turbopuffer

turbopuffer Python Client

The official Python client for accessing the turbopuffer API.

Usage

1. Install the turbopuffer package and set your API key.

$ pip install turbopuffer

Or if you're able to run C binaries for JSON encoding, use:

$ pip install turbopuffer[fast]

1. Start using the API

import turbopuffer as tpuf
tpuf.api_key = 'your-token'  # Alternatively: export=TURBOPUFFER_API_KEY=your-token
# Choose the best region for your data https://turbopuffer.com/docs/regions
tpuf.api_base_url = "https://gcp-us-east4.turbopuffer.com"

# Open a namespace
ns = tpuf.Namespace('hello_world')

# Read namespace metadata
if ns.exists():
    print(f'Namespace {ns.name} exists with {ns.dimensions()} dimensions and approximately {ns.approx_count()} vectors.')

# Upsert your dataset
ns.upsert(
    ids=[1, 2],
    vectors=[[0.1, 0.2], [0.3, 0.4]],
    attributes={'name': ['foo', 'foos']},
    distance_metric='cosine_distance',
)

# Alternatively, upsert using a row iterator
ns.upsert(
    {
        'id': id,
        'vector': [id/10, id/10],
        'attributes': {'name': 'food', 'num': 8}
    } for id in range(3, 10),
    distance_metric='cosine_distance',
)

# Query your dataset
vectors = ns.query(
    vector=[0.15, 0.22],
    distance_metric='cosine_distance',
    top_k=10,
    filters=['And', [
        ['name', 'Glob', 'foo*'],
        ['name', 'NotEq', 'food'],
    ]],
    include_attributes=['name'],
    include_vectors=True
)
print(vectors)
# [
#   VectorRow(id=2, vector=[0.30000001192092896, 0.4000000059604645], attributes={'name': 'foos'}, dist=0.001016080379486084),
#   VectorRow(id=1, vector=[0.10000000149011612, 0.20000000298023224], attributes={'name': 'foo'}, dist=0.009067952632904053)
# ]

# List all namespaces
namespaces = tpuf.namespaces()
print('Total namespaces:', len(namespaces))
for namespace in namespaces:
    print('Namespace', namespace.name, 'contains approximately', namespace.approx_count(),
            'vectors with', namespace.dimensions(), 'dimensions.')

# Delete vectors using the separate delete method
ns.delete([1, 2])

Endpoint Documentation

For more details on request parameters and query options, check the docs at https://turbopuffer.com/docs

Development

1. poetry run pytest
2. Bump version in turbopuffer/version.py and pyproject.toml
3. git tag vX.Y.Z && git push --tags
4. poetry build
5. poetry publish