NLU AI Engine - Traditional vs Hybrid
Reliable traditional NLU model vs. upgrading to the power of our new LLM-driven Hybrid NLU! 🌟
Last updated
Was this helpful?
Reliable traditional NLU model vs. upgrading to the power of our new LLM-driven Hybrid NLU! 🌟
Last updated
Was this helpful?
AI Studio’s Hybrid NLU introduces a significant advancement in language understanding by combining rule-based logic with transformer-based models from Google Gemini.
This hybrid approach enhances both intent recognition and entity extraction by leveraging contextual understanding and semantic nuance, rather than relying solely on predefined patterns. The result is a more accurate and flexible NLU system with minimal configuration required.
This guide provides step-by-step instructions for building, training, and optimizing Virtual Agents using the Hybrid NLU model.
When creating or duplicating an agent, you will see an AI Engine dropdown with two options:
Hybrid NLU 🚀
Traditional NLU
Important: Once a Virtual Agent is created, the AI Engine cannot be changed later. Choose carefully during setup. However, if you still need to change the AI engine, then you will have to duplicate/ import an existing VA and choose the appropriate AI engine.
With the Hybrid NLU, training just got a whole lot smarter! Keep these best practices in mind to get the most out of your intents -
Use full, rich sentences for training your intents to give better context.
Example:
✅ "I’d like to book a reservation for Saturday."
Don’t rely on single-word expressions unless absolutely necessary.
Example:
❌ "Reservation"
âś… Keep your user expressions meaningful but concise - quality over quantity.
❌ 10 near-identical expressions - don’t flood intents with redundant data.
Flatten your classification design - Hybrid NLU handles nuanced intent detection better without complex hierarchies.
Example: ✅ Use standalone intents like “Check account balance” instead of nested flows.
❌ "Banking > Balance > Checking" - don’t overcomplicate with rigid parent-child trees.
The New Hybrid NLU allows these methods to define custom entities:
Description
Write natural language rules about what the entity should capture. Use this when the entity has a clear structure but no predefined values.
Example:
“Order number consists of four letters, an underscore, and four digits.”
Both Description and Closed List
Provide a list of acceptable values and optional synonyms - Ideal for well-bounded value sets.
Example:
Entity: communication_type
Description: "The method someone wants to use to communicate."
Values and synonyms:
Phone → "call", "phone call", "mobile"
Fax → "facsimile", "fax message"
Email → "e-mail", "mail"
Using Closed Lists?
Toggle “Enable fuzzy matching” checkbox to leverage closed lists with low training data (i.e, synonyms for values in closed list). If this checkbox is not selected, then Hybrid NLU would try to search for exact matches of values/ synonyms provided in the closed list.
Define custom entities with a clear, natural-language Description (up to 500 characters). Think of it as writing a prompt for the AI - the more precise and detailed you are, the more accurate the extraction will be.
Example:
âś… entity name: delivery_method
✅ description: “The way the item is sent, such as mail, email, or fax.”
❌ No description or vague labels like "method" - poor prompts = poor extraction.
Use both Closed Lists and Descriptions for critical entities when maximum precision is needed.
Example: ✅ Closed list: “email, mail, fax” + description: “A method of communication.”
❌ Don’t write robotic, keyword-stuffed expressions - Hybrid NLU understands real conversation.
Test your entities thoroughly using the VA Tester after each major update.
Example: ✅ Simulate user input like “Can I get a transcript from last semester?” to verify behavior.
❌ Don’t assume the VA behaves the same as Traditional NLU - always revalidate.
Use the AI Studio Tester on the right of the canvas to:
During testing, compare the new Hybrid NLU behavior against the legacy NLU expectations: Expect better accuracy, but slightly slower response times (~100ms extra), which are imperceptible to end users.
Not ready to start from scratch? Duplicate an old VA to the new Hybrid NLU.
🚨 Important: Migrations should be tested thoroughly before going live.
Select "New" AI Engine during duplication.
Review and possibly simplify your classification nodes.
Flatten your classification design - Hybrid NLU handles nuanced intent detection better without complex hierarchies.
Add richer user expressions if your old VA relied heavily on single-word triggers.
Check out the above-mentioned best practices for intent classification with the new Hybrid NLU.
Redefine custom entities using the new "Description" field for better performance.
Traditional NLU to Hybrid NLU will show an error with custom entities, and this error will have to be resolved by adding a “Description” for custom entities.
Smarter AI needs smarter input: think about context and clarity in your expressions and entities.
The New Hybrid NLU minimizes manual effort, but a good initial design is still crucial.
Stay updated: More improvements will be rolled out, including further documentation and best practices based on community feedback.
