Skip to content

Conditional Routing with Switch Nodes

Intermediate

In this tutorial, you will build a workflow that classifies incoming messages and routes them to different specialized agents. A customer support scenario is used as the example: messages are categorized as billing, technical, or general, and each category is handled by a dedicated agent with domain-specific instructions.

Time: 20 minutes

What you will build:

flowchart LR
    CT[Chat Trigger] --> Cat[Categorizer]
    CatModel[AI Model<br/>gpt-4o-mini] -.->|llm| Cat
    Cat --> SW[Switch]
    SW -->|billing| BA[Billing Agent]
    SW -->|technical| TA[Technical Agent]
    SW -->|general| GA[General Agent]
    BAModel[AI Model] -.->|llm| BA
    TAModel[AI Model] -.->|llm| TA
    GAModel[AI Model] -.->|llm| GA

    style CT fill:#f97316,color:white
    style CatModel fill:#3b82f6,color:white
    style BAModel fill:#3b82f6,color:white
    style TAModel fill:#3b82f6,color:white
    style GAModel fill:#3b82f6,color:white
    style SW fill:#3b82f6,color:white

Prerequisites

Concepts

Before building, here is how conditional routing works in Pipelit:

  1. A Categorizer node uses an LLM to classify input text into one of several predefined categories.
  2. The Categorizer outputs a _route value matching the category name.
  3. A Switch node reads the route value from the workflow state.
  4. Conditional edges on the Switch node direct execution to different downstream nodes based on the route value.

Only the branch matching the current route executes -- the other branches are skipped.

Step 1: Create the workflow

  1. On the Dashboard, click New Workflow.
  2. Name it Support Router and click Create.

Step 2: Add a Chat Trigger

  1. From the Node Palette under Triggers, click Chat.
  2. A Chat Trigger node appears on the canvas.

Step 3: Add a Categorizer

The Categorizer uses an LLM to classify the incoming message into one of your defined categories.

  1. From the Node Palette under AI, click Categorizer.
  2. Click on the Categorizer node to select it.

  3. In the Node Details Panel, configure the categories in Extra Config:

    {
  "categories": [
    {
      "name": "billing",
      "description": "Questions about invoices, payments, refunds, pricing, or subscription plans"
    },
    {
      "name": "technical",
      "description": "Bug reports, errors, feature requests, integration help, or API questions"
    },
    {
      "name": "general",
      "description": "Greetings, general inquiries, feedback, or anything that does not fit the other categories"
    }
  ]
}

  4. Optionally, add a System Prompt to provide classification guidance:

    You are classifying customer support messages for a SaaS platform.
Be precise -- questions about feature limits should go to "technical",
not "billing". When in doubt, choose "general".

Connect an AI Model to the Categorizer

  1. Add an AI Model node from the Node Palette.
  2. Configure it with your credential and a model. A smaller, cheaper model like gpt-4o-mini works well for classification since the task is straightforward.
  3. Connect the AI Model's top diamond handle to the Categorizer's "model" diamond handle (blue).

Use a cheaper model for classification

The Categorizer makes a single LLM call to classify input. A smaller model reduces cost and latency without sacrificing classification quality for well-defined categories.

Step 4: Add a Switch node

The Switch node reads the route value set by the Categorizer and directs execution down the matching conditional edge.

  1. From the Node Palette under Routing (or Logic), click Switch.
  2. A Switch node appears on the canvas.

The Switch node does not require configuration for this use case -- it reads the _route value from the workflow state, which the Categorizer sets automatically.

Step 5: Add specialized agents

Create three agent nodes, one for each category.

Billing Agent

  1. Add an Agent from the Node Palette.

  2. Set its System Prompt:

    You are a billing support specialist. Help customers with:
- Invoice questions and payment issues
- Subscription plan details and upgrades
- Refund requests and billing disputes
- Pricing clarification

Be empathetic and solution-oriented. If the issue requires account
access you do not have, escalate to a human agent.

  3. Add an AI Model node and connect it to this agent's model handle.

Technical Agent

  1. Add another Agent.

  2. Set its System Prompt:

    You are a technical support engineer. Help customers with:
- Bug reports and error troubleshooting
- API integration questions
- Feature requests (acknowledge and log them)
- Configuration and setup assistance

Ask clarifying questions when the issue is ambiguous. Include
specific steps in your solutions.

  3. Add an AI Model node and connect it.

General Agent

  1. Add a third Agent.

  2. Set its System Prompt:

    You are a friendly general support agent. Handle:
- Greetings and casual conversation
- General product inquiries
- Feedback and suggestions
- Anything that does not fit billing or technical categories

Be warm and helpful. Redirect to specialized support if the
conversation turns technical or billing-related.

  3. Add an AI Model node and connect it.

Sharing AI Model nodes

Each agent needs its own AI Model connection. You can use the same credential and model across all of them, but each requires a separate AI Model node on the canvas.

Step 6: Connect the nodes

Data flow connections

  1. Chat Trigger -> Categorizer: Right handle to left handle.
  2. Categorizer -> Switch: Right handle to left handle.

Conditional edges from the Switch

Conditional edges are created differently from regular edges. Each conditional edge carries a condition_value that must match the route.

  1. Switch -> Billing Agent: Drag from the Switch's right handle to the Billing Agent's left handle. When the edge creation dialog appears, set the condition value to billing.
  2. Switch -> Technical Agent: Same process, with condition value technical.
  3. Switch -> General Agent: Same process, with condition value general.

Conditional edge labels

On the canvas, conditional edges display their condition value as a label on the edge (e.g., "billing", "technical", "general"). This makes it easy to see the routing logic at a glance.

The complete workflow

flowchart LR
    CT[Chat Trigger] --> Cat[Categorizer]
    CM[AI Model] -.->|llm| Cat
    Cat --> SW[Switch]
    SW -->|billing| BA[Billing Agent]
    SW -->|technical| TA[Technical Agent]
    SW -->|general| GA[General Agent]
    BM[AI Model] -.->|llm| BA
    TM[AI Model] -.->|llm| TA
    GM[AI Model] -.->|llm| GA

    style CT fill:#f97316,color:white
    style CM fill:#3b82f6,color:white
    style BM fill:#3b82f6,color:white
    style TM fill:#3b82f6,color:white
    style GM fill:#3b82f6,color:white
    style SW fill:#3b82f6,color:white

Step 7: Validate and test

Validate the workflow

Before testing, validate the workflow to catch any configuration errors:

  1. Click the Validate button (or use POST /api/v1/workflows/{slug}/validate/).
  2. Fix any reported issues (e.g., missing model connections, unconnected nodes).

Test with different messages

Open the Chat panel and try messages that should route to each category:

Billing message:

I was charged twice on my last invoice. Can you help me get a refund?

Expected: Categorizer outputs billing -> Switch routes to Billing Agent.

Technical message:

I am getting a 500 error when I try to use the API endpoint for user creation.

Expected: Categorizer outputs technical -> Switch routes to Technical Agent.

General message:

Hi there! I just wanted to say I love your product.

Expected: Categorizer outputs general -> Switch routes to General Agent.

Observe execution on the canvas

As each message is processed:

  1. The Chat Trigger lights up (running -> success).
  2. The Categorizer lights up and makes an LLM call to classify the message.
  3. The Switch node lights up and evaluates the route.
  4. Only the matching downstream agent lights up -- the other two stay idle.

Click the green "output" link on the Categorizer node after execution to see the classification result:

{
  "category": "technical",
  "raw": "{\"category\": \"technical\"}"
}

Variations

Using a Switch node with custom rules

Instead of using a Categorizer (which requires an LLM call), you can configure the Switch node with pattern-matching rules for simpler routing:

{
  "rules": [
    {"field": "text", "operator": "contains", "value": "invoice", "route": "billing"},
    {"field": "text", "operator": "contains", "value": "refund", "route": "billing"},
    {"field": "text", "operator": "contains", "value": "error", "route": "technical"},
    {"field": "text", "operator": "contains", "value": "bug", "route": "technical"}
  ],
  "default_route": "general"
}

This approach is faster and cheaper (no LLM call) but less flexible than LLM-based classification.

Adding a default/fallback route

If the Categorizer returns a category that does not match any conditional edge, the Switch node falls back to a default route. Add a conditional edge with condition value default to handle unrecognized categories:

flowchart LR
    SW[Switch] -->|billing| BA[Billing Agent]
    SW -->|technical| TA[Technical Agent]
    SW -->|default| FA[Fallback Agent]

Combining with conversation memory

Enable Conversation Memory on each specialized agent so they maintain context within their domain. A user who starts with a billing question and follows up with more details will get a continuous conversation thread with the Billing Agent.

Next steps