infrastructure · parser

Unstructured in. Evidence out.

The ingestion engine of the Verifiable Knowledge Layer: documents, scans, images, audio and video become structured, machine-readable evidence — layout, meaning, exact location and provenance. Standalone API, available in pilot; the first layer, not the final destination.

PDFDOCXXLSXHTMLimagesaudio+ any source
POST /v1/parse
# parse a document, keep every span
POST /v1/parse
{ "document": "s3://acme/contract_2024.pdf",
  "outputs": ["canonical_text", "structure", "spans"] }

# 200 OK
{ "atoms": 1284, "tables": 17,
  "reading_order": "resolved",
  "source_spans": true,
  "certificate": "cert_9f2…" }
The problem

Parsing is where AI systems start lying.

Most parsers flatten documents into a string. Structure, geometry, tables and provenance are destroyed before any model sees the content — and no downstream step can recover what parsing lost.

flattened tableslost reading orderorphan footnotesno source spanssilent OCR errors
the defaultdocument"one long string"LLMconfident, wrong
with Parserdocumentartifacts + spansLLMgrounded, checkable
What it produces

Artifacts, not strings.

Canonical text

  • deterministic
  • stable coordinates
  • addressable

Same document, same text — always.

Structure atoms

  • tables
  • lists · headers
  • reading order
  • geometry

Every element typed and addressable.

Built for verification

  • certificate
  • source spans
  • coordinates
  • provenance

Certificates in pilot; public verification at launch.

How it works

Three moves. Nothing hidden.

PDF inParserartifactsAPI responsecertificate
01

Submit

Send a document or point the API at a source. Nothing to install, nothing to host.

02

Parse

The measured champion stack for your document family runs — versioned and swappable.

03

Verify

Receive artifacts plus a certificate. Every element traces to its exact position in the source.

Inside the engine

From page to artifact, stage by stage.

Each stage is run by a measured champion — benchmarked per document family, versioned, and swapped the day a better engine wins.

Layout detectionregions, columns, figures — geometry first
OCRchampion engine per document family
Reading orderthe human path through the page, resolved
Tables & structurecells, headers, spans — typed, not flattened
Canonical textdeterministic — same document, same bytes
Structure atomsevery element addressable by position
Certificateengine version + inputs + outputs + spans, bound

Where this engine sits in the platform → architecture

Show me

One artifact, exactly as the API returns it.

Left: the element on the page. Right: what you receive — typed, located, verifiable.

table · reading_order 47 bbox [143, 220, 502, 640] p. 12
structure atom · 47 of 1,284
# artifact 47 — table, page 12
{ "type": "table", "page": 12,
  "bbox": [143, 220, 502, 640],
  "reading_order": 47, "cells": 84,
  "canonical_text": "Q2 revenue by segment…",
  "source_span": "contract_2024.pdf#p12/143,220,502,640",
  "certificate": "cert_9f2…" }

# artifact 51 — chart, page 14: the engine abstains
{ "type": "chart", "state": "ABSTAINED",
  "reason": "chart-to-table below the evidence bar" }

Representative artifact — the exact response contract is published with the pilot. The abstained chart is not a bug: fabricating a table would be.

Benchmarks

Measured before claimed.

Public benchmarks

Pre-registered evaluations against named competitors — published with the pilot, every number with its n.

Internal evaluation

Running today on proprietary gold sets. A number appears below only after it survives its gate — until then, the honest state is shown.

DimensionUnitStatusNotes
Reading orderF1 vs goldmeasured · internalpublication after pre-registered eval
Tablesstructure F1measured · internalgold set expanding — n reported with every number
OCRCER / WERmeasured · internalchampions measured per document family
Formsfield F1publication pendinggold set in construction
Long documentsfidelity at 60+ pagespublication pendinglong-document bench in progress
Chartschart-to-tableUNKNOWN · fail-closedthe engine abstains instead of guessing

Charts are the honest row: today the engine abstains rather than fabricate a table. Continuous benchmark →

Pricing

Credits per unit of work.

1 pageunit
1 creditdraft ratio

Certificates included with every artifact. How pricing works →

A property, not a product

Every artifact is verifiable.

Certification runs through everything Orves produces. Anyone can verify a certificate — free, forever, no account.

✓ provenance✓ lineageversioningevidencegovernanceresolution statestrust →
API first

Integrate in minutes.

Everything is delivered through the API. Nothing to install, nothing to host.

REST APIPythonTypeScriptGoJavaOpenAPIExamplesPlayground soon
Questions

FAQ

Why not an open-source parser — Docling, Unstructured, LlamaParse?

Those are engines. Orves is the measured layer above engines: a permanent championship benchmarks candidates per document family, and the winner ships behind one stable API with determinism, abstention and certificates. Engines keep changing; your contract doesn't — and you never operate, version or re-benchmark them yourself.

How deterministic is parsing?

Same document, same engine version — same output, byte for byte. Determinism is a platform invariant, checked with round-trip tests. It is what makes evidence reproducible and certificates meaningful.

What happens when parsing is uncertain?

The element is returned with an abstained state and the reason. UNKNOWN is a decision, not an error — nothing is fabricated.

How are tables represented?

As structure atoms: typed elements with geometry, cells, headers and reading order, addressable by coordinates. Not markdown soup — a table you can query, cite and verify cell by cell.

How are certificates generated?

Every parse records engine version, inputs, outputs and source spans; the certificate binds them, so any artifact traces back to its exact position on the page. Certificates are in pilot; the public verification endpoint ships at launch.

How do you benchmark the engine?

Pre-registered evaluations on gold sets, every number published with its n, champions re-measured as engines evolve. A claim appears on this site only after it survives its gate — until then you see the honest state.

How is pricing calculated?

1 page = 1 credit — flat. No multipliers for OCR, tables or file type, and certificates are never billed.

Can I self-host?

No. Orves is delivered exclusively as an API. Determinism, benchmarks and certificates are properties of the operated platform — we only guarantee them where we run the engines. Nothing to install, nothing to maintain.

Which formats are supported?

PDF, DOCX, XLSX, PPTX, HTML, images, scans and audio — plus databases, APIs and event streams through connectors. If a family is not yet measured, the API says so instead of silently degrading.

Can I use the Parser alone?

Yes. Parser is a standalone API with its own benchmarks — built for teams that already run their own pipelines, RAG or ingestion. Add Verifiable Knowledge when you need knowledge you can prove.

Turn your first document into structured evidence.

15,000 free credits · no credit card · the API activates with the pilot.