Few-Shot Prompting

Few-shot prompting is a technique where you include a small number of input-output examples in the prompt to guide the model's response, without updating model weights. The model uses these examples to infer the desired pattern or format. Operators encounter this when they want consistent output structure (e.g., JSON, bullet lists) or task-specific behavior without fine-tuning. The number of examples is typically 2-5; more than that may exceed context length or degrade performance.

Few-shot prompting leverages in-context learning, where the model generalizes from examples provided in the prompt. It sits between zero-shot (no examples) and many-shot (many examples, sometimes hundreds). The examples act as a task specification: the model sees a pattern and continues it. For operators, few-shot is a practical way to steer outputs without retraining. However, it consumes context window space—each example adds tokens, reducing room for other instructions or user input. The quality of examples matters: they should be representative and diverse. If the model misinterprets the pattern, adjusting examples or adding explicit instructions can help. Some runtimes (e.g., vLLM, llama.cpp) support system prompts that can include few-shot examples, but the same token budget applies.

An operator wants the model to extract names and dates from text into JSON. A few-shot prompt might include two examples: "Input: 'John was born on 1990-05-12.' Output: {"name": "John", "date": "1990-05-12"}" followed by the actual input. This costs about 30-40 tokens per example. On a 4K context window, using 5 examples leaves ~3.5K tokens for the actual task—enough for a few paragraphs.

In LM Studio, you can prepend few-shot examples in the user message or system prompt. For Ollama, you'd include them in the prompt string: ollama run llama3.2 'Translate English to French. Example: hello -> bonjour. Example: cat -> chat. Now translate: dog'. The model sees the pattern and outputs 'chien'. If the output is wrong, operators adjust examples or add formatting instructions.