Classifiers

A Classifier is an LLM-powered intent detection component. It analyzes user input and determines which actions should be triggered, along with any extracted parameters.

Structure

Field	Description
id	Unique identifier (within project)
projectId	Parent project
name	Display name
description	Optional description
prompt	Classification prompt template
llmProviderId	LLM provider for classification
llmSettings	LLM-specific settings
metadata	Arbitrary JSON
tags	Searchable labels for organization
archived	Whether the classifier is archived
version	Optimistic locking version

How Classification Works

During each conversation turn, classifiers run in parallel to analyze the user's input:

1. The stage's defaultClassifierId defines the primary classifier
2. Individual actions can specify overrideClassifierId to use a different classifier
3. The classifier receives:
   • The classification prompt
   • A list of available actions (filtered by conditions)
   • Each action's classificationTrigger, parameters, and examples
   • The current user input
4. The LLM returns which actions match and any extracted parameters

Classification Prompt

The prompt field is a Handlebars template that guides how the LLM classifies input. It has access to the full conversation context and the list of available actions for the current stage.

Template Context

In addition to all standard template variables, the classifier prompt receives:

Variable	Description
stage.availableActions	Array of actions eligible for classification (conditions already evaluated)
userInput	The current user's message

Each entry in stage.availableActions exposes:

Field	Description
id	Action ID — this is what the LLM must return as the JSON key
name	Display name
trigger	The classificationTrigger label describing when to fire this action
examples	Example user phrases
parameters	Parameter definitions (name, type, description, required)

Example Prompt

You are a classification assistant. Your task is to analyze user input and extract actions with parameters.

{{#if stage}}
Available actions in this stage:
{{#each stage.availableActions}}
- **{{name}}** (ID: {{id}})
  {{#if examples}}
  Examples: {{join examples ", "}}
  {{/if}}
  {{#if parameters}}
  Parameters:
  {{#each parameters}}
    - {{name}} ({{type}}){{#if required}} *required*{{/if}}: {{description}}
  {{/each}}
  {{/if}}
{{/each}}
{{/if}}

Instructions:
1. Determine the user's actions from their input using the defined actions above.
2. Extract any parameters that match the defined actions for this stage
3. For parameters extraction:
   - Only extract parameters that are explicitly mentioned or strongly implied in the user input
   - For "text" type parameters, extract the relevant text value
   - For "number" type parameters, extract numeric values
4. For action classification:
   - Prioritize defined actions when the user input matches their trigger descriptions
   - Consider the action associated with each intent when making classification decisions
   - Fall back to general actions when no specific intent matches
5. You can only use existing actions.

Required Output Format

The LLM must respond with a JSON object in the following format:

{
  "actions": {
    "<actionId>": {
      "paramName": "paramValue"
    },
    "<anotherActionId>": {}
  }
}

• Keys are action IDs (the id field from stage.availableActions), not trigger labels or action names
• The value is an object of extracted parameter key–value pairs
• Use an empty object {} when the action has no parameters (or none were extracted)
• Omit actions that were not matched — do not include them with empty objects unless they were genuinely triggered

Multiple Classifiers

A stage can effectively use multiple classifiers:

• One default classifier via defaultClassifierId on the stage
• Additional classifiers via overrideClassifierId on individual actions

When multiple classifiers are involved, they all run in parallel. Actions are distributed to their respective classifiers based on the override setting, and results are merged and deduplicated.

Classification Results

Each classifier returns:

• Matched actions — Which actions the user's input triggered
• Extracted parameters — Parameter values pulled from the user's input (name, type, value)

These results feed into the Actions & Effects execution pipeline.

Knowledge Integration

When a stage has useKnowledge enabled, knowledge categories are injected as synthetic actions into the classifier's consideration set. If the classifier matches a knowledge category, the relevant FAQ items are included in the response generation context.

Cloning

Classifiers can be cloned to create variations — for example, a specialized classifier for a particular domain that shares most behavior with a general-purpose one.