Qdrant

Qdrant stores vectors and payload metadata for similarity search. Vane writes query results through distributed point upserts, mapping columns to point IDs, vectors, and payload fields.

Example

This example writes two document embeddings to documents. Install the vane-ai[qdrant] extra and create a collection with one unnamed, 3-dimensional dense vector using cosine distance. The query supplies unsigned 64-bit IDs and float32 vectors.

example.py

import vane
from vane import QdrantSink


relation = vane.sql("""
    SELECT id::UBIGINT AS id, embedding::FLOAT[] AS embedding, title
    FROM (VALUES
        (1, [0.1, 0.2, 0.3], 'Vector search'),
        (2, [0.4, 0.5, 0.6], 'Data pipelines')
    ) AS documents(id, embedding, title)
""")


sink = QdrantSink(
    "documents",
    url="http://localhost:6333",
    point_id="id",
    vector_mapping="embedding",
    payload_mapping={"title": "title"},
    worker_count=2,
)
summary = relation.write_datasink(sink)

Use an endpoint reachable by all workers. Reusing an ID replaces the point's full vector and payload; successful batches are not rolled back if a later batch fails.