Text decision tree
Adapted from Appendix AC of Decision Trees: Frameworks for Technology Choices. Sources reviewed 2026-11-07.
START: Evaluate the smallest change that addresses the measured failure.
Q1. Prompting meets the evaluation target?
Yes -> Keep the base model and prompt. Serve and monitor the evaluated design. [STOP]
No -> Otherwise identify what information or behavior is missing. [Q2]
Q2. Need current or private knowledge?
Yes -> Add governed retrieval. Examples: Databricks AI Search or Foundry. [Q3]
No -> If knowledge access is not the failure, continue to the behavior question. [Q3]
Q3. Behavior still fails after evaluation?
Yes -> Evaluate supervised tuning. Example: NVIDIA NeMo fine-tuning. [Q4]
No -> Otherwise serve the evaluated prompt or retrieval design. [Q4]
Q4. Must a new base model be trained?
Yes -> Consider pretraining as a separate case. Require data, compute and expertise. [END]
No -> Otherwise return to evaluation and improve the selected design. [END]
END: Retrieval and tuning can coexist. Re-evaluate quality, latency, cost and access controls after each change.
AI-generated chapter illustration
Conceptual illustration for Choose an AI architecture pattern
The illustration arranges four interventions in a line, and its last panel also mentions changing the selected base model. The graph has a stop condition after successful prompting and treats training a new base model as a separate decision. Selecting another existing model does not by itself require pretraining.
The illustration was generated in BookWriter. Its visual arrangement is interpretive; use the explicit nodes and routes in the graph and text tree for the decision.