⛟ Export → Parquet

CLI tool to export a Qdrant collection to a Parquet file.

• All vector types: unnamed, named, multi-vectors and sparse.
• Parallel: ranges of points are exported by worker processes.
• Resumable: an interrupted export continues where it stopped.

Usage

Requires uv.

uvx qdrant-to-parquet my_collection out.parquet            # localhost

export QDRANT_URL=https://xyz.cloud.qdrant.io QDRANT_API_KEY=...
uvx qdrant-to-parquet my_collection out.parquet            # Qdrant Cloud

Option	Default
--url	$QDRANT_URL or http://localhost:6333	Qdrant URL
--api-key	$QDRANT_API_KEY or none	API key
--grpc-port	6334	gRPC port, on the host of --url
--rest	off (gRPC)	Use REST instead of gRPC
--workers	up to 4, by CPUs and collection size	Worker processes
--batch-size	256	Points per request
--timeout	60	Request timeout in seconds
--retries	3	Retries per failed request
--restart	off (resume)	Discard an interrupted export instead of resuming it
-f, --force	off	Replace the output file if it already exists.
-q, --quiet	off	No progress bar, warnings or summary

Exit codes: 0 done, 1 failed, 2 invalid arguments, 130 interrupted, 143 terminated.

Output

Column	Type
id	string: an integer ("42") or a UUID
vector (the unnamed vector) or vector_<name>	list<float32>
vector_<name> (multi-vector)	list<list<float32>>
vector_<name> (sparse)	struct<indices: list<uint32>, values: list<float32>>
payload	string: JSON, keys sorted

Rows are in point id order. A missing vector is null. The file metadata holds the collection's configuration.

-- DuckDB
SELECT id, payload->>'title' AS title, len(vector) AS dim FROM 'out.parquet';

Library

from qdrant_to_parquet.exporter import ClientConfig, export_collection

if __name__ == "__main__":
    config = ClientConfig(url="http://localhost:6333")
    export_collection(
        config.make(), "my_collection", "out.parquet", workers=4, client_config=config
    )

Development

uv run pytest                                      # without a server
QDRANT_URL=http://localhost:6333 uv run pytest     # also against a server, over gRPC and REST