LanceDB

Why LanceDB

LanceDB is an embedded vector database (like SQLite for vectors):

Single-file columnar Lance format on disk or S3/GCS.
Arrow native — zero-copy read into pandas, polars, DuckDB.
Versioned writes with time travel (checkout any past snapshot).
No server process; your Python / TypeScript process opens the DB directly.
Rust core, bindings for Python and JS/TS.

Pick LanceDB when:

You want a vector store without standing up another service.
Your corpus sits in S3 / GCS and you want to query it in place.
Data analytics (pandas, polars, DuckDB) is part of your retrieval pipeline.

Skip it for:

Multi-writer concurrency (it handles one writer at a time cleanly).
Very high QPS serving — embed into a server or use LanceDB Cloud.

Local vs Cloud vs S3

# pip install lancedb
import lancedb

# Local directory
db = lancedb.connect("./.lancedb")

# S3-backed (no server)
db = lancedb.connect("s3://my-bucket/lancedb",
                     storage_options={"region": "us-east-1"})

# LanceDB Cloud (managed)
db = lancedb.connect("db://my-project", api_key=os.environ["LANCEDB_API_KEY"])

S3 backing is a killer feature: many readers, one writer, all reading the same immutable Lance files. No replication to configure.

Creating a Table

import pyarrow as pa
import numpy as np

schema = pa.schema([
    pa.field("id", pa.string()),
    pa.field("vector", pa.list_(pa.float32(), 1024)),
    pa.field("text", pa.string()),
    pa.field("tenant_id", pa.string()),
    pa.field("created_at", pa.timestamp("us")),
])

table = db.create_table("docs", schema=schema, mode="overwrite")

Or infer from data:

data = [
    {"id": "d1", "vector": np.random.rand(1024).astype("float32"),
     "text": "OAuth uses refresh tokens.", "tenant_id": "acme"},
]
table = db.create_table("docs", data=data)

Embedding Functions (Auto-Embed on Insert)

from lancedb.pydantic import LanceModel, Vector
from lancedb.embeddings import get_registry

registry = get_registry()
embedder = registry.get("openai").create(name="text-embedding-3-small")

class Doc(LanceModel):
    id: str
    text: str = embedder.SourceField()
    vector: Vector(embedder.ndims()) = embedder.VectorField()
    tenant_id: str

table = db.create_table("docs", schema=Doc, mode="overwrite")
table.add([
    {"id": "d1", "text": "OAuth uses refresh tokens.", "tenant_id": "acme"},
    {"id": "d2", "text": "PKCE protects public clients.", "tenant_id": "acme"},
])

# Search by text
results = table.search("how to refresh a token").limit(5).to_pandas()

Registry includes OpenAI, Cohere, Voyage, HuggingFace Sentence Transformers, Ollama, and custom subclasses.

Indexing (IVF_PQ + HNSW)

Default brute-force search is fine up to ~50k vectors. Beyond that, build an ANN index:

# IVF_PQ — good for million+ scale with memory savings
table.create_index(
    metric="cosine",
    num_partitions=256,       # rule: sqrt(num_rows)
    num_sub_vectors=96,       # must divide dim; 1024/96 rounds, try 64 or 128
    index_type="IVF_PQ",
)

# HNSW — higher recall, more memory
table.create_index(
    metric="cosine",
    index_type="IVF_HNSW_SQ",  # IVF with HNSW inside each partition + scalar quantization
    num_partitions=256,
)

IVF_HNSW_SQ is LanceDB's current sweet spot for accuracy + memory.

Scalar indexes

table.create_scalar_index("tenant_id")     # bitmap; fast equality / IN filters
table.create_scalar_index("created_at")    # btree; range queries

Full-Text Search (FTS) + Vector Hybrid

table.create_fts_index("text", use_tantivy=True)

# Hybrid search
from lancedb.rerankers import RRFReranker

results = (
    table.search(query_type="hybrid")
    .vector(q_vec)
    .text("oauth refresh token")
    .rerank(reranker=RRFReranker())
    .where("tenant_id = 'acme'")
    .limit(10)
    .to_pandas()
)

use_tantivy=True enables the Rust Tantivy engine (BM25 + stemming + Unicode tokenization). FTS and vector query run in parallel, then merge via the reranker.

Filters with SQL WHERE

LanceDB accepts DataFusion SQL in .where():

table.search(q_vec).where(
    "tenant_id = 'acme' AND created_at > TIMESTAMP '2025-01-01' AND archived = false",
    prefilter=True,
).limit(10).to_pandas()

prefilter=True applies the filter before ANN search (fewer candidates, possibly lower recall if the filter is very selective); prefilter=False (default) filters after.

Versioning and Time Travel

Every write creates a new version of the dataset. Checkout any past version:

# Inspect versions
table.list_versions()
# [{'version': 1, 'timestamp': ..., 'metadata': {...}}, ...]

# Time-travel read
old = table.checkout(version=3).to_pandas()

# Restore to a past version
table.restore(version=3)

Versions are cheap (copy-on-write). Use them to:

Reproduce evals against a fixed corpus snapshot.
Roll back accidental bulk deletes.
Blue-green ingest: write new version, swap pointer.

Merge-on-Read Upserts

table.merge_insert("id").when_matched_update_all().when_not_matched_insert_all().execute([
    {"id": "d1", "text": "Updated text.", "tenant_id": "acme"},
    {"id": "d3", "text": "New doc.", "tenant_id": "acme"},
])

This is a proper UPSERT: match on id, update matching rows, insert the rest. Background compaction eventually rewrites files to remove tombstones.

Compaction and Optimization

Frequent small writes leave many tiny fragments. Run compaction:

table.optimize(cleanup_older_than=timedelta(days=7))

It rewrites fragments into larger files and purges unreachable versions. Schedule nightly in production.

Polars / DuckDB Integration

LanceDB tables are Arrow; polars and DuckDB read them zero-copy.

import polars as pl

df = table.to_polars()
df.filter(pl.col("tenant_id") == "acme").select(["id", "text"])

# DuckDB
import duckdb
duckdb.sql("SELECT id, text FROM lance_scan('./.lancedb/docs.lance') WHERE tenant_id = 'acme'")

Useful for offline eval, training set construction, bulk re-embedding.

Multi-Tenancy

Pick one of:

Single table + tenant_id column + scalar index + filter per query.
Table per tenant (db.create_table(f"docs_{tenant_id}", ...)).

Pattern 1 scales for thousands of tenants. Pattern 2 is cleaner when tenants have radically different schemas or data volumes.

Anti-Patterns

Anti-Pattern	Fix
Brute-force search on 1M+ vectors	Build IVF_PQ or IVF_HNSW_SQ index
Many concurrent writers	Serialize writes; LanceDB is single-writer per table
Forgetting to optimize	Nightly `table.optimize()` with retention
Storing raw PDFs in a row	Store text only; PDFs go to object storage
No scalar index on filter columns	Add `create_scalar_index` for filtered fields
Re-running embedding on every read	Register embedder; vectors are stored once
Ignoring versioning	Use `checkout` for reproducible evals
Full-text search without `use_tantivy=True`	Tantivy is substantially better than the legacy tokenizer

lancedb