Hyperlake Telemetry SDK for Python

Version 0.2.0 adds shared operations snapshots and redirect refusal. See SDK integration and release status. The 0.1.0 registry package does not contain these additions.

hyperlake-telemetry sends standards-compatible OTLP/HTTP JSON, attaches events and immutable artifacts, manages tenant-editable settings, and performs scoped event, artifact, video, Run, and evidence-graph investigations. It does not replace the OpenTelemetry SDK: applications with existing instrumentation should continue using their normal exporter and use this package only for run context, stable event identity, evaluations, and artifact attachment.

pip install hyperlake-telemetry

from hyperlake_telemetry import Client, Privacy

client = Client(
    base_url="https://us.hyperlake.cloud",
    protocol_token="write-only-protocol-token",
    tenant_jwt="short-lived-jwks-jwt",
    privacy=Privacy(capture_content=False),
)

with client.run("invoice-agent", goal_id="goal-42") as run:
    with run.span("lookup-account", kind="RETRIEVER"):
        pass
    run.evaluate("groundedness", score=0.97, label="pass")

client.flush()

result = client.semantic_query(
    "Show critical robot video and related faults from the last 2 hours"
)

detection = client.analyze_live_frame(
    "./frame.jpg",
    stream_session_id="camera-session-17",
    frame_id="frame-42",
    device_id="robot-17",
)

analyze_live_frame() uses the immutable model profile pinned to the JWT's pipeline. It requires a short-lived JWT with live:analyze and event:ingest. The response is a factual object-detection result; the attached live Rule Packs determine whether it becomes an alert or notification.

The JWT determines tenant and pipeline routing. Explicit routing IDs are optional with a JWT and are rejected when they conflict with its claims. Local claim decoding is not signature verification; the regional service verifies signature, issuer, audience, expiry, permissions, tenant, and pipeline for every request. Prompt, response, tool argument, and tool result content is suppressed by default. Set capture_content=True only after applying an approved data policy.

See the repository integration guides for Temporal, Langflow, OpenInference, Collector, and Alloy examples.

Reliable semantic queries

semantic_query() creates and validates a versioned, bounded query plan before calling any endpoint. Plans may target event retrieval, artifact/vector search, and evidence-graph traversal, but cannot contain tenant, pipeline, URL, token, SQL, or authorization fields. Every response includes the plan hash and scope provenance.

DeterministicSemanticPlanner works without an LLM. ModelSemanticPlanner accepts any callable that returns a JSON-like object; unsafe or invalid model output falls back to the deterministic plan. A model therefore interprets user intent but cannot expand authorization or execute arbitrary queries.

Artifact and video retrieval uses configured vector embeddings. Event retrieval is bounded structured/lexical retrieval, while graph results come from recorded scoped evidence edges. These signals remain separate and similarity is not treated as causation.

The client also exposes event listing/search, artifact search/upload/download, typed telemetry, video playback/timeline, Runs, evidence graphs, tenant config, protocol-token rotation, and Rule Pack validate/dry-run/publish helpers.