Few-shot learning lets AI models learn a new task from just a few labelled examples, usually by showing them in the prompt. Learn how it works and when to use it in chatbots.
More about Few-Shot Learning
Few-shot learning is the ability of an AI model to perform a new task after seeing only a small number of examples, usually somewhere between two and a handful. Rather than training or fine-tuning the model on thousands of labelled cases, you show it a few demonstrations and let it generalise from there.
In modern practice with a large language model, this almost always happens through the prompt. You place the examples directly in the input, followed by the actual query, and the model picks up the pattern on the fly. That specific style is also called few-shot prompting or in-context learning.
How Few-Shot Learning Works
A few-shot prompt has a simple structure:
- A short instruction describing the task.
- Two to five example input-output pairs.
- The real input the model should respond to.
For example, to classify customer support messages by intent, you might include three examples showing the format "message -> intent" before handing the model a new message. The model learns the task from the pattern of the examples, not from retraining.
This works because modern LLMs are pretrained language models that have already absorbed a huge amount of linguistic and task knowledge. The examples do not teach the model anything fundamentally new; they cue the model toward the right behaviour.
Few-Shot vs. Zero-Shot vs. Fine-Tuning
The three are points on a spectrum of how much guidance the model gets:
- Zero-shot learning: ask the model to perform a task with no examples.
- Few-shot learning: include a small number of examples in the prompt.
- Fine-tuning: update the model weights on thousands of examples.
Few-shot often closes most of the gap to a fine-tuned model at a fraction of the cost. When the examples are chosen well, the output quality is close to task-specific training. Where fine-tuning wins is on latency, token cost over time, and very narrow domains with consistent style requirements.
Why Few-Shot Learning Matters for Chatbots
Teams building AI chatbots reach for few-shot learning all the time, often without calling it that:
- Intent classification: show the model three examples of each intent category before asking it to classify a new message.
- Response formatting: include two examples of the tone, length, and structure you want, and the model mostly follows them.
- Data extraction: pin down a strict JSON schema by demonstrating two or three complete examples.
- Tool selection: show the model when to call a specific API versus when to answer directly.
SiteSpeak uses few-shot patterns internally to keep answer style consistent across customers. A handful of well-crafted examples in the system prompt can shift a chatbot's tone from generic to on-brand without any model training.
Choosing Good Few-Shot Examples
The quality of the examples matters more than the quantity. Useful heuristics:
- Cover the edge cases: include at least one tricky example, not just easy wins.
- Show the format you want: the model will mimic structure aggressively.
- Balance classes: if you are doing classification, do not stack the examples toward one label.
- Keep them short: every example eats context window and raises cost.
- Test and rotate: swap examples in and out and measure accuracy on a held-out set.
Limitations
Few-shot learning has real limits:
- Brittle to wording: small phrasing changes in the instruction or examples can shift outputs dramatically.
- Context window pressure: adding examples uses tokens. On long conversations you may run out of space.
- Not a silver bullet: on very specialised or low-resource tasks, the model may just fail regardless of examples.
- Inconsistent across turns: long multi-turn conversations can drift away from the demonstrated pattern.
For high-volume, high-precision tasks, fine-tuning or a more elaborate retrieval augmented generation setup usually outperforms pure few-shot prompting.
