[WSJF-5] One-feature-at-a-time: split AI extraction from parser generation #9

New issue

Open

opened 2026-06-27 10:45:28 +02:00 by bullitt · 0 comments

bullitt commented

2026-06-27 10:45:28 +02:00

Owner

Context

Article: 'One-feature-at-a-time reduces cognitive load on model.'

Problem

Our AI prompt asks the model to simultaneously: extract fields AND generate a reusable field_map. These are different tasks with different success criteria.

Proposal

Split into 2 sequential calls:
1. Extract fields from this email (cheap, focused)
2. Given this email and the extracted fields, generate a field_map (can reference extraction as ground truth)
Benefit: extraction quality improves (focused task), field_map can be validated against known-good extraction

WSJF: 5 (Medium value, Medium risk reduction, but larger size)

## Context Article: 'One-feature-at-a-time reduces cognitive load on model.' ## Problem Our AI prompt asks the model to simultaneously: extract fields AND generate a reusable field_map. These are different tasks with different success criteria. ## Proposal - Split into 2 sequential calls: 1. Extract fields from this email (cheap, focused) 2. Given this email and the extracted fields, generate a field_map (can reference extraction as ground truth) - Benefit: extraction quality improves (focused task), field_map can be validated against known-good extraction ## WSJF: 5 (Medium value, Medium risk reduction, but larger size)