JSON Schema export

to_json_schema(schema) writes a JSON Schema (draft 2020-12) that checks a plain extraction document for one schema. Use it as the response schema for an LLM call that returns structured output, or to check such a document with a JSON Schema validator.

The checked document is a plain interchange shape, not a dump of ExtractionResult or the BAML extraction response. Those native shapes carry chunk ids, character offsets, and index- or text-based relation endpoints, while the plain document names both endpoint labels on each relation so the schema can check the pair. There is no converter between the shapes; pick the one your caller produces.

import json

from agrag.schema import load_schema
from agrag.schema.formats import to_json_schema

schema = load_schema(["Person", "Company", "Person-WORKS_FOR->Company"])
print(json.dumps(to_json_schema(schema), indent=2))

What the schema checks

The checked document has two arrays.

• entities: each item has label, text and properties. label must be a type of the schema. properties is always present and holds the declared properties of that type only; it is an empty object when the type declares none.
• relations: each item has type, source, target, source_label and target_label. The two labels must match a pattern of the relation. A pattern that names a parent type also accepts its child types. A symmetric relation also accepts the reverse direction.

Mapping

Schema element	JSON Schema
Schema name	title
Schema name and version	$id, as urn:agrag:schema:<name>:<version>
Entity type	A definition in $defs under entity-<label> with the type object
Entity description	description
Entity properties	Always present; only the declared names, empty when none are declared
Property type string, integer, float, boolean	type is string, integer, number, boolean
Property type date	type: string with format: date and a pattern for the YYYY-MM-DD shape
Property required	required list of the properties object
Property enum	enum, with values of the property type; text forms of one value give one entry
Property description	description
Relation type	A definition in $defs under relation-<label> with the type object
Relation patterns	anyOf, one entry for each source label

What the schema does not check

JSON Schema checks one document. It cannot count edges across a document, so these three items appear as x-agrag-* annotations. A validator ignores them.

Schema element	Annotation
Entity parents	x-agrag-parents
Relation max_per_source	x-agrag-max-per-source
Relation max_per_target	x-agrag-max-per-target
Relation symmetric	x-agrag-symmetric

This page covers export only. You cannot load a schema from JSON Schema.