GoVec Python SDK

The Python client for GoVec, a compact vector search engine with HNSW, int8 quantization and hybrid search. Talk to it over REST or gRPC through the same API.

Features

• The whole server API: insert, batch insert, search with metadata filters and hybrid dense + sparse scoring, get, delete, stats, info, flush and reset.
• REST or gRPC, one argument apart. Both return the same results and raise the same exceptions; the test suite runs every operation against a real server over both.
• Typed. Responses are dataclasses, and the package ships py.typed.
• Batch inserts that scale: large lists are chunked for you, with per-vector errors reported instead of failing the whole batch.
• Safe against newer servers. Fields a newer GoVec adds are ignored, not fatal.

Installation

pip install govec
# or
uv add govec

Requires Python 3.12 or newer. You also need a GoVec server; the quickest way is Docker:

docker run -p 9697:9697 -v govec-data:/data ghcr.io/pradyothsp/govec:latest

Quick start

from govec import GoVecClient

with GoVecClient(host="localhost", port=9697, api_key="", protocol="rest", tls=False) as client:
    client.insert("doc-1", [0.12, 0.91, 0.20], metadata={"title": "Getting started"})
    client.insert("doc-2", [0.80, 0.10, 0.31], metadata={"title": "Release notes"})

    for hit in client.search([0.10, 0.88, 0.18], k=2):
        print(hit.id, round(hit.score, 3), hit.meta)

doc-1 1.0 {'title': 'Getting started'}
doc-2 0.287 {'title': 'Release notes'}

In a real application the vectors come from an embedding model (OpenAI, Sentence Transformers, ...); GoVec stores and searches them. Every vector in an index must have the same number of dimensions.

Connecting

GoVecClient(host, port, api_key, protocol, tls=True)

The constructor contacts the server straight away, so a wrong address fails at GoVecClient(...) rather than on the first call. Use the client as a context manager, or call client.close() when you're done.

To use gRPC, change two arguments; nothing else in your code changes:

client = GoVecClient(host="localhost", port=9698, api_key="", protocol="grpc", tls=False)

Guide

Inserting vectors

client.insert("doc-1", [0.12, 0.91, 0.20], metadata={"lang": "en", "year": 2024})

Inserting an existing ID replaces it. For many vectors, build InsertRequests and use insert_many, which sends them in chunks of batch_size:

from govec import InsertRequest

result = client.insert_many(
    [InsertRequest(id=f"doc-{i}", vector=vec) for i, vec in enumerate(vectors)],
    batch_size=500,
)
print(result.inserted_count)
for failure in result.errors:
    print(failure.id, failure.error)

Searching

hits = client.search([0.10, 0.88, 0.18], k=5)

Each hit has id, score (higher is closer) and meta (None if the vector has no metadata).

Filter by metadata. Every key must match exactly:

client.search([0.10, 0.88, 0.18], k=5, filter={"lang": "en"})

Hybrid search. Add a sparse vector (for example from BM25 or SPLADE) to both the insert and the query; the server blends the dense and sparse scores:

from govec import SparseVector

client.insert("doc-3", [0.3, 0.3, 0.3], sparse_vector=SparseVector(indices=[7, 42], values=[1.0, 0.5]))
client.search([0.3, 0.3, 0.3], sparse_vector=SparseVector(indices=[42], values=[1.0]), k=3)

Reading and deleting

record = client.get_by_id("doc-1")   # vector, sparse_vector and metadata, or None if missing
client.delete("doc-1")               # raises GoVecAPIError if the ID doesn't exist

Administration

client.info()        # server version, index type, quantization, metric, dimensions
client.get_stats()   # number of stored vectors
client.health()      # liveness
client.flush()       # write a snapshot to disk now
client.reset()       # delete every vector -- irreversible

API reference

Error handling

Both transports raise the same exceptions, all subclasses of GoVecError:

from govec import GoVecAPIError, GoVecError

try:
    client.insert("doc-9", [0.1, 0.2])   # wrong width for a 3-dimensional index
except GoVecAPIError as e:
    print(e.status_code, e.message)
except GoVecError:
    ...                                  # connection problems, timeouts

Compatibility

client.info().version tells you which server release you're connected to.

Two transport differences to know, both from protobuf:

• Metadata numbers come back as floats over gRPC: {"year": 2024} returns as 2024.0. REST preserves integers.
• Vectors come back at 32-bit precision over gRPC: GoVec stores 32-bit floats, so get_by_id over gRPC returns 0.12 as 0.11999999731779099. REST rounds it back to 0.12.

Roadmap

• Async client (planned for 0.2.0): an AsyncGoVecClient with the same API over httpx.AsyncClient and grpc.aio, for FastAPI services and async RAG pipelines. Until then, call the sync client from async code with await asyncio.to_thread(client.search, ...).

Contributing

Issues and pull requests are welcome; see CONTRIBUTING.md.

License

Apache License 2.0