Knowledge & behaviour

Training & Example Answers (Few-Shot)

In this area, you teach your agent how it should answer using concrete examples. You store pairs consisting of a typical user question and the ideal answer ("few-shot" examples). From these, the model learns your tone of voice, your phrasing and the right way to handle tricky questions — without you having to write a single line of instruction text. Use this area whenever the agent is correct in substance but off in style, or when a particular question keeps being answered badly.

You will find this area in the dashboard under Agents → (your agent) → Training. The card is headed "Training examples".

How an example works

Each example consists of exactly two fields:

"User says" — an example question or message, just as a real customer would write it.
"Agent should answer" — the ideal answer the agent should give in exactly this case.

The agent is given these pairs as learned models in every conversation (internally, they are placed at the start of the history as alternating user/assistant messages). It does not copy the answer word for word, but adopts the style, length and logic and applies them to new, similar questions.

Step by step

Step 1: Create your first example. If nothing has been stored yet, you will see the note "No examples yet". Click "Add example". A new, empty block appears, labelled "Example 1".

Step 2: Fill in the "User says" field. Enter a realistic customer question, e.g. "Hi, do you deliver to Switzerland?". Write the way customers actually do (including colloquial language — typos are fine) — the closer to the real wording, the better the example works.

Step 3: Fill in the "Agent should answer" field. Here, formulate the perfect answer, e.g. "Yes! We deliver to Switzerland — delivery takes 3–5 working days…". This field is deliberately a bit taller (3 lines), because the model answer is usually longer than the question. Pay attention to the tone, the level of politeness and the depth of detail you generally want — the agent imitates exactly that.

Step 4: Add more examples. Using the "Add example" button (bottom left) you can add as many further pairs as you like. They are numbered automatically ("Example 2", "Example 3", …).

Step 5: Remove individual examples. Each block has a bin icon in the top right ("Remove example"). One click deletes exactly this pair. There is no additional confirmation prompt — but the deletion only becomes final once you save (Step 6).

Step 6: Save. As soon as you change something, the safety bar appears at the bottom with the note about unsaved changes and the Save button. Only a click on it transfers your examples to the agent. The Save button stays greyed out as long as there is nothing new to save, and is acknowledged with a brief confirmation after a successful save. If you switch to another sidebar menu item before saving, your entries in this area are retained — but don't lose sight of them.

Fields, defaults and limits at a glance

"User says" (technically: user) — a mandatory part of an example. Default: empty. Recommendation: a short, typical customer question.
"Agent should answer" (technically: assistant) — a mandatory part of an example. Default: empty. Recommendation: the ideal, fully formulated answer.
Number of examples — Default: none (empty list). Recommendation: 3–8 well-chosen, varied examples. Upper limit: 50 pairs per agent. This limit applies above all to automatic promotion from the Decision-Flow (see below); once it is reached, no further pairs can be added automatically.

Promoting examples automatically from the Decision-Flow

You don't have to write all examples by hand. In the AI diagnostics / AI-Debug area you will find a Decision-Flow for every answer that was sent (the step-by-step decision view in n8n style) with thumbs-up/-down. Rate an answer with thumbs down and store a corrected version, and you can then "promote to a few-shot example" with one click. This automatically creates a new pair in exactly this Training area:

The original customer question lands in the "User says" field, your corrected answer in the "Agent should answer" field.
Identical pairs are not saved twice (automatic duplicate check) — so an accidental double-click does not create a second, identical example.
Likewise, from the thumbs-down feedback a correction can alternatively be promoted as a routing example (i.e. "this kind of question belongs to topic X" — e.g. knowledge, appointment booking, order, lead) instead of as an answer example. Few-shot trains the style/content of the answer, the routing example trains the classification of the question — choose the appropriate one depending on the problem.

This way your training base grows during live operation: spot a bad answer, correct it, lock it in with one click — and at the next similar conversation the agent answers better.

Tips & pitfalls

Empty fields are ignored. If either the question or the answer field of an example is left empty, the entire pair is silently skipped during training. A half-finished, half-empty example therefore achieves nothing — always fill in both fields or delete the block.
Only saving makes it take effect. Adding, editing and deleting only take effect after a click on Save. As long as the note about unsaved changes is visible, nothing has reached the agent.
Few-shot is not a knowledge store. Examples teach the agent style and approach, not current facts. Opening hours, prices or delivery countries belong in the knowledge base — otherwise the details in the example answers go out of date and the agent passes on outdated information.
Quality beats quantity. A few, clearly distinct examples have more effect than many almost identical ones. Cover different situations (greeting, difficult question, polite refusal) instead of storing the same question in ten variants.
Avoid contradictions. If you give contradictory answers in two examples for similar questions, this confuses the agent. Keep your model answers consistent with one another — and in line with the system prompt/persona (tone of voice, level of politeness).
Capped at 50 pairs. More than 50 examples per agent is not possible. Anyone who promotes a lot via the Decision-Flow should tidy up occasionally and delete outdated pairs instead of blindly adding more.
Choose realistic phrasing. Write the customer question the way customers really type. An overly smooth, "perfect" example question works less often, because real enquiries rarely sound like that.