LanceDB

The Lance file

A Lance file is a directory of fragments plus metadata. Each fragment has N rows; inside a fragment, columns are stored as Arrow IPC chunks. A separate row index maps primary keys to fragment + offset. The result is a file that supports fast column scans (for vector search) and fast row lookups (for primary-key reads).

Fragments can be added in O(1) and the file stays a single object for the consumer. This makes Lance popular for lakehouse-style data over S3 or local disk.

ANN search

LanceDB ships IVF-PQ (good for very large datasets) and HNSW (better recall). Both are built into the Rust core; queries are dispatched there directly. The query plan is simple: load the index, traverse it, fetch the top-k rows from the column store by their row-id, return.

              import lancedb

db = lancedb.connect("./.lancedb")
tbl = db.create_table("docs", data=rows, mode="overwrite")
tbl.create_index(num_partitions=256, num_sub_vectors=96)  # IVF-PQ

results = (
  tbl.search([0.1, 0.2, 0.3, ...])
    .metric("cosine")
    .limit(10)
    .to_pandas()
)
            

Why serverless wins for many workloads

For RAG over a few million embeddings, the cost of running a separate Qdrant or Milvus cluster is hard to justify. LanceDB runs inside the same Python process as the rest of your pipeline: no socket, no auth, no separate ops surface. You pay for the Rust binary and the file. When you outgrow that, you turn on server mode without changing the data layout.