What's changed: Added per-section figures (cert-figure-retrofit). New AI-300 Chapter 5 (Domain 5 "Optimize generative AI": RAG optimization = similarity threshold/chunk size/retrieval strategy/embedding model selection-fine-tuning/hybrid search/relevance metrics-A-B testing; advanced fine-tuning = methods/synthetic data/monitoring-optimization/dev-to-production lifecycle management)
5.1Optimizing RAG performance and accuracy
Understand optimizing retrieval performance by tuning similarity thresholds, chunk sizes, and retrieval strategies; selecting and fine-tuning embedding models for domain-specific accuracy; hybrid search combining semantic and keyword retrieval; and evaluating and improving RAG with relevance metrics and A/B testing.
RAG (retrieval-augmented generation) retrieves external context to ground answers and reduce hallucination. RAG quality depends strongly on retrieval quality, so optimizing retrieval is an AI-300 focus.
5.1.1Optimizing retrieval
Optimize retrieval by tuning the similarity threshold (how close context must be to be used), chunk size (how documents are split—too large adds noise, too small loses context), and retrieval strategy (top-k, re-ranking, etc.). The choice of embedding model also drives accuracy, and for domain-heavy terminology, fine-tuning the embedding model improves relevance. Tune iteratively while evaluating.
5.1.2Hybrid search and A/B testing
Hybrid search combines semantic (vector) search (by meaning) with keyword search (by term match) to cover each other’s weaknesses and improve recall and precision. Validate improvements quantitatively with relevance metrics (are the retrieved results appropriate) and A/B testing (compare configurations under near-production conditions). The basis of RAG optimization: "measure → tune → measure again."
| Knob | What | Effect |
|---|---|---|
| Similarity threshold | How close context must be | Balance precision/recall |
| Chunk size | Document split unit | Large=noise / small=lost context |
| Embedding model | Selection / fine-tuning | Domain accuracy |
| Search method | Hybrid (semantic + keyword) | Complement recall/precision |
| Validate improvement | Relevance metrics / A-B testing | Compare quantitatively |
Watch the mix-ups: (1) semantic (vector) search (meaning) vs keyword search (term match)—hybrid combines both. (2) chunks too large add noise; too small lose context—find the right size by evaluation. (3) Validate improvements with relevance metrics + A/B testing (not by gut feeling).
Map requirement → knob: "improve retrieval accuracy for domain terms" = embedding model selection/fine-tuning; "search by both meaning and term match" = hybrid search; "tune how close context must be" = similarity threshold; "tune lost context or noise" = chunk size; "validate improvement quantitatively" = relevance metrics + A/B testing.
5.1.3Section summary
- RAG quality depends on retrieval: tune similarity threshold/chunk size/retrieval strategy
- Improve domain accuracy via embedding model selection/fine-tuning; complement with hybrid search
- Evaluate/improve quantitatively with relevance metrics and A/B testing (measure → tune → measure)
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which RAG search method combines semantic (vector) and keyword matching to improve recall and precision?
Q2. For domain-heavy terminology, what improves retrieval relevance?
Q3. Which combination quantitatively validates RAG improvements?
Q4. Which is correct about RAG chunk size?
Q5. Which RAG knob controls how close retrieved context must be to be used?
Q6. What is the main reason RAG helps improve generative AI quality?
Keep track of your progress
The full study guide is free to read. Sign up free to practice with the question bank, track what you have read, review your mistakes, and highlight passages.