1.1.2. Embeddings and Vector Search from First Principles
💡 First Principle: An embedding is a location, not a summary. An embedding model converts text (or images) into a point in high-dimensional space, positioned so that similar meanings land near each other. Once meaning has coordinates, "find related content" collapses into "find nearby points" — pure geometry that a database can index. Every vector feature on this exam (Cosmos DB, pgvector, Redis) is just a different implementation of this one idea.
Why does the exam's biggest domain revolve around this? Because models have no memory of your data, and you can't paste your entire document corpus into a prompt. Retrieval-augmented generation (RAG) solves this: embed your documents once, store the vectors, and at question time embed the question, find the nearest document vectors, and hand only those documents to the model. The model answers from retrieved facts instead of hallucinating.
Two mechanical details drive many exam questions. First, similarity is measured mathematically — typically cosine similarity (angle between vectors), sometimes dot product or Euclidean distance. You don't need the formulas; you need to know that the query returns the nearest vectors, whether or not anything genuinely relevant exists. Garbage in the store means confident-sounding garbage retrieved. Second, at scale, comparing a query against every stored vector (exact search) is too slow, so databases build approximate nearest neighbor (ANN) indexes — structures like HNSW that check only promising regions of the space. ANN trades a little recall (it can miss a true neighbor) for enormous speed. This trade-off reappears in Cosmos DB index types (2.1.3), pgvector tuning (2.2.3), and Redis (2.3.2).
⚠️ Exam Trap: Vector similarity search returns the top-k nearest results, not the top-k relevant results — and ANN indexes return the approximately nearest. If a question asks why a RAG system retrieves plausible-but-wrong passages, look for answers about embedding quality, missing metadata filters, or ANN recall — not "the database is corrupted."
Reflection Question: Why does a RAG pipeline embed the user's question with the same embedding model used for the documents — what breaks if ingestion and query use different models?
