Before extraction
Define the entity, its stable identifier and the smallest useful schema before selecting a model or writing a prompt.
- Keep source URLs
- Record raw evidence
- Name required and optional fields
- Define acceptable unknowns
During normalization
Map vocabulary into controlled values without erasing meaningful differences or manufacturing precision.
- Preserve raw values
- Version mappings
- Attach confidence
- Separate inference from extraction
At the exception layer
Group failures by type, route ambiguous records for review and use recurring exceptions to improve the system.