What is Few-Shot Learning?
Few-shot learning (in-context learning) is providing 2-5 examples of desired input-output pairs before your actual query. The LLM learns the pattern from examples without any training—dramatically improving consistency and accuracy for specialized tasks.
Why It Matters
Describing a complex output format in words is error-prone. Showing 3 examples of exactly what you want is clearer for both humans and LLMs. Few-shot prompts consistently outperform zero-shot for classification, extraction, and formatting tasks.
Shot Terminology
Zero-Shot
No examples. Just instructions. Relies on model's pretraining.
One-Shot
Single example. Minimal context, better than zero.
Few-Shot (2-5)
Multiple examples. Optimal for most tasks.
Example: Risk Classification
# Few-shot prompt for contract risk classification
Example 1:
Input: "Party shall indemnify against all claims"
Output: {"risk": "HIGH", "type": "unlimited_liability"}
Example 2:
Input: "30-day written notice required"
Output: {"risk": "LOW", "type": "standard_terms"}
Example 3:
Input: "Liability capped at fees paid"
Output: {"risk": "MEDIUM", "type": "capped_liability"}
Now classify: "Vendor assumes no liability for consequential damages"