LanceDB

Warnings

• breaking Illegal instruction (SIGILL) crash on import on older Intel CPUs (pre-AVX2). lancedb/pylance wheels are compiled with AVX2 SIMD instructions. Affects Ubuntu 20.04 on older hardware and some VMs where CPU features are masked.
  Fix: Requires AVX2-capable CPU. Check with: grep avx2 /proc/cpuinfo. No workaround via pip — must use newer hardware or build from source without AVX2.
• breaking Some lancedb releases have pinned a pre-release version of pylance as a hard dependency (e.g., lancedb==0.17.1 required pylance==0.21.0b5). This breaks pip/uv installs in strict environments that disallow pre-release packages.
  Fix: If a version fails to resolve, pin to the previous minor version or add --prerelease=allow to uv. Check GitHub releases for known bad versions.
• breaking Python >=3.10 required as of lancedb 0.25+. Earlier Python versions raise install or import errors.
  Fix: Use Python 3.10, 3.11, 3.12, or 3.13.
• gotcha PyPI status is 'Development Status :: 3 - Alpha' despite being widely used in production. The API has had breaking changes between minor versions. Pin lancedb to a specific version in production.
  Fix: Pin in requirements: lancedb==0.29.2. Review CHANGELOG before upgrading.
• gotcha ANN index (create_index) must be created explicitly. Without it, all searches are O(n) brute-force regardless of dataset size. No warning is emitted — queries silently degrade at scale.
  Fix: Call table.create_index(metric='cosine') after loading data. For large datasets, tune num_partitions and num_sub_vectors for IVF_PQ.
• gotcha Automatic versioning creates a new Lance snapshot on every write operation. On high-frequency write workloads this accumulates many small version files rapidly, increasing storage and compaction overhead.
  Fix: Periodically run table.compact_files() and table.cleanup_old_versions() to manage storage. This is not done automatically.
• gotcha pylance (the LanceDB dependency) is a completely different package from pylance (Microsoft's Python language server for VS Code). pip install pylance without context installs Microsoft's package. lancedb's pylance is only installed as a transitive dependency.
  Fix: Never manually pip install pylance expecting lancedb's version. It is pulled in automatically by lancedb.

Install

• pip install lancedb Core install
• pip install lancedb[embeddings] With built-in embedding functions (OpenAI, HF, etc.)
• pip install lancedb[azure] With Azure Blob Storage support
• pip install --pre --extra-index-url https://pypi.fury.io/lancedb/ lancedb Preview/nightly builds

Imports

• lancedb

  import lancedb

  Top-level import works. Use lancedb.connect() for sync or lancedb.connect_async() for async.
• LanceModel (schema)

  from lancedb.pydantic import LanceModel, Vector

  LanceDB schemas require LanceModel, not plain pydantic BaseModel. Vector field uses lancedb's Vector type.

Quickstart

No server needed. Data is stored on disk as Lance files. create_index() is required for ANN performance — without it, all searches are exact (brute-force). Versioning is automatic: every add/delete creates a new version.

import lancedb
import numpy as np
from lancedb.pydantic import LanceModel, Vector

# Connect (creates directory if not exists)
db = lancedb.connect("/tmp/my-lancedb")

# Define schema using LanceModel
class Item(LanceModel):
    text: str
    vector: Vector(128)  # fixed dimensions

# Create table
table = db.create_table("items", schema=Item, mode="overwrite")

# Add data
data = [
    Item(text="hello world", vector=np.random.rand(128).astype('float32'))
    for _ in range(100)
]
table.add(data)

# Vector search (returns pandas DataFrame by default)
query_vec = np.random.rand(128).astype('float32')
results = table.search(query_vec).limit(5).to_pandas()
print(results)

# Create ANN index (required for scale)
table.create_index(metric="cosine")  # IVF_PQ by default