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.