Why we got rid of intent classifiers

One of the most common questions we get from potential customers and users is “Where are your intent classifiers?”

We don’t have any. And yeah, it’s on purpose. 

Botpress uses LLMs to identify user intent. Why? It’s magnitudes better for both the builders and users of an AI agent. 

We feel strongly about this stance, so I’d like to take a few minutes to explain our lack of intent classifiers. 

TLDR; It’s easier to build, it’s more accurate, and it’s simpler to maintain. 

The olden (pre-LLM) days

(If you’re familiar with what intent classifiers are and what they do, feel free to skip this section.) 

An intent classifier is a tool that categorizes user inputs into predefined intents based on training data.

Developers have to curate and label countless examples for each possible intent, hoping the system can match user inputs to these examples. 

For example, with an e-commerce chatbot, developers might define an intent like "TrackOrder". Their example utterances might include: "Where is my package?" "Track my order," and "Can you check the delivery status for me?"

They’re essentially training the AI agent to recognize the intent of the user by giving examples. And yes, they have to input all of those by hand.

Luckily, the need to do this manual mapping of possible utterances to an intent all but vanished as LLMs became more advanced.

But plenty of conversational AI platforms still use them. Why? We’ll get to that. 

4 drawbacks of intent classifiers

It’s not just that it’s a longer process – intent classifiers suck for a lot of reasons. Here are a few:

1. Data dependancy

Intent classifiers are data-hungry. They need a huge, representative dataset of user examples for each intent to work accurately. Without it, they struggle to classify inputs correctly.

And building these datasets is a slog. Developers spend endless hours gathering and labeling examples, which is – without a doubt – not a good use of their time.

2. Limited scalability 

Intent classifiers also aren’t built to scale. Adding new intents means collecting more data and retraining the model, which quickly becomes a bottleneck for development. Plus, they can be a maintenance headache – because as language use evolves, so do the utterances.

3. Poor language understanding 

Intent classifiers lack true language understanding. They struggle with variations in language, like:

• Synonyms
• Paraphrases
• Ambiguous phrasing
• Typos 
• Unfamiliar colloquial expressions
• Fragmented inputs

They also typically process each utterance in isolation, which means they lack the ability to maintain context throughout a conversation. 

4. Overfitting

Intent classifiers are prone to overfitting, where they memorize training examples instead of learning general patterns. 

That means that they perform well on exact phrases they’ve seen but struggle with new or varied inputs. This makes them way more brittle than appropriate for a professional use case.

6 reasons LLMs are better 

LLMs all but solved these issues. They understand context and nuance, and devs don’t need to fill them with training data to get them started. An LLM-based agent can start conversing from the moment it’s created. 

1. Zero-short learning capabilities 

LLMs don’t need examples to learn. Their extensive pretraining means they already understand context, nuance, and intent without needing developers to feed them specific examples.

2. A little thing called nuance

LLMs excel where intent classifiers fall short. They can interpret idioms, sarcasm, and ambiguous language with ease. 

Their extensive training on diverse datasets gives them the ability to grasp the subtle nuances of human communication that intent classifiers often miss.

3. Better context

LLMs don’t lose track of the conversation. They remember what was said earlier, which makes interactions flow naturally and feel more coherent.

This context also helps them clear up ambiguities. Even when the input is vague or complex, they can piece it together by looking at the broader conversation.

4. Scalability

LLMs are 100% better at scaling. They don’t need retraining to take on new topics, thanks to their broad understanding of language. 

That makes them ready to handle just about any use case right out of the box. For multi-agent systems, it’s a no-brainer to use an LLM instead of an intent classifier.

5. Flexibility

LLMs don’t rely on rigid templates. Their flexibility means responses feel natural, varied, and perfectly tailored to the conversation. They’re a much better experience for users than brittle intent classifiers.

6. Less training data

LLMs don’t need task-specific labeled data to get the job done. Their power comes from massive pretraining on diverse text, so they’re not reliant on painstakingly annotated datasets.

If needed, devs can always customize an LLM for their project. For example, LLMs can be fine-tuned with minimal data, so they can quickly adapt to specialized use cases or industries.

Why do other companies use intent classifiers?

Good question. If LLMs are so much better at classifying intents, then why do so many companies still use intent classifiers? 

The answer isn’t a pretty one, and it’s not the most diplomatic to say: it’s a problem of legacy tech. 

Most companies have a vested interest in using intent classifiers. They’ve built huge install bases that run on it. They have no reason to dissuade their users away from the system they’ve built.

But Botpress is LLM-first

LLMs are much better at identifying intents than old-fashioned intent classifiers. That’s why we rewrote from scratch to be LLM-first in 2020. 

We knew better tech had arrived, and instead of clinging to legacy tech, we invested in taking the leap.

Will we add intent classifiers?

No. We care too much about our builder experience and their users’ experience.

The future of intent classification

Intent classifiers are a tool of the past. That’s why we went all in on LLMs. 

As LLMs keep getting better, so will the AI agents built on Botpress. We’re ready and excited to keep raising the bar for what conversational AI can do.

If you’re looking to build flexible AI agents powered by LLMs, feel free to start building on Botpress. It’s free. 

Build AI Agents

Build custom autonomous agents

Start now

No credit card required