Text decision tree
Distinguish knowledge gaps from model behavior problems before choosing retrieval, fine-tuning or both.
Adapted from Chapter 92 of Decision Trees: Frameworks for Technology Choices. Sources reviewed 2026-11-05.
START
Q0 Is missing or stale knowledge the failure?
YES -> Improve grounding and retrieval. Use relevant, current and permitted context. Then Q1.
NO -> Otherwise investigate behavior, format, tool use or task performance. Then Q1.
Q1 Retrieval bottleneck and NVIDIA deployment fit?
YES -> Pilot NeMo Retriever. Measure extraction, embeddings and reranking. Then Q2.
NO -> Otherwise improve the existing retrieval pipeline before replacing it. Then Q2.
Q2 Evaluation demonstrates a required weight change?
YES -> Evaluate customization. Use suitable data and an explicit target metric. Then Q3.
NO -> Otherwise keep the model weights fixed and improve context or prompting. Then Q3.
Q3 Required training needs distributed scale?
YES -> NeMo Framework components. Select the training components the task needs. Then Q4.
NO -> Otherwise use a lighter trainer for a necessary small job, or no trainer when weights stay fixed. Then Q4.
Q4 Nemotron passes task, license and runtime checks?
YES -> Include Nemotron as a candidate. Validate the specific model and hardware. Then Q5.
NO -> Otherwise choose another model that satisfies those checks. Then Q5.
Q5 Quality, latency, privacy and cost tests pass?
YES -> Deploy the measured configuration. Keep evaluation and regression tests. Then END.
NO -> Otherwise return to the diagnosed failure and revise the configuration. Then STOP.
END: Deploy the configuration that passes the held-out tests. Retrieval components, a training framework and a Nemotron model can be combined, but none is automatic.
AI-generated chapter illustration
BookWriter-generated chapter 92 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
- Fine-tuning Nemotron for local rules — 2026-10-27.
- NeMo failure analysis and evaluation — 2026-11-04.
- Databricks AI Runtime and NVIDIA customization — 2026-10-21.