Few-Shot Prompting

What Is Few-Shot Prompting?

Few-shot prompting is a technique where you include a small number of complete input/output examples directly in the prompt to demonstrate to the LLM exactly what kind of response you expect. The model performs in-context learning — inferring the task pattern from the examples without any weight updates — and applies that pattern to the new input.

The term comes from the machine learning concept of "few-shot learning," but in the LLM context it refers entirely to prompt design rather than model training.

Prompting Variants by Example Count

A Practical Few-Shot Example

Zero-shot (may produce inconsistent output):

Extract the company name and founding year from this text:
"Apple was founded by Steve Jobs in 1976."

Few-shot (consistent, structured output):

Extract company name and founding year as JSON.

Text: "Microsoft was founded by Bill Gates in 1975."
Output: {"company": "Microsoft", "founded": 1975}

Text: "Stripe was founded by Patrick Collison in 2010."
Output: {"company": "Stripe", "founded": 2010}

Text: "Apple was founded by Steve Jobs in 1976."
Output:

The model now reliably produces {"company": "Apple", "founded": 1976}.

Why Few-Shot Prompting Works

LLMs have seen enormous amounts of pattern-matching examples during training. When you provide a few examples, you are activating the model's learned ability to continue a pattern rather than relying on its ability to interpret abstract instructions. This is especially effective for:

• Unusual output formats not covered well by zero-shot instructions alone.
• Edge case handling — showing how to treat missing values, ambiguous inputs, etc.
• Tone calibration — demonstrating the exact level of formality or conciseness you want.
• Classification tasks — showing label boundaries through examples rather than descriptions.

Few-Shot Prompting for Structured Extraction

Few-shot prompting is particularly powerful for web content extraction tasks. Rather than writing elaborate instructions, you show the model what "good extraction" looks like:

const systemPrompt = `Extract structured product data from web page content.

Example input:
"The ProWidget X1 retails for $299. Available in blue and red. Ships within 3 days."

Example output:
{"name": "ProWidget X1", "price": 299, "colors": ["blue", "red"], "shipping_days": 3}

Now extract from the following page content:`;

const { content } = await sdk.scrape("https://example.com/product");
const userMessage = content;

KnowledgeSDK's /v1/extract endpoint uses this approach internally — providing the model with structured examples of well-extracted knowledge to ensure consistent, clean output across a wide variety of web page formats.

Tips for Effective Few-Shot Examples

• Use real, representative examples — synthetic edge cases may teach the model the wrong distribution.
• Cover your edge cases — include an example with a missing field, an ambiguous value, or an unusual format.
• Keep examples concise — long examples inflate token count; trim to the essential signal.
• Order matters — the last example before the real input has disproportionate influence; make it your cleanest one.
• Balance your examples — for classification tasks, include roughly equal examples per class.

Few-Shot vs. Fine-tuning

Few-shot prompting is dramatically easier to iterate on than fine-tuning: change an example, test immediately, no training run required. Start with few-shot prompting and only graduate to fine-tuning when you have 100+ examples and need consistent performance at high request volumes.