jammi-ai-platform 0.6.0

Python SDK for Jammi AI Platform — the commercial managed audit-native AI platform

jammi-ai-platform

Python SDK for the Jammi AI Platform — the commercial managed audit-native AI platform. Talks to a Jammi SaaS endpoint via gRPC with bearer-token authentication.

Install

pip install jammi-ai-platform

Wheel-only distribution. Python 3.9+ on Linux / macOS / Windows.

Quickstart

import jammi_ai_platform as jap

with jap.Connection(url="https://api.jammi.cloud", api_key="jt_…") as conn:
    print(conn.tenant_id)         # resolved via WhoAmI at connect time
    print(conn.token_label)       # label of the bearer in use

    registry = jap.Registry(conn)
    model_id = registry.register_model("recall-q4", task="embedding")
    version_id = registry.register_version(
        model_id,
        jammi_model_id="bge-small-en-v1.5",
        source=jap.ManualSource(metric_name="recall_at_k", metric_value=0.92),
    )

    jap.Gate(conn).check(
        version_id,
        jap.GateConfig(
            gate_name="ship-gate",
            source_id="patents",
            golden_set="golden-1",
            k=10,
            rule=jap.ThresholdRule(
                metric_name=jap.MetricName.RECALL_AT_K,
                threshold=0.85,
                op=jap.ThresholdOp.GE,
            ),
        ),
    )

Use http://… / grpc://… instead of https://… / grpcs://… for local insecure development.

Configuration is typed

Every control-plane config is a typed builder, not an untyped dict: enums for the *_type/direction/op fields (so an invalid value is a construction error), and tagged-union variants dispatched by type — MaxIterationsBudget / WallClockBudget / NoImprovementBudget for an ExperimentConfig budget, EmbeddingTarget / InferenceTarget for a MonitorConfig target, ThresholdRule / RangeRule / WelchTTestRule / MannWhitneyURule / ParetoDominatesRule / CompoundRule for a gate Rule, and ExperimentIterationSource / ManualSource for a registry version source. The open JSON/Struct payloads — an experiment's search_space, a validation's perturbation_params, an evidence config_json — stay plain dicts.

Data plane (composed from jammi-client)

connect, RemoteDatabase, and BearerCredentials are re-exported from jammi-client: the embed / search / sql data plane comes from there, the control plane lives here, and both ride jammi-client's bearer transport.

The data plane and the control plane open separate sessions. A jap.Connection is bearer-scoped — it resolves a tenant via WhoAmI and vends the control-plane handles. jap.connect(target, credentials=...) opens its own RemoteDatabase session (jammi-session-id scoped) for the data plane. They are two independent connections, not one shared channel; open each explicitly.

Surface

• Connection — control-plane gRPC channel + bearer auth, calls WhoAmI at connect
• connect / RemoteDatabase — re-exported jammi-client data plane (embed / search / sql)
• Registry — model + version + evidence catalog
• Gate — quality-gate checks
• Monitor — embedding / inference / divergence monitors + run streaming
• Experiment — Bayesian / Grid / Random experiments
• DeploymentBlocks — block / unblock sources from production traffic
• Resilience — perturbation-based validation runs
• Auth — issue / revoke / list API keys; WhoAmI
• Tenants — read-only tenant directory lookup

Errors

status_to_exception (in jammi_ai_platform.errors) maps every gRPC status code to a typed Python exception. The hierarchy:

• EnterpriseException
  • NotFoundError
  • AlreadyExistsError
  • TransitionNotAllowedError
  • TenantUnboundError
  • ConfigError
  • PermissionDeniedError

Docs

Full docs at https://docs.jammi.cloud.