Build your own semantic search

What you'll build

You'll build a complete semantic-search engine in TypeScript from the dot product up: write the vector math, a brute-force exact kNN index over 5,183 real scientific abstracts, and the retrieval metrics that actually matter — recall@k, MRR, and nDCG@k. No vector-DB API, no GPU, no Python — the embeddings ship precomputed as frozen Float32Array blobs, so every score is a deterministic function of bytes on disk. You finish with a CLI that takes a natural-language question and returns the abstracts that answer it, scoring the same nDCG@10 a published retrieval paper reports.

• A vector-math kernel written by hand over Float32Array — dot product, L2 norm, and cosine similarity
• A brute-force exact kNN index: score one query against all 5,183 docs and select the top-k with a deterministic tie-break
• Hand-rolled retrieval metrics — recall@k, MRR, and nDCG@k — checked per-query against pytrec_eval, BEIR's evaluator
• A search CLI that embeds your own query live on CPU with a 22M-param sentence model and returns ranked real abstracts

What you'll learn

• What an embedding actually is — a point in a metric space — and why cosine collapses to a plain dot product on normalized vectors
• How exact k-nearest-neighbor search works before anyone reaches for an approximate index, and why it's O(N·d) per query
• How retrieval quality is really measured — recall@k, MRR, nDCG@k — the vocabulary of every embedding bake-off and RAG regression test
• Why correctness (reproducing the exact nearest neighbors) and quality (matching human relevance judgments) are two different numbers you must never conflate

Prerequisites

• Comfortable reading and running TypeScript
• Basic vector arithmetic — adding and multiplying arrays of numbers (we derive the dot product, norm, and cosine from scratch)
• No machine-learning, information-retrieval, or Python background required