Omnist

Omnist ("omni-structure") is one canonical data model for JSON, YAML, TOML, XML, and its own native OML (Omnist Markup Language) — read any of them into a single tree, validate it against a schema, compare schema versions, and write it back out to any of the others.

from omnist import parse_schema, doc

s = parse_schema('''
    record Member { "name": string, "role": string }
    record Team   { "name": string, "members" [1,]: Member }
    root Team
''')

s.validate(doc({"name": "Platform",
                "members": [{"name": "Ann", "role": "dev"}]})).ok    # True

Why Omnist

Three things you can do with Omnist that are genuinely hard anywhere else:

• Gate schema changes in CI. v1.compatible_with(v2) answers "does every document the old schema accepts still validate under the new one?" with a decidable yes or no — an algorithm with a proof behind it, not a heuristic or a sampled guess. Put it in a pipeline and schema evolution stops being a code-review judgment call.
• One model across five formats. Read JSON, YAML, TOML, XML, or OML into the same tree, validate it against the same schema, write it back out to any of the others — one library, one mental model, and a written report of anything a lossy format had to adjust.
• Cut a big schema down to what one service needs. schema.extract(...) computes the minimal subschema recognizing only the labels you keep — the operation that trimmed industrial XML schemas to a fraction of their size in the research this library grew from.

What makes those possible is a deliberately small formal model (inspired by Lee & Cheung, CIKM 2010):

• A Document is a tree — an ordered list of labeled edges. Arrays are just repeated labels, so the same Document represents JSON, YAML, TOML, XML (including its interleaved repeated elements), and OML — Omnist's own format, the only one with zero loss in either direction.
• A Schema is named record definitions (closed named fields, each with a cardinality), where every field's type is always exactly one fixed scalar (optionally nullable), one Ref to a named record — referenced by name for reuse and recursion — or any, an explicitly-marked unchecked subtree. Written as OSD (Omnist Schema Definition).
• Closed by default; open only where explicitly marked — records are closed, and scalar types are never composed into enums or unions. The one deliberate opening is the any type (v0.5.0), and it carries its cost openly: compatibility checking ends exactly where any begins. That discipline is not a constraint bolted on top; it is what makes compatible_with, equivalent, normalize, extract, and infer well-defined, decidable operations instead of best-effort heuristics. See why Omnist for the case in full — including a verified capability matrix and the honest non-goals.

Boringly correct

The claims above are checked, not asserted: 100% line coverage, a mypy --strict CI gate, property-based fuzzing, formal ABNF grammars verified against the parsers, every documentation example executed in CI, and the schema algebra verified three independent ways — bidirectional inclusion, the paper's minimize-and-isomorphism theorem, and brute-force enumeration against set-theoretic ground truth. Performance is measured and published, not implied: a 100k-edge document validates in about a third of a second (the numbers). See testing for how the suite is built.

The model is defined formally in spec.omnist.dev; see the quickstart for the shortest possible example, or the user guide for the practical tour.

A 60-second tour

from omnist import Doc, parse_schema, infer, doc, read_json

# OML is omnist's own format -- see docs/formats/oml.md
Doc.from_oml('id: 1\ntags: "a"\ntags: "b"\n').to_oml()

# converting from other formats is just read one, write another
Doc.from_json('{"id": 1, "tags": ["a", "b"]}').to_yaml()

# describe a shape and check data against it; errors carry exact paths
s = parse_schema('record R { "id": integer, "tags" [0,]: string }\nroot R')
print(s.validate(doc({"id": "x", "tags": ["a"]})))
#   invalid:
#     at $.id: expected integer, got string ('x')

# learn a schema from examples
print(infer([doc({"id": 1, "tags": ["a"]})]).to_osd())
#   record Root {
#       "id": integer,
#       "tags": string,
#   }
#   root Root

# is a schema change backward-compatible? (operations are Schema methods)
v1 = parse_schema('record R { "host": string }\nroot R')
v2 = parse_schema('record R { "host": string, "port" [0,1]: integer }\nroot R')
v1.compatible_with(v2)        # True -- adding an optional field is safe

# schema-directed deserialization: upgrade leaves to match the schema
s2 = parse_schema('record R { "d": date }\nroot R')
read_json('{"d": "2024-01-01"}', schema=s2)   # [('d', datetime.date(2024, 1, 1))]

Run the full demo: python3 examples/canonical_model.py.

Installation

Requires Python 3.11+ (it uses the standard-library tomllib). The core and JSON support have no dependencies.

pip install omnist                      # core + JSON
pip install omnist[all]                 # + pyyaml, tomli_w, defusedxml

Installing also provides an omnist CLI command; see the CLI docs.

Or from a checkout:

git clone https://github.com/omnist-dev/omnist.git
cd omnist
python3 -m venv .venv && source .venv/bin/activate
pip install .                    # core + JSON
pip install pyyaml tomli_w defusedxml   # YAML / writing TOML / hardened XML

Documentation

Full index: docs/, also browsable as a site at py.omnist.dev.

• Quickstart — the shortest possible example: one OML snippet, one schema, validate(), infer().
• Why Omnist — the differentiation case: a falsifiable thesis, a verified capability matrix against JSON/YAML/TOML/XML, a worked compatible_with comparison against jsonschema, and the honest non-goals (including XML attribute/namespace dropping).
• User guide — the practical tour: documents, OML (the native format), OSD, the Python builder, validation, operations, other codecs, inference.
• OML — Omnist's own format, designed alongside the model so every Document round-trips with zero adjustments, and how it maps onto the Python Document.
• The Schema model & OSD — Omnist's other central feature: record definitions, cardinality, the Python builder, and the comparison/inference operations.
• API reference — every public name, with signatures.
• CLI — the omnist command-line tool.
• Schema-directed deserialization — what changes (and what doesn't) about a Document's Python types when a schema is, vs. isn't, passed to a reader.

Status

Omnist is beta. The surfaces you build on — the Document model, the OSD and OML grammars, the public API importable from omnist, and the error types — change only through a deprecation cycle; see the stability policy for exactly what is covered and what may still change. Beta is not 1.0: 1.0 remains gated on the any-type openness decision, spelled out in the same policy. See the PyPI badge above (or CHANGELOG.md) for the current version.

Feedback and bug reports welcome: https://github.com/omnist-dev/omnist/issues. See SECURITY.md for the trust model if you parse untrusted input.

License

Apache-2.0 — see LICENSE and NOTICE.

Background

The model is inspired by Lee & Cheung, "XML Schema Computations: Schema Compatibility Testing and Subschema Extraction" (CIKM 2010), simplified for the JSON family of formats. You don't need the paper to use Omnist — spec.omnist.dev is self-contained.