Vector databases and embeddings

About this course

A course on turning text, images, or product cards into vectors and building search that doesn't lie to you. It covers similarity metrics and their traps, chunking strategy, exact versus approximate search, hybrid search with keywords, and retrieval quality metrics.

Five modules move from what an embedding is through similarity and its traps, data in the index, search at scale, and quality. In "Similarity" you fix a metric and normalization and build diagnostics on 20 vectors. In "Search at scale" you design a retrieval layer, justifying exact versus approximate indexing. In "Quality" you assemble a labeled set with a share of negative queries and an update policy.

For people building search or recommendations on their own data — you'll need a real domain and at least a draft collection of objects. Requires Python and basic familiarity with model APIs. This isn't a full RAG course — the focus is specifically vectors, indexes, and retrieval.

What you'll learn

• Choose a similarity metric — cosine, dot product, Euclidean — to fit the task
• Recognize when a similarity metric lies: length, language, domain
• Chunk data and decide what belongs in a chunk beyond its text
• Choose between exact search and an approximate index based on collection size
• Combine metadata filters with vector search
• Build hybrid search: vectors plus keywords
• Assemble a query set with reference answers and compute retrieval metrics
• Reindex a collection on model change without downtime

What you'll end up with

• A similarity-properties report with norm checks on 20 vectors
• A "data in the index" document with a chunking and enrichment strategy
• A designed and measured retrieval layer sized for your collection
• A vector-search quality policy with a labeled dataset (paid tiers include reviewer feedback and a certificate)

Not for you if you want a turnkey RAG recipe — this digs into how vector search itself works, not the full answer-generation pipeline.