@openparser/schema

Zod schemas and TypeScript types for openparser@1 document graphs and the generic raw OCR result envelope.

Use this package to validate parse results at runtime, type document graphs in TypeScript, and share one canonical graph definition across providers and runtimes.

Full reference · OpenParser

Install

npm install @openparser/schema

Import

Import from the package root (@openparser/schema):

import {
  ParsedDocumentSchema,
  RawParseResultSchema,
  OcrOutputFormatSchema,
  type ParsedDocument,
  type RawParseResult,
  type ParsedDocumentWithElementKinds,
} from '@openparser/schema';

Validate at runtime with ParsedDocumentSchema.parse(json). Narrow adapter output with ParsedDocumentWithElementKinds<'text' | 'table'> when you know which element kinds a converter may emit.

openparser@1

The canonical result is a document graph:

• pages define explicit coordinate spaces and reading order.
• elements preserve text down to words/symbols plus tables, figures, formulas, key-values, query answers, selections, signatures, barcodes, links, sections, and provider-defined fallback elements.
• relations retain hierarchy and semantic links without duplicating trees.
• text_annotations preserve range-based languages and styles.
• assets retain page and figure images by URI or base64 payload.
• normalized confidence records retain their original value and scale and are explicitly uncalibrated across providers. Page-level OCR aggregates use pages[].confidence; image-quality scores/defects use pages[].quality.

text is plain reading-order text. Every span is a half-open UTF-16 code-unit range into that string, so JavaScript consumers can use text.slice(start, end) without provider-specific indexing logic. markdown is the canonical best-effort rendering of the graph; the untouched provider response remains available through the separate raw output format.

pages[].reading_order is an ordered subset of that page's element_ids and contains top-level content, not both a parent and all of its descendants. Word- and symbol-level elements are present only when the provider and requested options return them. Relation directions are explicit: contains points from parent to child, while caption_of and footnote_of point from the annotation to its target.

Geometry always includes a bounding box and may additionally retain the native polygon.

All published object schemas are strict. Adding a field or otherwise changing a shape in a way that an existing schema rejects requires a new output_format revision; openparser@1 will not silently grow incompatible fields.

Table validation rejects overlapping cells and enforces a structural row-coverage limit during overlap checks.

License

Apache-2.0