Intent Detection Using Semantic Search

Where to find it in Automate

To use Semantic Search for intent classification in Automate:

1. Open your bot or assistant.
2. Go to NLU Settings for your Project.
3. Select Semantic Search.
4. Choose a model from the list.
5. Train and test your assistant.

For most users, this is the only part that matters: choose Semantic Search, then select the model that best fits the use case.

How to choose a model

The available built-in models are:

• auto
• MultilingualE5Small
• MultilingualE5Base
• BGEM3
• ParaphraseMLMiniLML12V2

They all serve the same purpose, but they differ in speed, quality, and typical use case.

auto

This is the easiest option when you do not want to choose a model manually.

• Compares the built-in models during training
• Selects the option that works best for the current dataset
• May also choose the most suitable internal classification strategy
• Usually takes longer to train than selecting one fixed model directly

Best for:

• new assistants
• teams that want the safest default
• use cases where it is not yet clear which model performs best
• projects where quality matters more than training speed

MultilingualE5Small

This is the safest starting point for most teams.

• Fast
• Good general quality
• Works well in multilingual environments
• Practical for everyday chatbot and routing scenarios

Best for:

• new assistants
• high-volume traffic
• teams that want a reliable default

MultilingualE5Base

This is a balanced option when you want more quality than the small model.

• Better quality in many cases than MultilingualE5Small
• Still suitable for general production use
• Slightly heavier and slower

Best for:

• general customer support use cases
• assistants where understanding quality matters more than raw speed
• teams that want a step up from the basic default

BGEM3

This is the quality-focused option.

• Often the strongest option for harder classification tasks
• Useful when intents are similar to each other
• Slower and heavier than lighter models

Best for:

• more complex assistants
• important business flows where classification quality is critical
• cases where a slightly slower response is acceptable

ParaphraseMLMiniLML12V2

This is the lightweight option.

• Fast and efficient
• Good for short user messages
• Useful when you want a smaller, simpler model

Best for:

• smaller deployments
• short phrase classification
• cases where low resource usage is important

Recommended starting point

If you do not know which model to choose:

• start with auto if you want the system to choose for you
• start with MultilingualE5Small
• move to MultilingualE5Base if you want more quality
• try BGEM3 if quality is the highest priority
• try ParaphraseMLMiniLML12V2 if you want a lighter option

What happens in auto

When a user selects auto, the system does extra work during training.

In simple terms:

• it tests the supported built-in models
• it compares which option fits the current training data best
• it saves the best-performing result for later use

This means auto is useful when:

• you want the system to make the choice for you
• you are starting a new assistant
• you want to optimize for quality without manually testing models one by one

The main tradeoff is simple:

• auto is easier for the user
• training usually takes longer than with one fixed model

Classification strategies

You do not choose the strategy directly in Automate, but it helps to understand what it means.

In practice:

• when you choose a specific model from the list, the system uses knn
• when you choose auto, the system may internally use centroid or hybrid if that gives better results

knn

This means: "find examples most similar to the current message."

In simple business terms:

• the system compares the new message to known training examples
• then it chooses the intent that looks most similar

This is the standard behavior when a user selects one of the built-in models directly.

centroid

This means: "compare the message to the overall profile of each intent."

In simple business terms:

• the system looks at the general shape of each intent
• then it chooses the closest topic

hybrid

This means: "first shortlist the most likely topics, then compare more precisely."

In simple business terms:

• the system first narrows down the possible intents
• then it makes a more detailed comparison

Good practice for business users

Model choice matters, but training data matters even more.

Good practice:

• keep intent names clear and distinct
• use realistic user phrases
• avoid heavy overlap between intents
• add enough examples for each intent
• test the assistant on real user questions after training

Practical recommendation

For most business teams, the simplest approach is:

1. Select Semantic Search in NLU Settings.
2. Start with auto if you want the system to choose for you, or MultilingualE5Small if you want a fast and predictable default.
3. Train the assistant.
4. Test real user messages.
5. If quality is not good enough, try MultilingualE5Base or BGEM3.