Lilac was acquired by Databricks in March 2024 and the open-source repository was officially archived in July 2025. While the code remains available under an Apache 2.0 license, it is effectively frozen in time. If you are on the Databricks platform, you should look for these features inside Mosaic AI. If you are a standalone developer, you are looking at a powerful, free, but unmaintained "zombie" tool.
For local-first data curation, Lilac was—and technically still is—excellent. It spins up a local web server that visualizes your text datasets, automatically clustering them to find semantic duplicates, PII, and garbage patterns. Processing a 100k-row dataset locally costs nothing but your own compute time. The interface is intuitive: you filter by "toxic" or "near-duplicate" scores, select the bad rows, and export the clean JSONL. It feels like a spreadsheet possessed by a language model.
The friction comes from its abandoned state. As Python libraries evolve, Lilac’s frozen dependency tree will increasingly conflict with modern environments. You cannot expect security patches, new embedding model support, or bug fixes. It relies on local hardware, so if you don't have a decent GPU, the embedding generation step for large datasets will crawl.
Compare this to Nomic Atlas, which offers similar massive-scale visualization but as a managed service (or enterprise self-host), or Argilla, which is actively maintained and better suited for human-in-the-loop labeling.
Use Lilac today only if you need a quick, private, offline way to inspect a messy text file and you are comfortable pinning your Python environment to 2024 standards. For any production pipeline or long-term project, this tool is a dead end. You should migrate to active alternatives immediately.
Pricing
Lilac is free (Apache 2.0), but the hidden cost is maintenance debt. Since the repo is archived (read-only), there are no paid tiers or enterprise support options outside of buying the full Databricks platform. The "free" usage is limited by your local hardware; processing millions of rows requires significant local GPU VRAM or patience. Competitors like Nomic Atlas allow free visualization up to 1M points but charge for enterprise privacy features.
Technical Verdict
The library is a self-contained Python package that launches a local Flask/React app. Installation is a simple pip install, but expect dependency hell with newer versions of Pydantic or LangChain due to the lack of updates since 2025. It works best in an isolated virtual environment. The UI is snappy for datasets under 1M rows, but performance degrades without a GPU for the embedding steps.
Quick Start
# pip install lilac[all]
import lilac as ll
# Point to your local data
ll.set_project_dir('./my_lilac_data')
# Load a dataset (e.g., from a JSONL file)
ll.ingest('local_dataset', 'my_data.jsonl')
# Start the local web UI
ll.start_server()Watch Out
- Repository is archived (July 2025); no new updates or fixes will be released.
- Dependency conflicts are high; likely requires downgrading other libraries to 2024 versions.
- No native support for image or audio data; text only.
- Embedding generation is slow on CPU-only machines.