AIF-C01 FM Selection, Inference Parameters, RAG and Agents Guide

This lesson covers the main application patterns AWS wants candidates to classify correctly. The exam often asks whether the stronger answer is a different model, a different inference configuration, a retrieval step, or an agentic workflow.

RAG: Retrieval-augmented generation, where the system retrieves relevant context before the model generates an answer.

Agent: Model-driven workflow component that can reason about steps and call tools or actions.

Inference parameters: Runtime settings such as temperature that shape how deterministic or creative the output becomes.

High-yield chooser

Pattern separation that matters on AIF-C01

Why this matters

Many AIF-C01 questions are not really about one AWS product. They are about picking the right FM application pattern for the risk, freshness, and workflow requirement.

Common traps

• calling everything an agent when the real need is only retrieval-backed question answering
• trying to solve freshness problems by changing the model instead of adding RAG
• treating inference settings as if they change model knowledge rather than response behavior
• picking the most autonomous-looking option when the stem rewards control and simplicity

Harder scenario question

A support assistant answers well most of the time, but it keeps missing new policy updates stored in internal documents. The team does not need tool calling or multi-step action. Which change is strongest first?

• A. Add RAG so the model can retrieve current policy content
• B. Convert the whole app into an agent by default
• C. Increase temperature for more creative answers
• D. Ignore the freshness problem

Correct answer: A. The issue is missing current knowledge, which points first to retrieval and grounding rather than agency.

Decision order that usually wins

1. Decide whether the problem is stale knowledge, workflow/tool use, output randomness, or model fit.
2. Add retrieval when the knowledge needs freshness and grounding.
3. Use agent patterns only when the workflow truly needs multi-step reasoning or tools.
4. Tune inference parameters when the issue is style or determinism, not knowledge.
5. Revisit model choice only after pattern and parameter fit are clear.

Quiz