The Intent Classification node intelligently matches input text against configured intent descriptions and routes execution to the corresponding branch. By configuring multiple intents, the model evaluates the input and selects the most appropriate intent(s), then executes the associated downstream path(s). It supports both single-selection and multi-selection modes, making it suitable for intelligent customer service, intent routing, and multi-scenario processing.
Usage Example
The Intent Classification node automatically identifies the type of user query and routes it to the appropriate processing branch. For example, in an intelligent customer service scenario, it routes queries about product inquiries, after-sales support, or complaints/suggestions to different specialized large language model (LLM) nodes for targeted handling.
Adding the Node
Method 1: Add via the + Button
-
Hover over any existing node on the canvas (e.g., the Start node); a
+button appears automatically on its right side. -
Click the
+button, then select Logic > Intent Classification from the pop-up node list. The node is added to the canvas and automatically connected to the upstream node.
Method 2: Drag from the Node Library
- In the left-hand node library, locate and select Logic > Intent Classification.
- Drag the node onto the canvas at your desired location, then drag a connection line from the edge of an upstream node to the Intent Classification node.
Core Configuration
Input Variable
Specifies the variable(s) this node will process to determine the intent.
| Parameter | Required | Description |
|---|---|---|
| Input Variable | Yes | Input text or variable reference used for intent classification. |
Model Configuration
Selects the model used to perform intent classification. Supported models include dedicated intent classification models, Qwen series models (e.g., Qwen-Max), and third-party models (e.g., DeepSeek). Refer to the console for the currently available models.
Click the gear icon next to the model selector to enable advanced configuration and set the following parameters:
| Parameter | Default | Description |
|---|---|---|
| Max Response Length | 1024 | Maximum number of tokens the model may output. |
| top_p | 0.80 | Controls output diversity — higher values yield more varied results. |
| temperature | 0.70 | Controls output randomness — higher values yield more stochastic outputs. |
| enable_thinking | Enabled | Enables deep reasoning mode; outputs step-by-step rationale. |
| thinking_budget | 4000 | Maximum number of tokens allocated to chain-of-thought reasoning. |
| result_format | message | Format of the returned result. |
| enable_search | Disabled | Enables web search capability. |
Intent Definitions
Define the intents the model should recognize. Each intent includes a descriptive label that guides matching. Click Add Intent to configure new intents. Up to 60 intents are supported. An “Other” fallback branch is included by default and executes when none of the defined intents match the input.
Intent Mode
Determines whether the model returns one intent (single selection) or multiple intents (multi-selection).
| Mode | Description | Use Case |
|---|---|---|
| Single Selection (Default) | Selects the best-matching intent among all options. | Simple scenarios where each input maps clearly to exactly one intent. |
| Multi-Selection | Returns all matching intents. | Complex scenarios where a single input may satisfy multiple intents. |
Reasoning Mode
Controls whether the model outputs its internal reasoning process during classification. This is a node-specific setting distinct from the enable_thinking parameter in LLM node model configurations.
| Mode | Reasoning Output | Use Case |
|---|---|---|
| Fast Mode (Default) | No reasoning output; thought field is empty. | Simple intent classification or latency-sensitive applications. |
| Accuracy Mode | Outputs full step-by-step reasoning; thought contains detailed logic. | Complex intent classification requiring high accuracy. |
Memory
Disabled by default. When enabled, the model retains conversation history to support multi-turn interactions.
| Cache Type | Description |
|---|---|
| Node-local Cache | Model remembers only the dialogue history within this node. Configure memory turns (1–50, default: 3); one input-output pair counts as one turn. |
| Custom Cache | Model retains global context. Requires selecting a context variable (e.g., built-in historyList). |
Prompt Instructions
Provides additional instructions or constraints to guide the intent classifier. You can add explicit examples or stricter rules to improve alignment between model output and expected intent categories.
| Parameter | Required | Description |
|---|---|---|
| Prompt | No | Accepts plain text or references to upstream node variables. Providing concrete classification examples helps steer similar inputs toward the correct intent. |
Node Outputs
The Intent Classification node delivers two types of outputs: it acts both as a control node for intelligent branching and as a data node passing classification results downstream.
Branch Routing
Execution automatically follows the matched intent(s):
- Single Selection Mode: Executes only the branch corresponding to the top-matching intent.
- Multi-Selection Mode: Executes all branches corresponding to matched intents.
- No Match: Executes the “Other” fallback branch.
Variable Outputs
Outputs structured intent classification results for downstream use—e.g., further logic evaluation or prompt engineering in subsequent LLM nodes.
Single Selection Mode:
| Output Variable | Type | Description |
|---|---|---|
result | Object | Intent classification result object. |
result.subject | String | Name of the matched intent (as defined in configuration). |
result.thought | String | Reasoning trace: empty in Fast Mode; detailed in Accuracy Mode. |
| Output Variable | Type | Description |
|---|---|---|
result | Object | Intent classification result object. |
result.subjects | Array<String> | Array of all matched intent names. |
result.thought | String | Reasoning trace (same behavior as above). |
Common Patterns
The Intent Classification node serves dual roles: as a control node enabling intelligent routing, and as a data node propagating classification outcomes.
| Paired Node(s) | Description |
|---|---|
| Intent Classification → Multiple LLM Nodes | Route to different LLM nodes based on intent — e.g., product inquiry → Product Expert LLM. |
| Intent Classification → Intermediate Node → Variable Processor (Aggregate Group Mode) | Use Aggregate Group mode in the Variable Processor to automatically collect non-null outputs across parallel intent branches, preventing null-related errors downstream. |
Frequently Asked Questions
How do I improve intent classification accuracy?
- Refine intent descriptions: Keep them concise, unambiguous, and mutually exclusive.
- Add prompt instructions: Include representative examples and explicit constraints to guide classification.
- Choose the right mode: Use Accuracy Mode for complex cases; Fast Mode for simple, latency-critical ones.
- Select an appropriate model: Intent-specific models or Qwen-Max are strongly recommended.