jevframe

How we reviewed this: we ran its full CI, then drove the accessor against a stand-in that logged every request, with canary strings planted in the columns we did not pass. We made no Jev calls.

What it is

df["dissatisfied"] = await df.jev.noul(
    "Is this customer dissatisfied?",
    state=["title", "review"],
)

An accessor for pandas and Polars that asks a typed question about each row and gives you back ordinary Series and DataFrames. Built on the official typesafe-sdk.

What leaves the row

This is the design decision worth the review. state= is a list of column names, so the request is built from those columns and not from the row. We gave it a frame with four columns, named two, and put distinct strings in the other two:

state on the wire: {"title": "Charged twice", "review": "Please refund the duplicate charge today."}

  SSN_CANARY          on the wire: 0
  ID_CANARY           on the wire: 0
  Charged twice       on the wire: 1
  duplicate charge    on the wire: 1

Nothing from secret_notes or internal_id left the machine. That sounds obvious until you consider the alternative — a library that serialises the whole row because it is easier — which is what we have found elsewhere. Here the columns you send are the columns you typed.

The pandas index survived too: a frame indexed [10, 20, 30] came back as a Series indexed [10, 20, 30].

One request per row, several questions inside it

res = await df.jev.evaluate(
    state=["title", "review"],
    questions={
        "urgent": Noul(instructions="Does the customer need help today?"),
        "topic": Choice(instructions="What is the main issue?",
                        criteria={"billing": "Charges and refunds", "bug": "Broken behavior", "other": None}),
    },
)

Three rows, two questions each, produced three requests of two questions. Questions about a row share its request; rows do not share with each other.

That is the right shape for a dataframe accessor and the wrong shape for bulk work, which is worth saying plainly: a million rows is a million requests. If that is your job, jev-ultralightspeed exists for exactly it. jevframe is for the notebook, where you have a few thousand rows and want them back as columns.

And you do get them as columns — the whole distribution, not just the winner:

['urgent__probability', 'topic__label', 'topic__confidence',
 'topic__p__billing', 'topic__p__bug', 'topic__p__other']

When a row fails

No silent nulls, and it tells you which row:

what the provider did	what happened
HTTP 500	EvaluationError: Row 2: inference failed (TypeSafeInternalServerError)
a 200 with an empty answers	EvaluationError: Row 0: inference failed (ValueError)
no API key	No API key was provided. Pass api_key or set the TYPESAFE_API_KEY environment variable.

For a library whose output is a column you are about to filter on, a NaN that means “the API was down” and a NaN that means “the model was unsure” would be indistinguishable and dangerous. This one raises, and names the row.

Engineering

MIT, Python 3.10+, pandas and Polars both optional extras — importing an integration registers its accessor. 108 tests pass, ruff check and ruff format --check are clean across 22 files, marimo check --strict passes on the example notebook, and uv build works. That is its whole CI and it ran green on the first try.

Our first attempt at the multi-question API failed because we passed the instruction positionally; the constructors are keyword-only, and the README’s examples are correct. Worth a note only because it is the kind of thing a reader will hit too.

Thirteen stars, one author, version 0.1.0, four days old when we tested it.

Verdict

A small, well-made thing that does one job and does not overreach. The column-selection design means the privacy question has a precise answer — the columns you name, and nothing else — and we checked it rather than assuming it.

Use it in a notebook on the scale a notebook implies. Do not reach for it to label a million rows; that is a different tool, and the author has not claimed otherwise.

For bulk work over the same kind of data, see jev-ultralightspeed; for checking that the question you are about to run over a column actually works, jev-calibrate.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 22, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.