What's changed: Added per-section figures (cert-figure-retrofit). New AI-200 Chapter 1 (vector data = embeddings/distance metric/ANN index/semantic vs keyword search, Azure vector-enabled data services = Azure AI Search/Cosmos DB for NoSQL/PostgreSQL+pgvector/Azure Managed Redis and selection criteria)
1.1Vector Data and Similarity Search
Understand the "vector" concepts behind the data layer of AI apps (especially RAG)—vectorization via embeddings, semantic search by similarity (distance), and approximate nearest neighbor (ANN) indexes—from a developer’s view.
AI-200 succeeds AZ-204 and certifies a developer who builds AI cloud solutions with a back-end focus. Where AZ-204 centered on generic Azure development (App Service, Functions, storage), AI-200 centers on the data layer, compute, and integration that power AI. The starting point is "vector data." For grounding (RAG) in generative AI apps, text is converted to vectors (arrays of numbers) via embeddings, and search uses the property that closer meaning = closer distance.
1.1.1Semantic search by similarity
When the query and documents are vectorized the same way, both live in the same vector space. Search finds the document vectors nearest the query vector, where nearness is measured by a distance metric like cosine similarity or Euclidean distance. That is why search works by semantic closeness rather than keyword match. With many documents, brute force is costly, so production uses approximate nearest neighbor (ANN) indexes (e.g., HNSW) to quickly narrow near candidates.
| Term | Meaning |
|---|---|
| Embedding | Convert text, etc. into a meaning vector |
| Vector space | Place query and documents in one space |
| Distance metric | Cosine similarity, Euclidean distance, etc. |
| ANN index | Fast near-candidate retrieval (e.g., HNSW) |
Common: (1) basis for "find semantically close docs" = embeddings (vectors) + distance metric (e.g., cosine). (2) fast search over many docs = ANN index (e.g., HNSW). (3) vector search is semantic closeness, not keyword match. (4) RAG data layer = storing vectors and similarity search.
Watch out: (1) embedding (vectorize meaning) vs generation (write text) are different roles. (2) vector search (semantic) vs keyword search (match)—hybrid is effective in practice. (3) ANN is approximate—a speed/recall trade-off (not exact nearest neighbor). (4) Vector dimensionality depends on the model and must match the store’s configuration.
1.1.2Section summary
- AI-200 succeeds AZ-204 with a focus on the data layer, compute, and integration powering AI
- Vectorize via embeddings; semantic search by a distance metric (e.g., cosine)
- Speed up many docs with ANN indexes (e.g., HNSW) (approximate = speed/recall trade-off)
- Vector search (semantic) vs keyword search (match) differ; hybrid is practical
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which technique converts text into meaning-capturing vectors for similarity search?
Q2. What is commonly used to measure "semantic closeness" between two vectors?
Q3. Which index is used to quickly retrieve near vectors from many document vectors?
Q4. Which best describes vector search vs keyword search?
Q5. Which is correct about approximate nearest neighbor (ANN)?