Use cases
Structured extraction
Documents in, JSON that matches your schema out.
Invoices, claims, contracts and forms all follow patterns. A small model trained on documents you have already processed learns where your fields live and returns them in your schema.
The problem
Long inputs, short outputs
Documents are long, so cost per document grows with every page you send to a hosted model.
Malformed output breaks pipelines
One missing key or wrong type stops the job downstream. Output must match the schema every time.
Documents are sensitive
Contracts and claims often can't leave your environment at all.
Where it applies
- Invoice and receipt fields
- Insurance claims
- Contract clauses and dates
- Forms and applications
- Emails to CRM records
How the switch is checked
Before you switch, the fine-tuned model and the LLM you use today answer the same held-out examples from your data.
- Field accuracy: the share of reference fields extracted correctly, with numbers and text normalised
- Valid JSON rate, and the share of outputs that match your JSON schema
- Accuracy per field, so you see which field needs more examples
What you need
Each example's document text, and the JSON object you expect from it as the answer.