LLM

The LLM node sends a prompt to a Large Language Model to generate, transform, or analyze text and returns the response.

What does the LLM node do?

The LLM node sends a prompt to a Large Language Model (GPT, Claude, Gemini, Mistral, and other configured providers) and returns the model response as a string. It is the foundation block for any AI-driven step in a workflow: generation, transformation, analysis, classification, structured extraction.

Common use cases:

• Generate content (articles, summaries, emails, ad copy) from variables coming from upstream nodes.
• Transform text (rewrite, translate, reformat HTML/Markdown, normalize tone).
• Extract structured information from unstructured text using an Output JSON Schema.
• Classify or score input data with a deterministic prompt and low temperature.
• Add a reflection loop so the model critiques and iteratively improves its own answer before returning.

Quick setup

Follow these steps to add and configure the LLM node:

Configuration parameters

Configuring the LLM node combines three things: the model identity, sampling parameters, and the instructions sent at runtime.

Required fields

Optional fields

What does the node output?

The node returns the model response as plain text. The execution engine also records a conversation trace (system message, prompt, AI response, reasoning tokens when applicable) for observability — exposed in run details, not in the downstream variable.

How to use the output

In Draft & Goal you do not need to look up a system-generated variable name:

1. Draw a connection from the LLM node’s output port.
2. Connect it to the next node’s input.
3. In that next node, create and name your own variable (for example, summary or extracted_json). The model response is injected automatically.

Usage examples

Example 1: Summarize an article into bullet points

Use a low temperature for consistent output and a clear instruction.

Prompt:

Summarize the following article in 3 bullet points.
Each point must be under 20 words.
Focus on actionable takeaways for marketers.

Article:
{{content}}

Settings:

• Temperature: 0.2
• Max output tokens: 4000
• System Message: You are a senior content strategist. Be concise.

Expected output:

- AI tools cut content production time by half.
- Automation works best on repetitive editing tasks.
- Integration with the existing CMS removes manual export steps.

Example 2: Extract structured data with an Output JSON Schema

Combine a precise prompt with Output JSON Schema to get parseable JSON for downstream nodes.

Prompt:

Extract the following information from this company page.

Page:
{{content}}

Output JSON Schema:

{
  "company_name": "string",
  "industry": "string",
  "employee_count": "string",
  "products": ["string"]
}

Expected output:

{
  "company_name": "Acme Technologies",
  "industry": "SaaS",
  "employee_count": "50-100",
  "products": ["Project Management", "Time Tracking"]
}

Pipe this output into a JSON Path Extractor to read individual fields, or into a Find and Replace node to strip residual Markdown fences.

Common issues

The workflow fails with `LLM Instructions is missing !`

Cause: The Prompt field is empty when the node runs.

Solution: Open the node settings, write the instructions in the prompt editor, and save. Even a one-line prompt is enough to make the node valid.

The response is cut off mid-sentence

Cause: Max output tokens is too low for the requested generation.

Solution: Increase Max output tokens. As a rule of thumb, count ~1.3 tokens per English word and add headroom. For thinking models, also raise Thinking budget tokens so reasoning does not eat the response budget.

The model returns invalid or wrapped JSON

Cause: Without a schema, models often wrap JSON in Markdown fences (```json) or add commentary.

Solution: Fill the Output JSON Schema field, lower the Temperature (0.0–0.3), and add Return only valid JSON, no Markdown to the prompt. Place a Find and Replace node after the LLM if stray fences remain.

The output is too random or inconsistent across runs

Cause: Temperature (or Top P) is too high, so the model samples broadly.

Solution: Lower Temperature to 0.2–0.4 for analytical tasks, or reduce Top P toward 0.5. Be specific in the prompt: state the exact format, length, and style expected.

The selected provider or model is missing from the dropdown

Cause: The provider’s API key is not connected to your organization, or the feature flag for the model is off.

Solution: Open Settings → Integrations and connect the provider, then reload the workflow editor. Reach out to your workspace admin if the model is gated.

Best practices and pitfalls

How does it fit into a workflow?

The LLM node typically sits between an input/scraping step that gathers context and a downstream node that consumes the response.

graph LR
    Scraper[Web Scraper] --> LLM[LLM
generates JSON]
    LLM --> FR[Find and Replace
strips Markdown]
    FR --> Extractor[JSON Path Extractor]
    Extractor --> Sheets[Google Sheets Writer]

Related nodes