Intent Recognition with External Embedding Model - Cognigy Documentation

Documentation Index

Fetch the complete documentation index at: /llms.txt

Use this file to discover all available pages before exploring further.

An embedding is a sequence of numbers that represents the meaning of sentences and is used by the NLU for tasks such as Intent classification. The Intent Recognition with External Embedding Model feature allows using the text-embedding-3-large hosted by OpenAI and Azure OpenAI, replacing the default Cognigy embedding models.

The Azure OpenAI and OpenAI models are multilingual, and the NLU will work even if the Cognigy Flow is set in a different language than the user input. We recommend creating individual Flows for each localization.

Key Benefits

The feature is designed for complex projects with a large number of Intents and example sentences.
It enables you to set up the Intent model faster because you spend less time selecting and refining example sentences.
There is a cost associated with using third-party models. It involves making external requests, which may incur expenses and latency.

Prerequisites

Add the text-embedding-3-large model provided by Azure OpenAI or OpenAI. Note that if you use an external embedding model provider, Cognigy will send user inputs to the selected service. Ensure that your data processing policy reflects this practice.
Consider that using the text-embedding-3-large external NLU model increases Snapshot size, leading to longer download and upload times.

Set Up a Third-Party NLU model

To set up a third-party NLU model, follow these steps:

In the left-side menu of the Project, go to Manage > Settings.
Go to the NLU Settings section.
From the Embedding Model Provider list, select a model that you added as part of the Prerequisites.
Confirm changes.

Train Flows

Changing the NLU embedding model requires all Flows in the Project to be retrained. Otherwise, Intent recognition will fail because the new embeddings won’t match the old ones used for training. During Flow training, only enabled Intents are considered. Intents marked as disabled will be skipped. You can train a single Flow or multiple Flows by navigating to Build > Flows.

More Information

Last modified on April 21, 2026