Vector search for your data lake: Query embeddings where they live.
Your embeddings shouldn't leave the lake. LanceDB adds vector search directly on object storage with up to 1000x faster random access than Parquet. Works alongside ClickHouse, Trino, DuckDB.
Tomorrow's AI is being built on LanceDB today
ClickHouse for analytics. LanceDB for vector serving.
ClickHouse is an analytics powerhouse. But vector search is a serving workload, not an analytics workload. Use ClickHouse for analysis, LanceDB for real-time retrieval with native hybrid search.
| ClickHouse | LanceDB | |
|---|---|---|
| Optimzed for | OLAP analytics | Vector search and serving. |
| Latency | Analytics batch queries | Real-time millisecond queries. |
| Search | Vector via approximate functions | Native vector + full-text + SQL hybrid in one query. |
| Random access | Scan-based | O(1) row lookups. Up to 1000x faster than Parquet. |
| Best for | Vector analytics on historical data | Real-time vector retrieval with hybrid search. |
LanceDB: the best storage layer for Trino vector queries
Trino is a query engine, not a storage format. LanceDB's Lance format is optimized for the fast random access that vector search requires. Query Lance tables through Trino.
| Trino | LanceDB | |
|---|---|---|
| Role | Query engine | Storage format + query engine. |
| Storage | Queries Parquet, Iceberg, etc. | Lance format - optimized for vectors and random access. |
| Random access | Scan-based | O(1) row lookups. Up to 1000x faster than Parquet. |
| Search | SQL queries over columnar data | Native vector + full-text + SQL hybrid in one query. |
| Best for | Federated queries across formats | Vector-native storage for Trino and beyond. |
DuckDB for laptop analytics. LanceDB for production AI.
DuckDB is fantastic for local analytics. LanceDB has the same embedded simplicity, but scales to production — cloud storage, distributed queries, billions of vectors, native hybrid search.
| DuckDB | LanceDB | |
|---|---|---|
| Deployment | Local/embedded only | Embedded to cloud-scale. |
| Scale | Single-node | 20 PB largest table. 20K+ QPS. Billions of vectors. |
| Search | Basic similarity via extensions | Native vector + full-text + SQL hybrid in one query. |
| Random access | Scan-based on Parquet | O(1) row lookups. Up to 1000x faster than Parquet. |
| Best for | Local analytics and prototyping | Prototype locally, scale to production. |
Talk to Engineering
Or try LanceDB OSS — same code, scales to Cloud.