Schema extract

Free text in, a typed object out — and an honest signal about what wasn't there.

Inputs

• schema — a JSON Schema describing the target object (fields, types, required).
• text — the source prose.

Method

1. Field-by-field, not free-form: walk the schema and locate evidence for each field in the text. One field's absence never blocks the others.
2. Never fabricate. If the text doesn't support a field, emit null and record a reason — a guessed value is worse than a missing one for a downstream agent.
3. Normalize types: dates → ISO-8601, numbers → numeric (strip currency/units into their own fields), enums → the closest schema-allowed value or null.
4. Score confidence per field in [0,1]: explicit verbatim match is high; inferred or ambiguous is low. Surface low-confidence fields so the caller can re-prompt.
5. Validate the assembled object against the schema before returning. On a type or required violation, report it — don't coerce silently.

Output

{ data: <object matching schema>, confidence: { <field>: number }, missing: [{ field, reason }], valid: boolean }

Determinism over creativity: same text and schema should yield the same object.