Databricks Model Serving

Deploy MLflow models and AI agents to scalable REST API endpoints.

Quick Decision: What Are You Deploying?

Prerequisites

• DBR 16.1+ recommended (pre-installed GenAI packages)
• Unity Catalog enabled workspace
• Model Serving enabled

Foundation Model API Endpoints

ALWAYS use exact endpoint names from this table. NEVER guess or abbreviate.

Chat / Instruct Models

Embedding Models

Common Defaults

• Agent LLM: databricks-meta-llama-3-3-70b-instruct (good balance of quality/cost)
• Embedding: databricks-gte-large-en
• Code tasks: databricks-gpt-5-1-codex-mini or databricks-gpt-5-1-codex-max

These are pay-per-token endpoints available in every workspace. For production, consider provisioned throughput mode. See supported models.

Reference Files

Quick Start: Deploy a GenAI Agent

Step 1: Install Packages (in notebook or via MCP)

%pip install -U mlflow==3.6.0 databricks-langchain langgraph==0.3.4 databricks-agents pydantic
dbutils.library.restartPython()

Or via MCP:

execute_code(code="%pip install -U mlflow==3.6.0 databricks-langchain langgraph==0.3.4 databricks-agents pydantic")

Step 2: Create Agent File

Create agent.py locally with ResponsesAgent pattern (see 3-genai-agents.md).

Step 3: Upload to Workspace

upload_to_workspace(
    local_path="./my_agent",
    workspace_path="/Workspace/Users/you@company.com/my_agent"
)

Step 4: Test Agent

execute_code(
    file_path="./my_agent/test_agent.py",
    cluster_id="<cluster_id>"
)

Step 5: Log Model

execute_code(
    file_path="./my_agent/log_model.py",
    cluster_id="<cluster_id>"
)

Step 6: Deploy (Async via Job)

See 7-deployment.md for job-based deployment that doesn't timeout.

Step 7: Query Endpoint

query_serving_endpoint(
    name="my-agent-endpoint",
    messages=[{"role": "user", "content": "Hello!"}]
)

Quick Start: Deploy a Classical ML Model

import mlflow
import mlflow.sklearn
from sklearn.linear_model import LogisticRegression

# Enable autolog with auto-registration
mlflow.sklearn.autolog(
    log_input_examples=True,
    registered_model_name="main.models.my_classifier"
)

# Train - model is logged and registered automatically
model = LogisticRegression()
model.fit(X_train, y_train)

Then deploy via UI or SDK. See 1-classical-ml.md.

MCP Tools

If MCP tools are not available, use the SDK/CLI examples in the reference files below.

Development & Testing

Deployment

Querying

Common Workflows

Check Endpoint Status After Deployment

get_serving_endpoint_status(name="my-agent-endpoint")

Returns:

{
    "name": "my-agent-endpoint",
    "state": "READY",
    "served_entities": [...]
}

Query a Chat/Agent Endpoint

query_serving_endpoint(
    name="my-agent-endpoint",
    messages=[
        {"role": "user", "content": "What is Databricks?"}
    ],
    max_tokens=500
)

Query a Traditional ML Endpoint

query_serving_endpoint(
    name="sklearn-classifier",
    dataframe_records=[
        {"age": 25, "income": 50000, "credit_score": 720}
    ]
)

Common Issues

Critical: ResponsesAgent Output Format

WRONG - raw dicts don't work:

return ResponsesAgentResponse(output=[{"role": "assistant", "content": "..."}])

CORRECT - use helper methods:

return ResponsesAgentResponse(
    output=[self.create_text_output_item(text="...", id="msg_1")]
)

Available helper methods:

• self.create_text_output_item(text, id) - text responses
• self.create_function_call_item(id, call_id, name, arguments) - tool calls
• self.create_function_call_output_item(call_id, output) - tool results

Related Skills

• databricks-agent-bricks - Pre-built agent tiles that deploy to model-serving endpoints
• databricks-vector-search - Create vector indexes used as retriever tools in agents
• databricks-genie - Genie Spaces can serve as agents in multi-agent setups
• databricks-mlflow-evaluation - Evaluate model and agent quality before deployment
• databricks-jobs - Job-based async deployment used for agent endpoints

Resources

• Model Serving Documentation
• MLflow 3 ResponsesAgent
• Agent Framework