Classifier Component

Use the Classifier component to route unstructured input (for example user messages, logs or email content) into clear, predefined categories using a prompt powered by a language model.

Why this matters

Instead of writing complex if-else logic, you can use this component to classify text into buckets such as “support”, “sales” or “urgent” based on simple natural-language descriptions.

What You’ll Configure

• Choose a Model
• Write or Update the Prompt
• Add Inputs
• Define Output Categories
• Advanced Routing (Optional)
• Debug and Preview
• Best Practices
• Troubleshooting Tips
• What to Try Next

Step 1 - Choose a Model

Field	Required?	Description	Tips
Model	Yes	The LLM used for classification	By default you will see GPT 4o Mini (SmythOS 128K). You can also choose GPT 4.1 Nano, GPT 4.1 Mini, GPT o1, Claude Sonnet 3.7, Gemini 2.5 Pro, Sonar or other supported engines.
Custom Model	No	Use your own hosted model or external API	You can connect to OpenAI, Claude, Google AI, Grok, or any endpoint compatible with your workflow.

Token limits

Default models have token windows from 128K to 1M tokens (shown next to each model name). If you need to classify very large inputs or include extra context, use a custom model with extended limits.

Step 2 - Write or Update the Prompt

Default Classification Prompt

classifierroutingprompt

Classify the {{Input}} content to one of the categories. Set the selected category to true and the others to an empty value.

Make prompts specific

You can add context such as examples of each category or definitions of what qualifies for a category. For instance, you might write:

Specific Classification Prompt Example

classifierexample

Classify the {{Input}} text as 'support' if it mentions troubleshooting or technical issues, 'sales' if it talks about pricing or upgrades, 'marketing' if it refers to promotions or events, and so on. Return "true" under the matching category and leave other categories blank.

Step 3 - Add Inputs

Field	Required?	Description	Notes
Input Name	Yes	A descriptive label for the input, for example user_message or email_body.	This name becomes a variable reference in your prompt (use {{InputName}}).
Input Type	Yes	The data type of the input value. Options typically include Text, Number, Boolean, Any etc.	Choose Any if you do not need type‐specific validation.

INFO

You must have at least one primary input that the model will classify. You can optionally add more inputs (for example sender, channel or metadata), but they are not strictly required to drive classification.

Advanced Options for Each Input

1. Required / Optional

   • Choose whether this input must be provided (Required) or can be skipped (Optional).
   • If you set it to Optional, and no value is provided at runtime, the model will receive an empty value for that input variable.

2. Color Label

   • Click the color swatch to pick a hex code (for example #F35063) that helps visually organize inputs on the workflow canvas.
   • This is purely for organizational clarity... colors do not affect model behavior.

3. Default Value

   • If no value is supplied by the user or upstream component, the model will use whatever you enter here.
   • Useful if you want to ensure the prompt always has at least a placeholder (for example "No message provided").

Step 4 - Create Outputs

Each Output corresponds to one classification bucket. The model will assign exactly one bucket (set to true) and leave all others blank. By default, when you first open the Classifier node, no categories are defined... you must click + Outputs to begin defining them.

Field	Required?	Description	Example
Name	Yes	A unique identifier for this category.	support, sales, critical_alert
Description	No	A short explanation of what content belongs in this bucket.	“Password reset requests or technical issues.”
Expression	No	(Advanced) A custom expression that can post‐process or filter which messages go here. Leave blank to rely purely on the LLM.	For example, you could write a JSONPath or Liquid expression to extract a field or to validate the model’s output.
Color	No	Pick a hex code (for example #3C89F9) to color-code this output on the canvas. Helpful when you have many buckets.	Use contrasting colors for easy identification.

Example: Adding a New Category

1. Click + Outputs on the Classifier card to open the Add Output dialog.
2. In Output Name, type support.
3. (Optional) In Description, type “Customer support requests, including troubleshooting or technical issues.”
4. (Optional) In Expression, leave blank if you don’t need extra filtering logic.
5. Click the color swatch and choose a hex value like #3C89F9.
6. Click Create.

7. You will now see “support” appear in the Categories list on the Classifier node. It will use the color you picked. Repeat these steps to add as many categories (outputs) as you need.

Example Output Configuration (JSON)

[
  {
    "Name": "positive_sentiment",
    "Color": "green",
    "Description": "Text that shows appreciation, praise or good feedback"
  },
  {
    "Name": "support",
    "Color": "#3C89F9",
    "Description": "Customer support inquiries or help requests"
  },
  {
    "Name": "escalation",
    "Color": "red",
    "Description": "Urgent or high-risk issues requiring immediate attention"
  }
]

Step 5 - Advanced Routing (Optional)

You can configure each category’s Format or attach downstream logic by using custom expressions, Webhook calls or conditional fields. For example, you might want the “sales” bucket to forward message content to a CRM, while “engineering” sends technical logs to a different queue.

Example: Return Different Formats per Category

[
  {
    "Name": "sales",
    "Description": "Routes message to the sales inbox",
    "Format": "text"
  },
  {
    "Name": "engineering",
    "Description": "Forwards technical messages to the dev team",
    "Format": "json"
  },
  {
    "Name": "marketing",
    "Description": "Extracts campaign name from the input",
    "Format": "text"
  }
]

TIP

You can reference parts of the input or return fixed values by writing Liquid or JSONPath expressions under Expression for that output. This is useful if you need to extract a substring, call an API or reshape the response.

Step 6 - Debug and Preview

Live testing

1. Click Debug on the Classifier node.
   2. Paste or type a sample message to see how the model routes your input.

Example Input:
“Hi, I’m trying to reset my password but the link isn’t working.”

Configured Outputs:

[
  { "Name": "support", "Description": "Customer support requests" },
  { "Name": "sales", "Description": "Product pricing or upgrade inquiries" },
  { "Name": "marketing", "Description": "Messages about promotions or events" }
]

Expected Result: The LLM should mark the support category as true and leave the other outputs empty. You will see this result returned in the debug console along with any token usage details.

Best Practices

• Use detailed, mutually exclusive descriptions for each output
• Avoid vague labels like "misc" or overlapping categories like “help” and “support”
• Always test edge cases in Debug
• Use fallback categories if you expect unpredictable input
• Stick to consistent naming like snake_case or kebab-case
• Validate using real user data if possible

Troubleshooting Tips

If classification doesn’t behave as expected

• Double-check that your output descriptions are specific and non-overlapping
• Review the prompt: vague instructions lead to unpredictable classification
• Confirm your input field is correctly referenced as {{Input}} in the prompt
• Test with Debug to see the full model output and how it matched categories
• If all outputs are skipped, add a fallback category or simplify inputs
• Ensure the model you’re using supports the prompt + input token count

What to Try Next

• Combine with Agent Skill to classify natural user queries
• Store classifications in RAG Remember to build memory or logs

TIP

Want to go deeper? Try chaining Classifier with a GenAI LLM to auto-reply to each category differently.