Skip to content
NautAI

Quickstart

Push a dataset, review a diff, label a queue and run your first eval. It takes about fifteen minutes and works on the free plan.

NautAI is in private preview. Commands below show the developer experience; console access opens soon. Questions? Contact us.

Install

The CLI ships with the Python package. It needs Python 3.10 or later.

shell
$ pip install nautai
$ naut login

Create a dataset

Run naut init in a folder with your data. NautAI detects the format and writes a naut.yaml describing the schema and cleaning steps.

naut.yaml
dataset: support-tickets
schema:
text: string
label: enum[billing, refund, bug, cancellation]
clean:
- dedupe: { method: embedding, threshold: 0.97 }
- redact: [email, phone, card_number]
- filter: "len(text) > 12"

Push and diff

naut push runs your cleaning steps, then commits the result. Compare any two versions with naut diff.

shell
$ naut push ./tickets -m "Initial import"
# Published support-tickets v1.0 (1,180,422 rows)
$ naut diff v1.0..main --stat

Label

Create a queue from a filter, attach a guideline, and optionally pre-label with a model. Reviewers work in the console; results land as a commit on a branch.

shell
$ naut label create refunds \
--where "label in ('billing','refund')" \
--guideline ./guidelines/refunds-v3.md \
--prelabel my-classifier@latest --reviewers 3

Evaluate

Attach an eval suite once and it runs on every commit. Add a merge check to block changes that lower the score.

shell
$ naut eval add intent-accuracy --model https://models.internal/classify
$ naut eval run intent-accuracy --on v2.4
# intent-accuracy: 0.94 (+0.02 vs v2.3)

Python SDK

Use the SDK in training scripts, and to send production feedback from your app.

train.py
import nautai
ds = nautai.dataset("support-tickets", version="v2.4")
train, test = ds.split(test=0.1, seed=7)
# Later, in your app:
nautai.feedback("support-tickets",
input=message, prediction=label,
signal="thumbs_down")

CLI reference

naut CLI commands
CommandWhat it does
naut init <name>Create a dataset in the current directory
naut push [path] -m <msg>Hash, clean and commit new or changed rows
naut diff <a>..<b>Show row-level changes between two versions
naut branch <name>Start a branch for an experiment
naut mr openOpen a merge request for review
naut label <queue>Create or open a labeling queue
naut synth --fill <field>Generate rows for under-represented values
naut eval run <suite>Run an eval suite against a version
naut tag <version> <name>Pin a version, for example to a training run
naut checkout <version>Download any version locally