industry}} company."
outputMode: text
outputColumn: email_subject
temperature: 0.8
maxTokens: 50

Structured Output Example

Extract company metadata as JSON:

outputMode: json_schema
outputSchema:
  type: object
  properties:
    industry: { type: string }
    employee_range: { type: string, enum: ["1-10", "11-50", "51-200", "201-1000", "1000+"] }
    is_b2b: { type: boolean }
  required: [industry, employee_range, is_b2b]

Notes

• Each record generates one LLM call — costs scale with dataset size.
• LLM runs are logged to the audit trail; confidence scoring applies.
• Token limits include the rendered prompt — large record fields injected via templates count toward the limit.
• json_schema mode uses structured outputs where available; falls back to JSON parsing with validation.
• Failed calls (rate limits, timeouts) are retried with exponential backoff; records that exhaust retries are flagged in the run log.