---

Jev is a new kind of model from TypeSafe: it only returns structured output, it's blazing fast, and it's cheap. That's great, but it means Jev doesn't speak JSON Schema, and most LLM apps use JSON Schema for structured output.

This library sits in between. Give it your schema and your content, and you get back JSON that validates against the schema you started with.

JSON Schema  ──▶  Jev questions  ──▶  [ Jev ]  ──▶  answers  ──▶  JSON Schema output

Quick Start

pip install jev_jsonschema   # or: uv add jev_jsonschema

from jev_jsonschema import JevClient

schema = {
    "type": "object",
    "properties": {
        "sentiment": {"type": "string", "enum": ["positive", "neutral", "negative"]},
        "is_spam": {"type": "boolean", "description": "The message is spam."},
        "quality": {
            "type": "integer",
            "minimum": 1,
            "maximum": 5,
            "description": "Overall writing quality.",
        },
    },
}

with JevClient() as jev:  # reads TYPESAFE_API_KEY, or pass api_key="..."
    result = jev.evaluate(schema, state="Loved it. Shipped in a day.")

result.output
# {"sentiment": "positive", "is_spam": False, "quality": 5}

result.output validates against schema. Hand it to the same code that used to parse your model's JSON.

There's an AsyncJevClient with identical methods:

from jev_jsonschema import AsyncJevClient

async with AsyncJevClient() as jev:
    result = await jev.evaluate(schema, state="Loved it. Shipped in a day.")

Converting a schema is pure work, so if you're calling the same schema in a loop, convert once and reuse it:

question_set = jev.convert(schema)

for review in reviews:
    result = jev.ask(question_set, state=review)

What Maps to What

Jev has three question types. Here's the JSON Schema that reaches each one:

You get back the type in the first column. One question per schema property, in the order the schema declares them.

Jev needs to know what it's judging, so each question gets instructions from the property's description, falling back to its title, then to the property name. A real description is the biggest lever you have on answer quality: is_spam alone is a thin thing to ask about. Set instructions_fallback_to_key=False if you'd rather the library reject a boolean or number that has neither.

What You Get Back

Jev answers with distributions, not just values, and none of that is thrown away:

result.output
# {"sentiment": "positive", "is_spam": False, "quality": 5}

result.confidence
# {"sentiment": 0.97, "is_spam": None, "quality": 0.81}

result.probabilities
# {"sentiment": {"positive": 0.97, "neutral": 0.02, "negative": 0.01},
#  "is_spam": {"true": 0.03, "false": 0.97},
#  "quality": {"1": 0.0, "2": 0.0, "3": 0.02, "4": 0.1, "5": 0.88}}

result.usage
# SystemOneUsage(input_tokens=120, output_tokens=12)

result.response
# the raw SystemOneResponse, if you want it

probabilities is keyed by your schema's values, not Jev's internal labels, so a score of 1–5 reads as "1"–"5" and not "0"–"4". Noul questions carry no confidence of their own, so confidence is None for booleans and numbers.

Errors

from jev_jsonschema import IncompatibleSchemaError, JevApiError

try:
    result = jev.evaluate(schema, state=review)
except IncompatibleSchemaError as e:
    ...  # your schema has properties Jev can't answer. See below.
except JevApiError as e:
    ...  # e.status_code, e.retryable, e.request_id

JevApiError messages are written to be shown to your users as-is, and retryable tells you whether trying again could help (timeouts, 429s, 5xxs). The client does one POST and never retries on its own, so the backoff policy stays yours.

What Isn't Supported

Jev answers questions from a fixed set of options. Plenty of JSON Schema doesn't fit, and this library refuses it loudly rather than inventing a mapping:

• Free-form string (anything without an enum), array, object, null
• anyOf, oneOf, allOf, $ref, const, not, and multi-type "type": [...]
• number with any bounds other than minimum: 0 / maximum: 1
• integer ranges wider than 10 values, and enums with more than 255 values
• integer without both minimum and maximum

You find out before anything is sent, and you find out about every bad property, not just the first:

try:
    jev.evaluate(schema, state=review)
except IncompatibleSchemaError as e:
    for failure in e.failures:
        print(failure.key, failure.reason)
# summary uses 'anyOf', which is not supported
# tags type 'array' is not supported

Error messages here are also written to be shown to your users as-is.

Without the Client

The conversion is a separate, pure layer. If you'd rather make the HTTP call yourself (your own retries, your own auth, TypeSafe's official SDK), use the two converters directly and skip JevClient entirely:

from jev_jsonschema import JSONSchema2Jev, JevResult2JsonSchema

question_set = JSONSchema2Jev().convert(schema)

body = question_set.request(
    state="Loved it. Shipped in a day.", model="jev-latest"
).to_body()
answers = your_http_post("https://api.typesafe.ai/v1/systemone", json=body)["answers"]

result = JevResult2JsonSchema().convert(question_set, answers)
result.output

The QuestionSet is the thing to hold onto between the two halves: it remembers how each property was mapped, which is why decoding needs it. It's a plain frozen dataclass.

Options

from jev_jsonschema import JevClient, MappingOptions, ScoreDecode

options = MappingOptions(
    noul_threshold=0.5,  # where a noul probability becomes True
    max_score_levels=10,  # lower the cap on integer ranges
    score_decode=ScoreDecode.argmax,  # or ScoreDecode.expected, for the rounded mean
    instructions_fallback_to_key=True,  # use the property name when there's no description
)

jev = JevClient(options=options)

Using the converters directly? Pass the same options to both halves. The decoder needs to know how the questions were built.

Design Notes

• Two layers, and you can take just one. JSONSchema2Jev and JevResult2JsonSchema are pure and know nothing about HTTP. JevClient is a thin wrapper that adds the POST.
• Small dependency footprint: httpx and pydantic, both of which most apps already have.
• No hidden retries, no hidden concurrency. One call is one POST.
• Never logs your data. Failures log the status and TypeSafe's request id, never the body, which would echo your state and questions.
• Fully typed, ships a py.typed marker.

Development

Uses uv, ruff, and ty.

uv sync           # install everything
uv run pytest     # tests
uv run ruff check --fix && uv run ruff format   # lint + format
uv run ty check   # typecheck
uv build          # build the wheel and sdist

No test hits the network. The client's tests run against respx. CI runs all of the above on Python 3.10 through 3.14.

License

MIT. See LICENSE.