User Story

As Maya, in order to improve extraction quality on complex government forms by teaching the model what good output looks like, I want an extraction variant that uses curated few-shot examples without requiring fine-tuning.

Preconditions

• #58 (Story 10 variant picker) merged to main

Acceptance Criteria

• New variant extraction/few-shot-sonnet that prepends 2-3 canonical (PDF description → spec) exemplar pairs to the extraction prompt
• Exemplars committed under src/services/extraction/exemplars/ (or similar); each includes a short description, a compact spec, and rationale for inclusion
• Extraction tab in Settings → Variants lists the few-shot variant
• Evaluation run comparing sonnet baseline vs few-shot-sonnet on all three fixtures with the LLM-judge scorer
• Both scorers (deterministic + LLM-judge) reported
• New catalog page catalog/experiments/pdf-field-extraction/few-shot-sonnet.md with exemplar descriptions, approach, metrics, and findings on what the exemplars seemed to help with
• catalog/experiments/_roadmap.md updated with shipped status and one-line finding

Success Metrics

• Non-trivial delta (positive or negative) in at least one metric vs the Sonnet baseline — teaches us something either way
• Exemplars are documented well enough that another contributor could add more

Notes

• Class topic: prompt conditioning (Ch 8)
• Keep exemplar count small (2-3) to avoid blowing token budget
• Consider exemplars that deliberately demonstrate edge cases (nested groups, sensitivity labels, conditional fields)

Definition of Done

• Acceptance criteria met
• Tests pass
• Type checking passes
• CI pipeline green
• Deployed and demoable