Example selection, format consistency, balancing example quality vs context cost.

Few-Shot Prompting — Teaching by Example

Few-shot prompting provides 3-10 input-output examples in the prompt to demonstrate the desired task, format, and tone. The model infers the pattern from examples rather than explicit instructions. Key factors: example quality (gold standard examples beat average ones), example diversity (cover different edge cases), example order (recent examples are weighted more), and format consistency (violating the format in one example confuses the model).

Interactive Pipeline Flow

UnderstandImplementOptimise

📋 Task — define what to show

Define the task clearly through examples: sentiment analysis, data extraction, code review, classification. Examples communicate format, depth, and style without explicit instruction.

Architecture Diagrams

Retrieving relevant examples at query time.

Loading diagram...

Key Design Decisions

⭐Example quality matters more than quantity

3 carefully curated, human-verified examples outperform 20 auto-generated ones. Each example is a data point teaching the model. One wrong example can skew the entire task.

🔄Dynamic few-shot improves accuracy on diverse inputs

Fixed few-shot examples are static templates. Dynamic examples retrieved from a vector DB pick the most relevant ones for each query. Especially valuable for classification tasks with many classes.

📐Format violations in examples are silent bugs

If 4 examples output JSON and 1 outputs plain text, the model randomly picks one. Audit your example bank regularly for format consistency. A linter for your example bank prevents this.

Discussion