NLU Intents Classification with LLM

NLU intents classification with LLM enables advanced intent classification using a Large Language Model (LLM). This model is integrated into the Flow Predict Engine, enhancing intent recognition accuracy and adaptability.

NOTE: This feature is available starting with Druid 8.10 and higher. NLU intents classification with LLM currently does not support Named Entity Recognition (NER).

Instead of relying solely on manual rules, the LLM determines the correct user intent by analyzing three key flow components:

  • Flow Name
  • Description LLM (the technical instructions and context provided for the model)
    • NOTE: In Druid versions prior to 9.18, the LLM used the Description field (now renamed to Display Name) for intent classification. To ensure a seamless transition and maintain classification logic, Druid 9.18 automatically migrated the existing content from the legacy Description field into the new Description LLM section.
  • Training Phrases (the sample utterances mapped to the flow)

By automatically learning from these elements, the model reduces the manual classification effort while providing a more nuanced understanding of user queries.

Supported LLM providers

You can choose from the following large language model (LLM) providers:

  • AzureOpenAI
  • OpenAI
  • Google
  • Mistral
  • Mesolitica – MaLLaM - This LLM provider is available in DRUID 9.1 and higher.
  • AWS Bedrock - This LLM provider is available in DRUID 9.5 and higher. Contact your DRUID representative to activate it on your tenant.

DRUID-dedicated LLM resources

  • Druid Becus 3.0 / 1.0 (Proprietary LLM)
  • Azure OpenAI - gpt-4o-mini

  • AwsBedrock - mistral-large-2407-v1:0

  • Google Vertex AI

If you want to use DRUID-dedicated LLM resources, contact your sales representative to activate them for your tenant.

How It Works

When NLU intents classification with LLM is configured, the system leverages a LLM for intent classification using two key prompts:

  • System prompt: Instructs the LLM to classify a user query by scoring provided intents based on relevance, adhering to strict JSON formatting and predefined scoring rules while considering both user-supplied and system intents.
  • User prompt: Provides the intent list and user query, ensuring the model has the necessary context for classification.

Configure NLU Intents Classification with LLM

NOTE: Before you start, make sure you have an LLM resource set up in the Druid Portal. For more information, see Create LLM Resources.

To configure NLU intents classification with LLM, follow these steps:

  1. Navigate to NLU > Configurations > Intents tab.
  2. Click on Thresholds and parameters.
  3. Add the NLU.NER.Classification.ModelType parameter and set it to LLM.

    4. Click the link that appears in the message below the NLP parameters.

    The NLU Intents Classification with LLM expands.

  4. From the Endpoint Type field, select one of the following options:
    • Druid if you have an LLM subscription with Druid.
    • My resource if you are using your own generative endpoint.
  5. Select the Model Name. In the dropdown, you will see all your resources created in Administration > LLM Resources.
  6. The Api Type field appears dynamically in Druid 9.19 and higher when an Azure OpenAI model is selected (e.g., AzureOpenAi/gpt-4o-mini). Select one of the following:
    • Responses: Recommended for Azure OpenAI models. This is the preferred API for Azure OpenAI models, enabling better performance and more intelligent interactions.
    • Chat Completions (Legacy): Use this for backward compatibility with existing configurations. Note that this endpoint is now considered legacy.

    HINT: For other providers, the platform defaults to the standard completion settings compatible with those models.
  7. Select Use Knowledge Base Results to allow the LLM to use Knowledge Base data for intent classification. When enabled, the system queries the Knowledge Base with the user message and provides the results to the LLM to help identify the most relevant intent.
  8. Select Disable Ssl Validation when you configure your own LLM resource and the endpoint uses a self-signed certificate.
  9. For both prompts, click the Set default button. The default Druid prompts will be automatically filled in.
    10. HINT: Since this feature is in technology preview, you can experiment with the prompts, provided their JSON structure remains unchanged. In an upcoming release, the prompts will become non-editable.
  10. Click Save to apply the NLU configuration.
  11. Train the bot.

To configure NLU intents classification with LLM, follow these steps:

  1. Navigate to NLU > Configurations > Intents tab.
  2. Click on Thresholds and parameters.
  3. Add the NLU.NER.Classification.ModelType parameter and set it to LLM.

    4. Click the link that appears in the message below the NLP parameters. The NLU Intents Classification with LLM expands.

  4. From the Endpoint Type field, select Druid if you have a LLM subscription with DRUID or Custom if you're using your own generative endpoint.
  5. Select the Client Type.
  6. For custom endpoints, enter the URL of the generative API.
  7. Provide the secret key generated in your generative account.
  8. Specify the generative model name.
  9. If you selected Google as Client Type, enter the Location of the location for a Vertex AI and your Google Project Id.
    10. NOTE: Google Vertex AI is available in DRUID 8.13 and higher.
  10. For both prompts, click the Set default button. The default DRUID prompts will be automatically filled in.
    11. HINT: Since this feature is in technology preview, you can experiment with the prompts, provided their JSON structure remains unchanged. In an upcoming release, the prompts will become non-editable.
  11. Click Save to apply the NLU configuration.
  12. Train the bot.

Once configured, the model uses the two prompts to classify user intents automatically.

IMPORTANT! When using NLU intents classification with LLM, it’s important to separate natural language training phrases from technical commands. To avoid introducing noise into intent detection, move any technical training phrases—such as exact keywords, system triggers, or command-like inputs — into the Commands section of the respective flow. These phrases will then be matched exactly, without affecting the NLP model's understanding of user intent. Commands are available in DRUID 8.14 and higher.