The fastest way to improve your agent is to teach it from conversations that already happened. When a visitor asked something and got a wrong or incomplete answer, that's a direct signal of what your agent needs to learn. Vosaire lets you turn those moments into training data with one click.
How training works
When you mark a conversation response as a training example, you're adding a question-answer pair to your agent's knowledge. The next time a visitor asks something similar, your agent draws on that trained response instead of trying to figure it out from scratch from your knowledge sources.
Finding conversations to train from
Go to Conversations and look for:
- Conversations where the visitor followed up with "that's not what I meant" or "that didn't help"
- Conversations where the visitor asked a simple question but the agent gave a long, over-explained answer
- Questions your agent gets frequently that would benefit from a standardized, polished answer
- Any conversation where you, reading it back, know you'd answer it differently
Training on a response
- Open a conversation
- Find the specific message exchange you want to train on
- Click the Train button (the graduation cap icon) next to the agent's response
-
The training modal opens. You'll see:
- The visitor's question (what the agent needs to recognize)
- The agent's current answer (what it said)
- A text field where you write the ideal answer
-
Write a clear, complete answer — the way you'd want your support team to respond
-
Click Save training
Writing good training answers
The quality of your training answer directly affects how well your agent responds in the future. A few things that help:
Be specific. Vague answers ("it depends on your plan") aren't useful training data. Give the actual details.
Match your brand voice. Write the answer the way you want your company to sound — not too formal, not too casual.
Cover the full answer. If the question has multiple parts, answer all of them. Don't leave the visitor guessing.
Don't over-explain. Training answers that are too long train the agent to be verbose. Keep it concise.
Reviewing your training data
Go to your agent → Sources tab → find the "Training" source. This is where all your manually trained question-answer pairs live. You can edit or delete any of them from here.
If you notice the agent regressed on something you trained — read that training entry. Sometimes earlier training answers need to be updated when your product changes.
How often to train
There's no minimum — train whenever you notice something worth improving. Some teams do a weekly review of the past week's conversations and train on anything that stood out. That cadence tends to produce noticeable improvements within a few weeks.