oci-data-science

name: oci-data-science description: OCI Data Science service patterns including Jobs, Pipelines, Model Catalog, authentication, and the ADS SDK beyond AQUA. Triggered when user asks about OCI Data Science Jobs, ML Pipelines, model artifacts, conda environments, or general ADS SDK usage. user-invocable: true disable-model-invocation: false

OCI Data Science with ADS SDK

Use this skill for general OCI Data Science service operations beyond AI Quick Actions (AQUA), including Jobs, Pipelines, Model Catalog, and authentication.

Authentication

import ads

# In OCI Notebook Sessions (recommended)
ads.set_auth("resource_principal")

# Local development with API key (~/.oci/config)
ads.set_auth("api_key")

# Local development with security token (session-based)
ads.set_auth("security_token")

# Get the current signer for raw OCI API calls
from ads.common.auth import default_signer
auth = default_signer()

OCI Client Factory

from ads.common.oci_client import OCIClientFactory

# Create service clients
ds_client = OCIClientFactory(**auth).data_science
os_client = OCIClientFactory(**auth).object_storage
compute_client = OCIClientFactory(**auth).compute
identity_client = OCIClientFactory(**auth).identity

Data Science Jobs

Jobs allow running Python scripts, notebooks, or containers on managed infrastructure.

Create and Run a Job

from ads.jobs import Job, DataScienceJob, ScriptRuntime

job = (
    Job(name="my-training-job")
    .with_infrastructure(
        DataScienceJob()
        .with_compartment_id("ocid1.compartment.oc1..xxx")
        .with_project_id("ocid1.datascienceproject.oc1.iad.xxx")
        .with_shape_name("VM.Standard.E4.Flex")
        .with_shape_config_details(memory_in_gbs=16, ocpus=2)
        .with_block_storage_size(50)
        .with_log_group_id("ocid1.loggroup.oc1.iad.xxx")
        .with_log_id("ocid1.log.oc1.iad.xxx")
    )
    .with_runtime(
        ScriptRuntime()
        .with_source("path/to/script.py")
        .with_service_conda("pytorch21_p39_gpu_v1")
    )
)

job.create()
job_run = job.run()
job_run.watch()  # Stream logs

Job with Container Runtime

from ads.jobs import Job, DataScienceJob, ContainerRuntime

job = (
    Job(name="my-container-job")
    .with_infrastructure(
        DataScienceJob()
        .with_compartment_id("ocid1.compartment.oc1..xxx")
        .with_project_id("ocid1.datascienceproject.oc1.iad.xxx")
        .with_shape_name("VM.GPU.A10.1")
    )
    .with_runtime(
        ContainerRuntime()
        .with_image("iad.ocir.io/my-namespace/my-image:latest")
        .with_environment_variable(MY_VAR="value")
    )
)

Job with Git Runtime

from ads.jobs import Job, DataScienceJob, GitPythonRuntime

job = (
    Job(name="my-git-job")
    .with_infrastructure(
        DataScienceJob()
        .with_compartment_id("ocid1.compartment.oc1..xxx")
        .with_project_id("ocid1.datascienceproject.oc1.iad.xxx")
        .with_shape_name("VM.Standard.E4.Flex")
    )
    .with_runtime(
        GitPythonRuntime()
        .with_source("https://github.com/user/repo.git")
        .with_entry_point("train.py")
        .with_service_conda("generalml_p311_cpu_v1")
    )
)

Watch Job Logs

ads opctl watch <job_run_ocid> --auth resource_principal

ML Pipelines

Pipelines chain multiple steps into reproducible ML workflows.

Create a Pipeline

from ads.pipeline import Pipeline, PipelineStep, ScriptStep

step1 = (
    ScriptStep("data-prep")
    .with_infrastructure(
        DataScienceJob()
        .with_shape_name("VM.Standard.E4.Flex")
        .with_block_storage_size(50)
    )
    .with_runtime(
        ScriptRuntime()
        .with_source("prep_data.py")
        .with_service_conda("generalml_p311_cpu_v1")
    )
)

step2 = (
    ScriptStep("train-model")
    .with_infrastructure(
        DataScienceJob()
        .with_shape_name("VM.GPU.A10.1")
    )
    .with_runtime(
        ScriptRuntime()
        .with_source("train.py")
        .with_service_conda("pytorch21_p39_gpu_v1")
    )
)

pipeline = (
    Pipeline("my-ml-pipeline")
    .with_compartment_id("ocid1.compartment.oc1..xxx")
    .with_project_id("ocid1.datascienceproject.oc1.iad.xxx")
    .with_step_details([step1, step2])
    .with_dag(["data-prep >> train-model"])
    .with_log_group_id("ocid1.loggroup.oc1.iad.xxx")
)

pipeline.create()
pipeline_run = pipeline.run()
pipeline_run.watch()

Model Catalog

Store, version, and manage ML models in a centralized catalog.

Save a Model

from ads.model import GenericModel

model = GenericModel(estimator=my_trained_model, artifact_dir="./model_artifact")
model.prepare(
    inference_conda_env="generalml_p311_cpu_v1",
    inference_python_version="3.11",
    model_file_name="model.pkl",
)
model.verify(test_data)
model.save(
    display_name="my-classification-model",
    compartment_id="ocid1.compartment.oc1..xxx",
    project_id="ocid1.datascienceproject.oc1.iad.xxx",
)

Load a Model

from ads.model import GenericModel

model = GenericModel.from_model_catalog(
    model_id="ocid1.datasciencemodel.oc1.iad.xxx",
    artifact_dir="./downloaded_artifact",
)
predictions = model.predict(test_data)

Model Deployment (non-AQUA)

from ads.model.deployment import ModelDeployment, ModelDeploymentContainerRuntime

deployment = (
    ModelDeployment()
    .with_display_name("my-model-deployment")
    .with_compartment_id("ocid1.compartment.oc1..xxx")
    .with_project_id("ocid1.datascienceproject.oc1.iad.xxx")
    .with_model_id("ocid1.datasciencemodel.oc1.iad.xxx")
    .with_shape_name("VM.Standard.E4.Flex")
    .with_shape_config_details(memory_in_gbs=16, ocpus=2)
    .with_bandwidth_mbps(10)
)
deployment.deploy()
deployment.predict(test_data)

LangChain Integration

Chat with OCI Model Deployment

from ads.llm import ChatOCIModelDeployment

llm = ChatOCIModelDeployment(
    model="odsc-llm",
    endpoint="https://modeldeployment.us-ashburn-1.oci.customer-oci.com/<deployment_ocid>/predict",
    streaming=True,
)

response = llm.invoke("Tell me about Oracle Cloud.")
print(response.content)

With Tool Calling (LangChain)

from langchain_core.tools import tool
from langchain.agents import create_tool_calling_agent, AgentExecutor

@tool
def get_weather(city: str) -> str:
    """Get weather for a city."""
    return f"Weather in {city}: Sunny, 72F"

agent = create_tool_calling_agent(llm, [get_weather], prompt_template)
executor = AgentExecutor(agent=agent, tools=[get_weather])
result = executor.invoke({"input": "What's the weather in Austin?"})

Object Storage Helpers

from ads.common.object_storage_details import ObjectStorageDetails

# Parse an OCI Object Storage path
oss = ObjectStorageDetails.from_path("oci://my-bucket@my-namespace/path/to/file.csv")
print(oss.bucket)     # "my-bucket"
print(oss.namespace)  # "my-namespace"
print(oss.filepath)   # "path/to/file.csv"

Conda Environments

OCI Data Science provides pre-built conda environments:

List available:

from ads.common.conda_environment import CondaEnvironment
envs = CondaEnvironment.list()

Project Structure Reference

ads/                    # Main SDK source
├── aqua/              # AI Quick Actions (AQUA) - GenAI/LLM lifecycle
├── common/            # Shared: auth, OCI clients, serializers, utils
├── model/             # Model framework, artifacts, deployment
├── jobs/              # OCI Data Science Jobs
├── pipeline/          # ML Pipelines
├── opctl/             # CLI operators, log watching
├── llm/               # LangChain integration
├── feature_store/     # Feature Store
└── catalog/           # Model catalog helpers

Key Documentation

• ADS User Guide
• ADS API Reference
• OCI Data Science Docs
• AQUA Docs
• Samples Repo