How to fine-tune intents

Learn how to fine-tune agent intents in AI Chatbot Hub to enhance chatbot performance and ensure accurate intent classification.

When your chatbot includes multiple user-facing Agents, AI Chatbot Hub’s AI Supervisor automatically selects the best Agent for each user query. It does this by analyzing recent chat history to understand user intent and then checks each Agent’s name, description, and prompt for relevance.

In our Multi-Agent Chatbot guide, we discuss best practices for writing Agent descriptions and provide examples. Despite the AI Supervisor’s efforts, some situations may still arise where Agent descriptions don’t fully cover all practical use cases.

When user-facing Agents are created or modified, a set of intents and sample user queries are generated automatically to guide the AI Supervisor. You can review, customize, or delete these intents through the AI Supervisor Settings interface, as they are key to the routing algorithm.

To fine-tune intents, go to AI Supervisor Settings -> Intents:

Here, you’ll see intents generated by the AI Supervisor based on its interpretation of each Agent’s role. When a user submits a query, if multiple user-facing Agents are active, the AI extracts the user’s intent and matches it to the closest Agent intent. The selected Agent (without Overrides) will be the one with the closest matching intent.

To add a new intent, select the desired Agent, click “Add intent,” and define the intent with example user queries.

Click “Create” to see the intent added under “Fixed Intents.” Scroll down if needed to view the complete list. You can also edit AI-generated intents by selecting “Edit” from the three-dot menu next to the intent. Once saved, the edited intent moves to “Fixed Intents.”

AI-generated intents are refreshed each time you update the Agent name, description, or prompt, whereas Fixed Intents remain until manually deleted.

When to fine-tune agent intents

For most general Q&A setups, fine-tuning isn’t necessary, as AI-generated intents generally suffice. However, controlling intent classification at a granular level allows for enhanced performance, particularly with multiple user-facing Agents or high-precision requirements.

For multi-agent setups, we recommend following a “MECE” (Mutually Exclusive, Collectively Exhaustive) approach, ensuring no overlaps and no gaps in intent coverage.

While it’s impossible to cover all potential intents and natural language variations, you can refine your chatbot’s intent classification to achieve high accuracy.

To monitor which Agents are selected for each query, use “Debug mode.” With multiple user-facing Agents, you’ll see “Intent” and “Active Agent” next to each user query that underwent routing.

If the wrong Agent was selected, add the intent and corresponding user query to the correct Agent in AI Supervisor Settings -> Intents. This update ensures that similar future queries are routed correctly. Adding more examples to an intent will further improve the AI Supervisor’s ability to classify similar cases accurately.