Intent Classification
Work out what a user is actually asking for.
Intent Classification takes a natural-language query and matches it to one of the intents you've defined, so the workflow can route to the right action rather than guessing.
Reference layer pending. This page documents the skill's purpose and configuration as used in the builder. Its formal configuration schema, output schema, and error codes are not yet documented — see the Skill Library overview.
Worked example: routing inquiries in a financial chatbot
An agent powers a financial inquiry chatbot. A customer types "Can you please check my account balance?" or "I need to dispute a recent transaction." The workflow has to understand which, and route accordingly.
The problem. Customers phrase the same goal in different ways. A keyword search misses "How much money is in my account?" — no keyword matches, but the intent is obvious to a person.
The approach. Define a set of intents (CheckAccountBalance, DisputeTransaction) with examples of how each might be phrased. The skill uses an LLM to match the query to the best-fitting intent.
1. Choose the model. Select a model from the Model ID dropdown.
2. Define your intents. This is the part that matters. For each:
Name — a clear identifier, e.g.
CheckAccountBalance.Description — what the intent means: "Triggered when a user wants to know their current account balance."
Examples — several phrasings a user might actually use: What's my balance?, How much money is in my account?, Show me my current funds. These are what teach the model to recognize the intent.
Repeat for every intent the workflow handles — DisputeTransaction, ApplyForLoan, and so on.
3. Set the query. Point the Query field at the user's message from the previous step — $input.userQueryText.
The quality of the examples decides the quality of the classification. Write the phrasings your customers actually use — including the awkward ones — rather than the tidy version you'd write yourself.
Output
intentMatched—trueorfalse: whether a suitable intent was found.intents— the list of potential matches. The first item is the best match. Branch on this to route the workflow.query— the original query that was analyzed.
Handle intentMatched: false deliberately. A user whose intent you didn't anticipate is a real case, and falling through to nothing is a worse experience than saying you didn't understand.
To add this skill to an agent, see Adding a Skill to the Agent.
Last updated