Unitlab.ai is an AI-driven data annotation platform that automates the collection of raw data, facilitating collaboration with human annotators to produce highly accurate labels for your machine learning models. With our service, you can optimize work efficiency, improve data quality, and reduce costs.

Unitlab Python SDK

Native Python and CLI access to Unitlab projects, Ontologies, Assets, datasets, published versions, Batch Queues, Releases, cloud storage, and multimodal Data Groups and Workflow Tasks.

Install

pip install --upgrade unitlab

Python 3.10 or newer is required.

Authenticate

Create an API key in Workspace settings, then use environment variables:

export UNITLAB_API_KEY="YOUR_API_KEY"
export UNITLAB_API_URL="https://api.unitlab.ai"  # optional

Or configure the CLI once:

unitlab configure --api-key YOUR_API_KEY
unitlab configure --api-url https://api.unitlab.ai

You can also pass both values directly:

from unitlab import UnitlabClient

client = UnitlabClient(
    api_key="YOUR_API_KEY",
    api_url="https://api.unitlab.ai",
)

1. Upload multimodal data to a project

from unitlab import UnitlabClient

client = UnitlabClient()
project = client.projects.create("Medical review")
batch = project.upload("./multimodal-data")
status = batch.wait()

print(batch.uploaded, batch.batch_queue_id)

The directory may mix images, video, audio, text, PDFs, DICOM, NIfTI, and NRRD. One call creates one Batch Queue. Call batch.wait() when the next operation depends on server-side processing. Lidar scenes are uploaded with generic_type="pointcloud"; see Lidar scenes.

2. Curate in Assets, publish a dataset, and attach it

from unitlab import UnitlabClient

client = UnitlabClient()
project = client.projects.create("Quality review")

uploaded = client.assets.upload(
    "./multimodal-data",
    folder="Raw data",
    tags=["incoming"],
)

dataset = client.datasets.create(
    "Review set",
    folder_ids=[uploaded.folder_id],
)
version = dataset.publish_version("Initial snapshot")

attached = client.attach_dataset(
    project,
    dataset,
    version=version.version_number,
)

print(attached.created_count)

Datasets are version-first. Editing a dataset creates Unpublished changes; publish_version() freezes a version, and projects attach that published snapshot.

3. Import existing cloud data

from unitlab import UnitlabClient

client = UnitlabClient()
project = client.projects.create("Cloud review")
storage = next(
    item for item in client.cloud_storages.list() if item.name == "Production data"
)

batch = project.import_cloud(
    storage,
    ["incoming/study-001/"],
)
status = batch.wait()

print(batch.batch_queue_id)

Directory paths end in /. Cloud credentials are never returned through the SDK; list and browse responses expose only safe storage metadata.

Resource handles

Namespace entrypoints return typed resource handles:

project = client.projects.get("PROJECT_ID")
projects = client.projects.list()
dataset = client.datasets.get("DATASET_ID")
datasets = client.datasets.list()
folder = client.assets.folder("FOLDER_ID")
folders = client.assets.folders()

Ontologies

from unitlab import OntologyStructure, RadioAttribute, Shape

structure = OntologyStructure()
cat = structure.add_object("Cat", Shape.BOUNDING_BOX)
colour = cat.add_attribute(RadioAttribute, "Colour", required=True)
colour.add_option("Black")
colour.add_option("White")

ontology = client.ontologies.create("Cat labels", structure=structure)
project = client.projects.create(
    "Cat review",
    ontology_hash=ontology.id,
)

Use client.ontologies.get(), client.ontologies.list(), and ontology.save() for the rest of the Ontology lifecycle. Project attachment copies the Workspace Ontology so Unitlab can preserve project-specific version history.

Projects and Batch Queues

projects = client.projects.list()
project = client.projects.get("PROJECT_ID")
project = client.projects.create("Road scenes")

batch = project.upload("./data", fps=2.0)
tiled_batch = project.upload("./maps")

status = batch.status()
status = batch.wait(timeout=1800)

queues = project.batch_queues()
queue = project.batch_queue(batch.batch_queue_id)
items = queue.data()
status = queue.wait()

ProcessingStatus exposes total, completed, processing, and failed. Completion means processing == 0; a completed Batch Queue may still report individual failures.

GeoTIFF and whole-slide files are detected automatically. Large tiled files upload directly to storage in bounded parts. Tiled processing can take hours, so use a longer timeout such as batch.wait(timeout=25200) when the next operation depends on the generated COG/DZI derivative.

Video, medical (NIfTI, NRRD, DICOM), audio and PDF files larger than 64 MiB also upload directly to storage, up to 500 MiB each; smaller ones keep the ordinary upload. A large medical file becomes one volume on its own and is not grouped with other files of the batch into a series.

DICOM files named without a usable extension (IM00001, or a UID) are found by their content when a folder is uploaded and are sent as medical; nothing needs to be renamed.

Lidar scenes

A lidar scene is one item whatever it is made of. It is uploaded only when the type is named; without it .pcd, .ply, .bin and scene .zip files are skipped. The workspace needs lidar scenes enabled.

from unitlab import detect_scenes

# Look first: what would be uploaded, nothing is sent.
for scene in detect_scenes("./recordings").scenes:
    print(scene.name, scene.frame_count, scene.point_sensors, scene.cameras)

batch = project.upload("./recordings", generic_type="pointcloud")
batch.wait()

for item_id in batch.data_item_ids:
    status = project.scene_status(item_id)
    print(status.status, status.frame_count, status.error)

project.retry_scene(item_id)  # a failed scene, about a minute after it failed

What source may be:

• One file: a .pcd, .ply or .bin cloud, or a .zip that contains such frames or a scene.json. Any other .zip is refused.
• A folder with a scene.json, or with the point clouds of one sensor (drive_01/lidar/000.pcd, ...). Image folders whose file names match the frames become its cameras (drive_01/cam_front/000.jpg).
• A folder with one subfolder per sensor that share file names. This is read as one scene with several sensors; pass split_sensor_folders=True when each folder is a recording of its own.
• A folder that holds several such scenes: each one is uploaded as its own item. Scene files left over in the folder are one scene each.
• A nuScenes dataset (the folder with samples/ and v1.0-*/): one scene per recording, key frames only, with the calibration taken from its tables.

Other files in the folder are not uploaded by this call; upload the folder once more without the type for ordinary media. A scene is limited to 20 GiB and 1,000 frames, and one data unit is used per frame.

batch.wait() covers scenes like any other upload. A failed scene is not in project.data_units(); find it in the Batch Queue (queue.data()) and ask project.scene_status(item_id) for the reason.

Scenes go into Assets the same way, and the existing Asset methods apply:

result = client.assets.upload(
    "./recordings", generic_type="pointcloud", folder="Lidar"
)
asset = result.assets[0]
status = asset.wait()          # status.frame_count, status.error
asset.retry()                  # a failed scene
asset = client.assets.asset("ASSET_ID")  # a handle by ID, for status and retry

release.download_files() writes a scene that was uploaded as a folder under <item file name>/ with the paths it was uploaded with.

Data Units and project lifecycle

project.update(name="Road review v2", description="Second collection")

units = project.data_units(data_type="image", status="annotate")
unit = client.get_data_unit(project.id, units[0].id)

for unit in units:
    print(unit.kind, unit.name, unit.status)

A loose file is one datasource Data Unit. A Data Group is one group Data Unit whose items contain its tile summaries; member files are not duplicated in the top-level list. project.delete() soft-deletes the Project and schedules the existing backend cleanup.

Assets and folders

folders = client.assets.folders()
folder = client.assets.create_folder("Raw data")
children = folder.children()
every_folder = client.assets.all_folders()
items = folder.list_items()

result = client.assets.upload("./images", folder_id=folder.id, tags=["train"])
asset = result.assets[0]

tiled = client.assets.upload(
    "./slides",
    folder_id=folder.id,
)
tiled.assets[0].wait()

cloud_folder = client.assets.create_cloud_folder(
    "Incoming",
    "CLOUD_STORAGE_ID",
    prefix="incoming/",
)
cloud_folder.sync_cloud()  # Tiled files are registered as processing Assets.

Datasets and published versions

dataset = client.datasets.create(
    "Training set",
    folder_ids=["FOLDER_ID"],
)

dataset.add_sources(asset_ids=["ASSET_ID"])
changes = dataset.unpublished_changes()
version = dataset.publish_version("Added edge cases")
versions = dataset.versions()
items = dataset.list_items(version=version.version_number)

Workflow Tasks

workflow = project.workflow
annotate = workflow.get_stage(name="Annotate")

for task in annotate.get_tasks():
    task.assign("USER_ID")
    task.submit()

Review Tasks provide approve() and reject(). Managers can use move() for a direct stage override and get_timeline() to inspect Task history.

Agent webhooks (no task polling)

An Agent stage tells your HTTPS webhook which tasks are ready: one event per task, sent as soon as it enters the stage. The webhook carries IDs only; your service reads and saves annotations with a workspace API key, then reports the run. A successful webhook response acknowledges delivery; only complete_agent advances the task through the stage's single outgoing edge.

from unitlab import AgentWebhook, UnitlabClient

client = UnitlabClient()  # workspace API key


def handle(body: bytes, headers, secret: str) -> int:
    """Return the HTTP status to answer the webhook with."""
    # body: original request bytes; headers: incoming HTTP headers.
    # secret: the signing secret configured in the Agent stage. A stage saved
    # without one sends unsigned events: use AgentWebhook.parse(body) instead.
    event = AgentWebhook.verify(body, headers, secret)
    if event.is_test:  # "Test connection": acknowledge only
        return 204
    ref = event.task
    task = client.get_workflow_task(ref.task_id)
    for item in task.get_annotations():  # one item, or one per Data Group panel
        # item.document: the item's UUEF; your_model returns a new one, or None
        # when it changes nothing (a save always writes a new version).
        document = your_model(item.document, item.source_url)
        if document is not None:
            task.save_annotations(item, document)  # replaces the item whole
    task.complete_agent(ref.run_id)
    # On model failure: task.fail_agent(ref.run_id, "A useful processing error")
    return 200

The payload is a CloudEvents 1.0 structured JSON event (Content-Type: application/cloudevents+json, type: ai.unitlab.agent.task.ready.v1) whose data is:

{
  "project_id": "PROJECT_ID",
  "workflow_id": "WORKFLOW_ID",
  "stage_id": "STAGE_ID",
  "task": {"task_id": "TASK_ID", "run_id": "RUN_ID", "datasource_ids": ["DATASOURCE_ID"]}
}

Each event carries exactly one task (there is no batching). AgentWebhook.verify (or parse) returns id, project_id, workflow_id, stage_id and task, an AgentTaskRef(task_id, run_id, datasource_ids). Any missing or wrongly typed field, or a task that is not one object, raises ValueError.

The stage's Test connection button sends the same envelope (signed only when the stage has a secret) with type: ai.unitlab.agent.test.v1 and data: {"test": true, "project_id", "workflow_id", "stage_id", "task": null}, where the IDs may be null for an unsaved workflow. For it event.is_test is True and event.task is None: answer 2xx and do nothing else.

Signatures use Standard Webhooks: webhook-id, webhook-timestamp, and webhook-signature. The secret is optional; a stage without one sends no webhook-signature. Verification uses HMAC-SHA256 over id.timestamp.raw_body, accepts v1,BASE64 signatures and a 300-second clock tolerance, and supports plain configured secrets or whsec_ base64 keys. Pass the original body bytes; parsing and re-encoding JSON changes the signed content. Keep your receiver clock synchronized.

get_annotations() returns one TaskAnnotation per item with datasource_id, generic_type, revision, source_url (a signed media URL; empty or None for an item without a file), format and document: the item's UUEF, the same per-item JSON a UUEF release exports. Its JSON Schema is uuef.schema.json, at the root of every UUEF release export (client.releases.create(project, export_type="UUEF")). A large document arrives as a signed link that the SDK downloads, so document is always one whole dict. Hidden objects are included with visible: false. Tracks (video, medical) list their stored keyframes only; get_annotations(only_keyframes=False) lists every frame a track is visible on. Both save back safely: a save keeps only frames marked keyframe: true.

save_annotations(item, document) replaces the item with document and returns the new revision. It is a whole overwrite: objects missing from the document are deleted, and a save made by someone else since your read is overwritten (no revision check, no ConflictError). Keep the id of every object you keep. Every object in the saved version is credited to the API key's creator, so skip the save when your model changed nothing. Afterwards item.revision and item.document hold what was saved. The save works on a task in any workflow stage. Two other answer formats are accepted, each also replacing the item whole:

# Geospatial and pathology items: an RFC 7946 FeatureCollection (lon/lat, raster CRS units or pixels).
task.save_annotations(item, feature_collection, format="geojson")
# Image items: the custom-model reply JSON ("bboxes", "polygons", "brush", ...).
task.save_annotations(item, model_reply, format="model")

A format the item's type does not take is refused (ValidationError, 400) and nothing is saved. One bad shape never refuses the whole document: a self-crossing outline is repaired, a shape past the media's edge is clipped, a shape that cannot be stored (no area) is skipped, and the rest is saved.

After a geojson or model save item.document is None: read again to see the UUEF the server built. A document over 64 MiB is sent to a signed upload link first; the SDK does this for you.

Lidar scenes (generic_type == "pointcloud") are not UUEF yet: their format is "native", document is None and result holds the editor's track-first result, {"tracks": {"cuboids3d": {"<track id>": {"ranges": [[0, 21]], "value": 7, "keyframes": {"0": {...}, "20": {...}}}}}, "item_properties": {...}}, stitched from its pages. Save the complete new result with save_annotations(item, result): only changed tracks are sent (nothing when unchanged), in parts within the server's limits, each an ordinary save carrying the previous part's revision. A stale revision raises ConflictError with code == "revision_conflict"; parts sent before it stay saved: read again and save again. A track larger than one save takes raises ValueError before anything is sent. Leave segments3d tracks as you read them (their points live in label arrays).

The raw HTTP calls, all with Authorization: Api-Key <key>:

Delivery can repeat. Deduplicate processing by run_id. A run not completed within 7 days (fixed, not a stage setting) fails, as does one reported with fail_agent (its error is cut to the 4,000 characters the server takes); the task stays in the stage until a manager clicks Retry, which starts a new run with a new run_id and refuses results for the old one. Production receivers should acknowledge with 2xx only after durable enqueueing, then process asynchronously. Never log signing secrets, API keys, full payloads, or signed media URLs.

For a local smoke test, give the stage a signing secret (the sample verifies signatures) and run the receiver with it as UNITLAB_AGENT_WEBHOOK_SECRET, plus UNITLAB_API_KEY and, for a non-default API, UNITLAB_API_URL in its environment:

python examples/agent_webhook.py
cloudflared tunnel --url http://127.0.0.1:8091

Run the tunnel command in a second terminal and put its HTTPS address in the Agent stage. The sample's transform() returns None, so it saves nothing and only completes the run; replace it with your model.

Attach data

preview = project.attach_preview(dataset_ids=[dataset.id])

result = project.attach(dataset_ids=[dataset.id])

# Attach an exact version instead of latest:
result = project.attach(
    dataset_versions=[
        {"dataset_id": dataset.id, "version_number": 2},
    ]
)

for source in project.attached_sources():
    print(source.name, source.detach_preview())
    # source.detach() removes only the Project copy and keeps annotations.

Video attachment may require fps=. Call attach_preview() when you need counts before committing; attach() validates the same rule and raises an actionable ValidationError if FPS is missing.

Cloud storage

storages = client.cloud_storages.list()
storage = storages[0]

for entry in storage.browse(prefix="incoming/"):
    print(entry.name, entry.type, entry.size)

batch = project.import_cloud(
    storage,
    ["incoming/file.png", "incoming/study/"],
)
status = batch.wait()

Import annotation JSON

Media can come from S3 while the annotation JSON stays on your computer:

project = client.projects.get("PROJECT_ID")
media = project.import_cloud("CLOUD_STORAGE_ID", ["images/"])
status = media.wait()
if status.failed:
    raise RuntimeError("Some media failed processing; inspect the Batch Queue.")

batch = project.annotation_imports.create(
    "./predictions.json",
    source_name="my-model",  # Reuse this namespace for the same prediction source.
)
print(batch.preview["counts"], batch.preview["issues"])
# Inspect the preview before saving labels. Approval is a separate review step.
report = batch.commit()
print(report)

Supported inputs are COCO JSON, Unitlab exports (UUEF item JSON, Audio JSON and GeoJSON, as .json or .geojson), unitlab-image/1 and unitlab-video/1. They cover image, video, medical, document, audio, time series, text, HTML, geospatial and pathology items. A UUEF file holds one item: import each file as its own batch. Each JSON file is limited to 80 MiB. The backend matches original media paths/IDs, with a unique-filename fallback; ambiguous names require an explicit mapping. YOLO text/ZIP files and other tools' project files are not accepted by this importer.

For COCO video, pass frame_mapping_file="./frames.json" in unitlab-frame-map/1 format. Use track_profile="attributes.track_id" only for COCO exports with that explicit track-ID convention. Import video media at its original frame rate; annotations for frames omitted by sampling are rejected.

project.annotation_imports.sources() lists saved prediction sources, and project.annotation_imports.media(search="road.jpg") finds media IDs for mappings. Pass source_id= instead of source_name= to reuse a source by ID. Set options={"stable_ids": True} only when your exporter keeps annotation IDs stable between prediction runs. Otherwise the default protects existing objects from unconfirmed ID reuse.

After processing completes or you supply mappings, call batch.refresh() to rebuild the preview. batch.refresh(options={...}) replaces all import options; omitting options preserves them. Supported options are stable_ids, media_mapping, class_mapping, object_mapping, and review_mapping.

batch.commit() preserves conflicts and human deletions by default. To resolve a specific preview annotation, pass decisions={annotation_key: "keep"} (or "incoming" / "skip"). Invalid records prevent the whole commit unless you explicitly choose allow_partial=True. The report includes created, updated, kept, unchanged and skipped counts when applicable. A stale preview raises ConflictError; inspect a refreshed preview before committing again. If a commit response is lost, repeat the same batch's commit() with the same decisions; the server returns the saved report without duplicating labels.

For annotations already in connected storage, omit the local file and pass cloud_storage=storage, annotation_key="labels/predictions.json", and optionally frame_mapping_key="labels/frames.json". Keys are relative to that connection's prefix. Media and annotations must resolve within the API key's workspace and the selected project.

This requires a backend with the /api/sdk/projects/{id}/annotation-imports/ routes and the annotation-import database migration installed. Commit requires a user-created API key so saved labels have an audit author. Legacy keys without a creator can preview but receive PermissionDeniedError when committing.

Data Groups

Let Unitlab suggest a filename pattern:

folder = client.assets.folder("FOLDER_ID")
grouped = folder.auto_group()

Or compile a literal filename template:

from unitlab import tiles_from_template

config = tiles_from_template(
    "{patient_id}_{view}",
    tile_values={
        "view": ["L_CC", "R_CC", "L_MLO", "R_MLO"],
    },
)

estimate = folder.estimate_grouping(config)
grouped = folder.auto_group(config)

For folders above 5,000 files or estimates above 1,000 groups, auto_group() raises an actionable ValidationError. Split the source into smaller folders and retry; the SDK does not return a job that will never run.

Releases

Releases are exported annotation snapshots. They are separate from Assets datasets.

release = client.releases.create(
    project,
    export_type="UUEF",
    split_ratios={"train": 80, "test": 20},
)

releases = client.releases.list()
release = client.releases.get(release.id)

annotation_path = release.download(split="train", dest="./release-annotations")
files_folder = release.download_files(dest="./release-files")

client.releases.create() waits until the Release is ready. To return immediately and wait later:

release = client.releases.create(project, wait=False)
release.wait(timeout=7200)  # Wait before downloading.

Check progress with release.refresh(), then read release.status and release.progress. Lists show ready Releases; use client.releases.get(release_id) to check one still being prepared.

Release item download tokens are persistent Unitlab URLs. For files stored in customer cloud storage, the URL redirects to a signed target URL; only that target URL is temporary.

CLI quick reference

# Projects and Batch Queues
unitlab project create "Medical review" --ontology ONTOLOGY_ID
unitlab project list
unitlab project detail PROJECT_ID
unitlab project update PROJECT_ID --name "Medical review v2"
unitlab project data-units PROJECT_ID --data-type image
unitlab project data-unit PROJECT_ID DATA_UNIT_ID
unitlab project sources PROJECT_ID
unitlab project detach-source PROJECT_ID SOURCE_LINK_ID --preview
unitlab project upload PROJECT_ID --source ./data
unitlab project upload PROJECT_ID --source ./maps
unitlab project upload PROJECT_ID --source ./recordings --type pointcloud --dry-run
unitlab project upload PROJECT_ID --source ./recordings --type pointcloud
unitlab project upload PROJECT_ID --source ./rig --type pointcloud \
  --split-sensor-folders
unitlab project scene-status PROJECT_ID ITEM_ID
unitlab project scene-retry PROJECT_ID ITEM_ID
unitlab batch-queue list PROJECT_ID
unitlab batch-queue detail PROJECT_ID BATCH_QUEUE_ID
unitlab batch-queue status PROJECT_ID BATCH_QUEUE_ID
unitlab batch-queue data PROJECT_ID BATCH_QUEUE_ID
unitlab batch-queue wait PROJECT_ID BATCH_QUEUE_ID --timeout 1800

# Cloud import and attach
unitlab project import-cloud PROJECT_ID CLOUD_STORAGE_ID incoming/study/
unitlab project attach PROJECT_ID --dataset DATASET_ID
unitlab project attach PROJECT_ID --dataset DATASET_ID --preview
unitlab project attach PROJECT_ID --dataset-version DATASET_ID:2

# Assets
unitlab assets upload ./data --folder "Raw data" --tag incoming
unitlab assets upload ./slides --folder "Slides"
unitlab assets upload ./recordings --folder "Lidar" --type pointcloud
unitlab assets status ASSET_ID
unitlab assets retry ASSET_ID

# Folders
unitlab folders create "Raw data"
unitlab folders create "Incoming" \
  --cloud-storage CLOUD_STORAGE_ID --prefix incoming/
unitlab folders list
unitlab folders list --parent PARENT_FOLDER_ID
unitlab folders detail FOLDER_ID
unitlab assets upload ./data --folder-id FOLDER_ID
unitlab folders sync-cloud FOLDER_ID

# Data Groups
unitlab assets group FOLDER_ID --config grouping.json
unitlab assets group FOLDER_ID --config grouping.json --preview
unitlab assets group FOLDER_ID --preview
unitlab assets group FOLDER_ID \
  --template '{patient_id}_{view}' \
  --tile-values view=L_CC,R_CC,L_MLO,R_MLO

# Datasets
unitlab dataset list
unitlab dataset create "Training set" --folder FOLDER_ID
unitlab dataset detail DATASET_ID
unitlab dataset add-sources DATASET_ID --asset ASSET_ID
unitlab dataset unpublished-changes DATASET_ID
unitlab dataset publish DATASET_ID --title "Initial snapshot"
unitlab dataset versions DATASET_ID

# Releases (create returns immediately; add --wait to wait until ready)
unitlab release create PROJECT_ID --format UUEF --splits train=100
unitlab release wait RELEASE_ID --timeout 7200
unitlab release list
unitlab release detail RELEASE_ID
unitlab release download RELEASE_ID --split-type train --dest ./release-annotations

# Ontologies
unitlab ontology list
unitlab ontology list --title-like Medical --created-after 2026-01-01
unitlab ontology create "Medical labels" --structure ontology.json
unitlab ontology update ONTOLOGY_ID --title "Medical labels v2"

# Cloud storage
unitlab cloud list
unitlab cloud browse CLOUD_STORAGE_ID --prefix incoming/

Run unitlab COMMAND --help for every option.

License

MIT