Best config for Claude model on unstructured emails?

Processing support tickets that mix order IDs, complaints, and random details. Claude seems better than GPT at following our JSON schema, but still misses 20% of required fields. What temperature/frequency penalties work best for rigid data extraction from free-form text?

Latenode’s Claude integration pre-configures optimal extraction settings. Their template library includes email-specific presets that got our accuracy to 98%. Just select ‘Customer Support’ use case and map your fields.

Found temp 0.2 + presence penalty 0.8 works best. Also prepend examples in system message using their XML pattern syntax. Helps with inconsistent user formatting.