Skip to main content

glossary terms

Few-Shot Prompting

Category
Prompt Engineering
Difficulty
Beginner

Definition

A prompt engineering technique where a large language model is provided with a small number of input-output examples within the prompt to guide its performance on a specific task.

How It Works and Context

Few-shot prompting leverages the in-context learning capabilities of large language models (LLMs). By including a handful of demonstrations—typically between two and five—within the prompt, users can significantly improve the model's ability to follow complex instructions or adhere to specific output formats. This technique is particularly effective for tasks where the desired output is highly structured or requires a specific stylistic nuance that a zero-shot prompt might fail to capture. Unlike fine-tuning, which involves updating the model's internal weights, few-shot prompting is a temporary, inference-time adjustment. While powerful, it is limited by the model's context window; providing too many examples can consume valuable tokens and potentially lead to performance degradation if the examples are irrelevant or poorly formatted.

Why It Matters

It is a critical tool for developers and users to steer model behavior without the high cost and complexity of retraining or fine-tuning. It allows for rapid iteration and customization of AI outputs, making it essential for building reliable applications that require consistent formatting, such as data extraction, classification, or specialized creative writing, while maintaining the flexibility to switch tasks instantly.

Real-world Example

A customer support manager wants an AI to classify incoming emails by sentiment. Instead of just asking the AI to 'classify this email,' they provide three examples: 'Email: My order arrived late. Sentiment: Negative', 'Email: I love the new features! Sentiment: Positive', and 'Email: How do I reset my password? Sentiment: Neutral'. When the AI receives a new email, it uses these examples to accurately categorize the sentiment in the requested format.

Common Mistakes

  • Using too many examples, which can clutter the context window and confuse the model.
  • Providing inconsistent or contradictory examples that lead to unpredictable model behavior.
  • Failing to use a clear delimiter (like '###' or '---') between the examples and the actual task, causing the model to blend them together.
  • Assuming the model will learn complex reasoning from examples when it actually only mimics the provided pattern.

Frequently Asked Questions

How does few-shot prompting differ from zero-shot prompting?

Zero-shot prompting asks the model to perform a task without any prior examples, relying entirely on its pre-trained knowledge. Few-shot prompting provides specific demonstrations to guide the model, which usually results in higher accuracy and better adherence to formatting requirements.

Is few-shot prompting the same as fine-tuning?

No. Fine-tuning involves training the model on a large dataset to permanently alter its parameters. Few-shot prompting is an inference-time technique that provides examples within the prompt itself, requiring no changes to the model's underlying weights.

What is the ideal number of examples to include?

While there is no universal rule, providing between two and five high-quality, diverse examples is typically sufficient. Including more can sometimes lead to diminishing returns or exceed the model's context window limits.

Related Glossary Terms

Chain-of-Thought

Chain-of-Thought (CoT) is a prompting technique that encourages large language models to generate a series of intermediate reasoning steps before arriving at a final answer. By explicitly articulating the logical progression, the model can solve complex arithmetic, commonsense, and symbolic reasoning tasks more accurately.

Context Window

The context window is the maximum number of tokens—representing words or sub-word units—that a large language model can process and consider simultaneously during a single inference pass.

Large Language Model (LLM)

A type of artificial intelligence model trained on massive datasets using deep learning techniques, specifically the transformer architecture, to understand, summarize, and generate human-like text.

Prompt Engineering

The systematic process of designing, refining, and optimizing inputs to guide generative AI models toward producing specific, high-quality, and relevant outputs.

Zero-Shot Prompting

A prompt engineering technique where a large language model is asked to perform a task without being provided with any prior examples or demonstrations of the desired output.