Skip to content
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Fine-Tuning Intent Recognition using Generative AI

Make AI intents of the Answer node more accurate: write better meanings and descriptions, add context to the Intent recognition prompt and choose the model.

AI intents of the Answer node let a large language model decide which intent the customer's message belongs to. They need no training set, understand paraphrases and typos, and you can make them more accurate in two ways:

  1. Better intents – a clear Meaning and a helpful Description for each intent.

  2. Context in the Intent recognition prompt – what the model should know about the conversation and the customer.

How the recognition works

For every message, the node sends one request to the model with:

  • Behavior – the system message, by default: you are an assistant for intent recognition; classify the utterance and respond only with the intent name as JSON.

  • Text before list of intents, then the list of the node's AI intents, one line each: - NAME: Meaning [Description], then Text after list of intents, which asks for the answer as { 'recognized_intent': string } or NOT_FOUND.

  • User – what to classify, by default {current_utterance}, the customer's last message.

The model does not see the conversation history – only what is in these fields. Anything else it should know, you have to put there (see Add context).

Open the prompt with the gear icon next to the AI section title of the Answer node. It applies to all AI intents of the node.

1. Write better intents

  • Name – the value the model returns, for example INVOICE_NOT_RECEIVED. Use clear, unique names; the model reads them too.

  • Meaning – one sentence: what the customer wants. Customer did not receive the invoice for the last month.

  • Description – optional details that help the model decide. Keep it short: all descriptions of the node go into every request, so long descriptions add latency and cost.

What to put in the description:

Examples of what customers say

Customer asks about the weather forecast [Show me the weather in London; Will it rain in Seattle tomorrow; How hot will it be this weekend]

Keywords, synonyms and alias terms

Meaning: Complaint about a product. Description: claim, damage, broken, not working, won't turn on, defective

What the intent is not about

Draw the line between similar intents:

Customer asks whether a product is in stock. Not about product specifications or reviews.

Several facets of one intent

Customer is interested in discounts: promotional offers, seasonal sales or loyalty rewards.

Two intents that the model keeps mixing up usually have meanings that overlap. Rewrite the meanings so that each describes a different situation, or merge the intents and decide later with a Condition.

2. Add context to the Intent recognition prompt

Short answers like yes, the second one or 10 GB can be classified only with context. Add it to Behavior or User, and use variables – type { to insert one. Keep Text after list of intents as it is: the node reads the model's answer in exactly that JSON format.

The question the customer is answering

Store the last question in a variable (for example last_question) in the Message node that asks it, then:

A summary of the conversation
What you know about the customer

"I want 10 GB of mobile data" is an upgrade or a downgrade depending on the current tariff:

Date and time

You can also change User, for example to give the model the question and the answer together:

3. Choose the model and the timeout

In the same dialog:

  • Provider and Language Model – a smaller, faster model is usually enough for classification. Use a larger one only when the intents are subtle.

  • Reasoning › Effort (reasoning models) – keep it at the lowest option. Classification rarely needs more, and every step adds latency – on a call, the customer waits for it.

  • Temperature – GPT-5 and GPT-6 models use a fixed value; for other models keep it low for consistent results.

  • Timeout – how long the node waits for the model, 3 seconds by default. On a call, keep it short.

Last updated

Was this helpful?