Text decision tree
Choose model serving, AI Search and agent components while keeping identity, retrieval and runtime boundaries clear.
Adapted from Chapter 123 of Decision Trees: Frameworks for Technology Choices. Sources reviewed 2026-11-05.
START
Q0 Application needs a model inference endpoint?
YES -> Model Serving or supported model API. Match the model and deployment requirements. Then Q1.
NO -> Otherwise do not add a model endpoint simply because it is available. Then Q1.
Q1 Need semantic or hybrid retrieval over data?
YES -> Databricks AI Search candidate. Formerly Vector Search; validate indexes and filtering. Then Q2.
NO -> Otherwise do not add a vector index. Then Q2.
Q2 Need tools, state or multi-step agent behavior?
YES -> Agent framework and governed tools. Add tracing and evaluation for the actual workflow. Then Q3.
NO -> Otherwise keep a simpler application workflow. Then Q3.
Q3 A required capability fails the chosen component?
YES -> Compare a dedicated external service. Check model, hardware, protocol and latency needs. Then Q4.
NO -> Otherwise retain the chosen component without another service boundary. Then Q4.
Q4 An external agent runtime already owns the workflow?
YES -> Connect through governed interfaces. Use supported APIs, tools or MCP integrations. Then Q5.
NO -> Otherwise evaluate the applicable Databricks-hosted deployment option. Then Q5.
Q5 Identity, retrieval, evaluation and cost checks pass?
YES -> Pilot the composed application. Keep clear ownership and regression tests. Then END.
NO -> Otherwise revise the failed component or boundary before deployment. Then STOP.
END: A complete application can combine Databricks components with dedicated external services. Keep identity, evaluation and interface ownership explicit.
AI-generated chapter illustration
BookWriter-generated chapter 123 illustration
This AI-generated illustration gives a visual overview. It can simplify labels or imply connections that do not preserve every decision. For exact branches, prerequisites and stop conditions, follow the video graph and the text tree above.
Reference links
- Predictive AI Functions — 2026-10-14.
- Genie One MCP general availability — 2026-09-22.
- AI Search table-source support — 2026-10-26.