Few-Shot Prompting

The fundamental mechanism by which a model temporarily adapts its behavior based on the input-output patterns provided in the prompt. Unlike fine-tuning, no gradient updates occur. The model's pre-trained weights remain frozen, and it uses the provided examples to infer a latent task description via its attention mechanism. This allows the model to recognize the desired format, tone, and reasoning structure for the specific legal task at hand.

The performance of few-shot prompting is highly sensitive to the choice of exemplars. Effective strategies include:

• Semantic similarity: Selecting examples whose embeddings are close to the target query
• Diverse sampling: Ensuring examples cover a range of edge cases and output structures
• Canonical examples: Using prototypical instances that clearly demonstrate the desired mapping Poor example selection can introduce spurious correlations or bias the model toward an incorrect pattern.

The model infers the output schema from the delimiter structure and labeling convention used in the examples. Consistent use of separators like ### Input: and ### Output: or XML-style tags creates a parseable template. In legal contexts, this consistency ensures the model generates outputs that match the expected structure—whether that's a contract clause classification, a case summary, or a structured JSON object for downstream processing.

Few-shot prompting works by providing the model with analogical reasoning anchors. Each example demonstrates a mapping from a specific input to a specific output. The model generalizes this mapping to the novel query at inference time. For complex legal tasks like statutory interpretation, examples can demonstrate how to decompose a statute into elements, apply a legal test, and format the conclusion—all without explicitly programming the reasoning steps.

Each few-shot example consumes tokens from the model's context window, creating a direct trade-off between the number of examples and the remaining capacity for the user query and generated output. For lengthy legal documents, practitioners must balance:

• The number of examples provided
• The length of each example
• The space required for the target document Efficient example compression and selection are critical for production legal AI systems.

The examples define the output label space for the model. In a legal classification task, the few-shot examples explicitly enumerate the valid categories—such as Force Majeure, Indemnification, or Limitation of Liability. The model is strongly conditioned to select from this demonstrated set rather than generating novel or hallucinated categories, significantly improving output reliability for downstream structured extraction pipelines.

The maximum token capacity a model can process in a single request. Few-shot examples consume precious context window space, creating a direct trade-off:

• More examples = better guidance but less room for the actual legal document
• Fewer examples = more document space but weaker behavioral control Effective few-shot design requires token budgeting to balance example count against the length of contracts or case law being analyzed.

The process of reducing token length while preserving semantic meaning. For few-shot legal prompts, compression techniques include:

• Truncating verbose example outputs to essential elements
• Removing redundant formatting from demonstrations
• Using abbreviated citation formats in examples Compression allows more examples to fit within the context window, improving model guidance without sacrificing space for the target document.

The capability to generate responses in machine-readable formats like JSON. Few-shot examples are the primary mechanism for teaching a model the exact output schema required:

• Each example demonstrates the desired field structure
• Consistent formatting across all shots reinforces the pattern
• Mismatched example outputs cause schema drift and parsing failures For legal pipelines, few-shot examples must mirror the production JSON contract precisely.

The practice of tracking changes to prompt templates over time. Few-shot examples are sensitive assets that require version control because:

• A single poorly chosen example can introduce systematic bias
• Model updates may change how examples are interpreted
• Legal domain shifts require example set refreshment Versioned few-shot sets enable auditable rollback and systematic A/B testing of example configurations.