LaminDB - Data management for multimodal AI

LaminDB is an open-source data management tool that makes it easy to query, trace & govern datasets across diverse storage formats and locations. It gives you context through annotations, memory through lineage, and governance through branching and versioning. It uses a scalable lakehouse architecture that integrates files, tables, arrays, ontologies, and notes. It supports biological data through modules for bio-formats and registries by the creators of Scanpy.

Why?

1. Untraceable results cannot be trusted, especially in the age of agents.
2. Without effective access to multimodal data, models burn tokens or fail entirely.
3. We want to govern changes to data like we govern changes to code with git.

How?

• lineage → trace results across agent sessions, notebooks, scripts & workflows
• lakehouse → query across many datasets, manage tables & arrays schema-based and ACID
• LIMS & ELN → unified schema-based records management with support for ontologies & notes
• FAIR datasets → validate & annotate files, DataFrame, AnnData, SpatialData, zarr, …
• governance → manage changes via branching & by versioning data + code

Architecture?

• zero lock-in → uses open standards (metadata in SQLite/Postgres, data in parquet, zarr, etc.)
• scalable → hit storage & database directly through your pydata or R stack, no REST API involved
• simple → pip install lamindb or install.packages('laminr') - no Docker required, no separate backend
• unified → federate data across storage locations (local, S3, GCP, …) in any database
• distributed → federate data zero-copy & lineage-aware across databases
• reproducible → track agent traces, source code & compute environments
• idempotent → re-run logic without worries about duplications or overwrites
• integrations → bio ontologies git, nextflow, vitessce, redun, and more
• extensible → create custom plug-ins based on the Django ORM, the basis for LaminDB's registries

Who?

Scientists and engineers at leading research institutions and biotech companies, including:

• Industry → Pfizer, Altos Labs, Ensocell Therapeutics, ...
• Academia & Research → scverse, DZNE (National Research Center for Neuro-Degenerative Diseases), Helmholtz Munich (National Research Center for Environmental Health), ...
• Research Hospitals → Global Immunological Swarm Learning Network: Harvard, MIT, Stanford, ETH Zürich, Charité, U Bonn, Mount Sinai, ...

From personal research projects to pharma-scale deployments managing petabytes of data across:

entities	OOMs
observations & datasets	10¹² & 10⁶
runs & transforms	10⁹ & 10⁵
proteins & genes	10⁹ & 10⁶
biosamples & species	10⁵ & 10²
...	...

UI, permissions, audit logs? LaminHub is a collaboration hub built on LaminDB similar to how GitHub is built on git.

Quickstart

To install the Python package with recommended dependencies, use:

pip install lamindb

Install with minimal dependencies.

The lamindb package adds data-science related dependencies through the [full] extra, see here.

For a minimal install of the lamindb namespace, use:

pip install lamindb-core

Agent? The lamindb skill ships with the lamindb package at .agents/skills/. Pure markdown docs ship with the package in docs/ and are available from the web via llms.txt.

Query databases & datasets

You can browse public databases at lamin.ai/explore. To access laminlabs/cellxgene, run:

import lamindb as ln

db = ln.DB("laminlabs/cellxgene")  # a database object for queries
df = db.Artifact.to_dataframe()    # a dataframe listing datasets & models

To get a specific dataset, run:

artifact = db.Artifact.get("BnMwC3KZz0BuKftR")  # a metadata object for a dataset
artifact.describe()                             # describe the context of the dataset

See the output.

Access the content of the dataset via:

local_path = artifact.cache()  # return a local path from a cache
adata = artifact.load()        # load object into memory
accessor = artifact.open()     # return a streaming accessor

For broader queries of cellxgene, see docs.lamin.ai/cellxgene.

Save files & folders

You can create a LaminDB instance at lamin.ai and invite collaborators. To connect to an existing instance, run:

lamin login
lamin connect account/name  # tip: add flag `--here` to scope to current directory

Or init a new instance instead (no login required).

lamin init --storage ./quickstart-data --modules bionty

On the terminal and in a Python session, LaminDB will now auto-connect. For more configuration, see docs.lamin.ai/setup.

To save a file or folder via the API:

import lamindb as ln
# → connected lamindb: account/instance

open("sample.fasta", "w").write(">seq1\nACGT\n")        # create dataset
ln.Artifact("sample.fasta", key="sample.fasta").save()  # save dataset

To save a file or folder via the CLI, run:

lamin save sample.fasta --key sample.fasta

To load an artifact via the CLI into a local cache, run:

lamin load --key sample.fasta

Read more about the CLI: docs.lamin.ai/cli.

Trace data, code & agents

The lamindb skill ships with the lamindb package at .agents/skills/. Ask your coding agent to copy it to wherever it reads skills from — .claude/skills/ for Claude Code, .agents/skills/ for GitHub Copilot — so that it automatically tracks agent sessions.

To create a dataset in a script or notebook while tracking source code, inputs, outputs, logs, and environment:

import lamindb as ln
# → connected lamindb: account/instance

ln.track()                                              # track code execution
open("sample.fasta", "w").write(">seq1\nACGT\n")        # create dataset
ln.Artifact("sample.fasta", key="sample.fasta").save()  # save dataset
ln.finish()                                             # mark run as finished

Running this snippet as a script (python create_fasta.py) produces the following data lineage:

artifact = ln.Artifact.get(key="sample.fasta")  # get artifact by key
artifact.describe()      # context of the artifact
artifact.view_lineage()  # fine-grained lineage

Watch a mini video: youtu.be/yK3ODFZLL1A

Access run & transform.

run = artifact.run              # get the run object
transform = artifact.transform  # get the transform object
run.describe()                  # context of the run

transform.describe()  # context of the transform

Track a project or an agent plan.

Pass a project/artifact to ln.track(), for example:

ln.track(project="My project", plan="./plans/curate-dataset-x.md")

Note that you have to create a project or save the agent plan in case they don't yet exist:

# create a project with the CLI
lamin create project "My project"

# save an agent plan with the CLI
lamin save /path/to/.cursor/plans/curate-dataset-x.plan.md
lamin save /path/to/.claude/plans/curate-dataset-x.md

Or in Python:

ln.Project(name="My project").save()  # create a project in Python

You can track workflows by decorating functions:

import lamindb as ln

@ln.flow()
def create_fasta(fasta_file: str = "sample.fasta"):
    open(fasta_file, "w").write(">seq1\nACGT\n")    # create dataset
    ln.Artifact(fasta_file, key=fasta_file).save()  # save dataset

if __name__ == "__main__":
    create_fasta()

Beyond what you get for scripts & notebooks, this automatically tracks function & CLI params and integrates well with established Python workflow managers: docs.lamin.ai/track. To integrate advanced bioinformatics pipeline managers like Nextflow, see docs.lamin.ai/pipelines.

A richer example.

Here is an automatically generated re-construction of the project of Schmidt et al. (Science, 2022):

A phenotypic CRISPRa screening result is integrated with scRNA-seq data. Here is the result of the screen input:

You can explore it here on LaminHub or here on GitHub.

Label artifacts

You can label an artifact by running:

my_label = ln.ULabel(name="My label").save()   # a universal label
project = ln.Project(name="My project").save() # a project label
artifact.ulabels.add(my_label)
artifact.projects.add(project)

Query for it:

ln.Artifact.filter(ulabels=my_label, projects=project).to_dataframe()

You can also query by the metadata that lamindb automatically collects:

ln.Artifact.filter(run=run).to_dataframe()              # by creating run
ln.Artifact.filter(transform=transform).to_dataframe()  # by creating transform
ln.Artifact.filter(size__gt=1e6).to_dataframe()         # size greater than 1MB

If you want to include more information into the resulting dataframe, pass include.

ln.Artifact.to_dataframe(include=["created_by__name", "storage__root"])  # include fields from related registries

The query syntax for DB objects and for your default database is the same.

Here is an overview that illustrates how artifacts can be labeled by other entities:

Read more: docs.lamin.ai/organize.

Manage features & records

Let's define some features:

from datetime import date

gc_content = ln.Feature(name="gc_content", dtype=float).save()
experiment_note = ln.Feature(name="experiment_note", dtype=str).save()
experiment_date = ln.Feature(name="experiment_date", dtype=date, coerce=True).save()  # accept date strings

The most basic thing you can do with features is annotating artifacts, records, or runs with them:

artifact.features.set_values({
    gc_content: 0.55,
    experiment_note: "Looks great",
    experiment_date: "2025-10-24",
})

# query
ln.Artifact.filter(experiment_date == "2025-10-24").to_dataframe(include="features")  # query all artifacts annotated with `experiment_date`

You can create records for entities underlying your experiments (samples, perturbations, instruments, etc.):

ln.Record(name="Sample 1", features={gc_content: 0.5}).save()

You can dynamically create registries and relationships of entities via record types:

# create an experiments registry by defining a record type
experiments_registry = ln.Record(name="Experiments", is_type=True).save()

# create a record inside the Experiments registry
ln.Record(name="Experiment 1", type=experiments_registry).save()

# create a feature that links experiments, creating a relationship
experiment = ln.Feature(name="experiment", dtype=experiments_registry).save()

# create a sample record that links the sample to `Experiment 1` via the `experiment` feature
ln.Record(name="Sample 2", features={gc_content: 0.5, experiment: "Experiment 1"}).save()

# export a registry to a dataframe
experiments_registry.to_dataframe()

You can edit records like Excel sheets on LaminHub.

Lakehouse

Here is how you ingest a DataFrame:

import pandas as pd

df = pd.DataFrame({
    "sequence_str": ["ACGT", "TGCA"],
    "gc_content": [0.55, 0.54],
    "experiment_note": ["Looks great", "Ok"],
    "experiment_date": [date(2025, 10, 24), date(2025, 10, 25)],
})
ln.Artifact.from_dataframe(df, key="my_datasets/sequences.parquet").save()  # no validation

To validate & annotate the content of the dataframe, use the built-in schema valid_features:

ln.Feature(name="sequence_str", dtype=str).save()  # define a remaining feature
artifact = ln.Artifact.from_dataframe(
    df,
    key="my_datasets/sequences.parquet",
    schema="valid_features"  # validate columns against features
).save()
artifact.describe()

Watch a mini video: youtu.be/Ji6E7hTnReQ

You can filter for datasets by schema and then launch distributed queries or batch load distributed datasets. For tables, see: docs.lamin.ai/tables. For arrays, see: docs.lamin.ai/arrays.

To validate an AnnData, call:

import anndata as ad
import numpy as np
import pandas as pd

adata = ad.AnnData(
    X=np.ones((21, 10)),
    obs=pd.DataFrame({'cell_type_by_model': ['T cell', 'B cell', 'NK cell'] * 7}),
    var=pd.DataFrame(index=[f'ENSG{i:011d}' for i in range(10)])
)
artifact = ln.Artifact.from_anndata(
    adata,
    key="my_datasets/scrna.h5ad",
    schema="ensembl_gene_ids_and_valid_features_in_obs"
).save()
artifact.describe()

To validate a SpatialData or any other array-like dataset, you need to construct a Schema. You can do this by composing simple pandera-style schemas: docs.lamin.ai/curate.

Branching & versioning

LaminDB co-versions code and datasets for you. If edit and run the create_fasta.py script, you'll automatically create a new version of the transform and the sample.fasta artifact.

The edited script

# create_fasta.py
import lamindb as ln

ln.track()
open("sample.fasta", "w").write(">seq1\nTGCA\n")  # a new sequence
ln.Artifact("sample.fasta", key="sample.fasta", features={"experiment": "Experiment 1"}).save()  # annotate with the new experiment
ln.finish()

artifact_latest = ln.Artifact.get(key="sample.fasta")  # pass version for a previous version: ln.Artifact.get(key="sample.fasta", version="1.0")
artifact_latest.versions.to_dataframe()                # all versions of that artifact
artifact_latest.transform.versions.to_dataframe()      # all versions of the transform that created the artifact

To isolate changes, create a contribution branch and switch to it as in git:

lamin switch -c my_branch

To merge a contribution branch into main, run:

lamin switch main  # switch to the main branch
lamin merge my_branch  # merge contribution branch into main

Read more: docs.lamin.ai/manage-changes.

Watch a mini video: youtu.be/rzRwcMj6-fc

Data sharing

To share data in a lineage-aware way, transfer objects from a source database to your default database:

db = ln.DB("laminlabs/lamindata")
artifact = db.Artifact.get(key="example_datasets/mini_immuno/dataset1.h5ad")
artifact.save()

This is zero-copy for the artifact's data in storage. Read more: docs.lamin.ai/transfer.

Ontologies

Plugin bionty gives you >20 public ontologies as SQLRecord registries. This was used to validate the ENSG ids in the adata just before.

import bionty as bt

bt.CellType.import_source()  # import the default ontology
bt.CellType.to_dataframe()   # your extensible cell type ontology in a simple registry

You can then create objects, e.g. for labeling, analogous to ULabel, Project, or Record:

t_cell = bt.CellType.get(name="T cell")
artifact.cell_types.add(t_cell)

Read more: docs.lamin.ai/manage-ontologies.

Watch a mini video: youtu.be/3vpWjHj3Kw8

Manage notes

When in your development directory, you can save markdown files as records:

lamin save <topic>/<my-note.md>