pixano-inference-client

The lightweight client for a Pixano Inference server. Install this when your app only needs to call a running server over HTTP — it pulls in just httpx, pydantic, and numpy, not the server stack (ray[serve], fastapi, uvicorn, torch, …).

pip install pixano-inference-client

It ships the same client classes and wire schemas as the full pixano-inference package, so code written against either works unchanged.

Use

from pixano_inference_client import SyncPixanoInferenceClient, DetectionRequest, EmbeddingRequest

client = SyncPixanoInferenceClient("http://localhost:7463", api_key="…")

det = client.detection(DetectionRequest(model="my-detector", image="https://example.com/cat.jpg"))
print(det.data.boxes, det.data.scores, det.data.classes)

emb = client.embedding(EmbeddingRequest(model="clip", text="a photo of a cat"))
print(emb.data.dim, emb.data.embeddings.to_numpy().shape)

An asynchronous twin, PixanoInferenceClient, exposes the same methods with await. Both accept an optional API key, retry transient failures with backoff, and raise PixanoInferenceError carrying the server's {code, message, requestId} envelope.

What's included

• Clients: PixanoInferenceClient (async), SyncPixanoInferenceClient (sync), PixanoInferenceError.
• Request/response types for every /v1 capability: SegmentationRequest/Response, DetectionRequest/Response, VLMRequest/Response, NERRequest/Response, EmbeddingRequest/Response, TrackingRequestV1/TrackingResponse.
• Wire types: NDArray/NDArrayFloat (compact base64 arrays), CompressedRLE (masks), ModelInfo, ModelStatusInfo, JobStatus, DeployModelRequest, and the capability I/O models.

Optional: mask codecs

Decoding a segmentation response into a numpy mask (CompressedRLE.to_mask()), or building one from a mask (CompressedRLE.from_mask()), needs pycocotools (and Pillow for image inputs). These are optional — parsing a normal compressed-bytes response works without them:

pip install pixano-inference-client[masks]

Relationship to pixano-inference

The full server package pixano-inference depends on this package and re-exports it, so from pixano_inference.client import PixanoInferenceClient and from pixano_inference_client import PixanoInferenceClient refer to the same classes. Server operators install pixano-inference; third-party callers install just pixano-inference-client.