Chapter 07 — Comparisons & Use Cases

Overview

Choose the lightest strategy that meets reliability / transparency requirements. Escalate only when baseline performance or guardrail needs justify cost.

Technique selection matrix

Technique	Best For	Cost	Transparency	Failure Risk Reduced
Maieutic	Traceable commonsense QA	High	High (tree)	Inconsistent justification
Self-Refine	Draft polishing / code quality	Medium (rounds)	Medium	Surface-level errors
Least-to-Most	Compositional multi-step tasks	Low→Medium	Medium (explicit steps)	Premature leaps
Self-Consistency	Math / classification	Linear in K	Low (rationales sampled)	Random reasoning errors
RAG	Factual Q&A / summarization	Index + retrieval calls	High (citations)	Hallucinated facts

Hybrid patterns

LtM + Self-Refine: Decompose, then refine each step for correctness.
RAG + Maieutic: Retrieve passages, build explanation tree citing sources.
ReAct + Self-Consistency: Sample tool-augmented traces; majority answer.

Design hybrids only if each layer adds a distinct mitigation (e.g., hallucination reduction + reasoning robustness).

Escalation checklist

Baseline simple prompt measured & insufficient?
Quantified failure being mitigated (e.g., hallucination rate)?
Added latency within SLO budget?
Instrumentation captures step-level outputs?
Rollback path if hybrid underperforms?