Embeddings & Vector Search by Hand: Cosine Similarity, Top-k Retrieval & Recall
Vector length and normalisation, dot product vs cosine with SUMPRODUCT, top-k retrieval with RANK.EQ, LARGE and INDEX/MATCH, Euclidean vs cosine ordering, chunking cost, ANN recall@k, hybrid keyword + vector scoring and RAG retrieval evaluation with hit rate, MRR and a chart: 8 labs on tiny embeddings you can read.
The first lesson is free — no account needed.
About this course
An embeddings and vector search course that runs retrieval on tiny embeddings you can read in a spreadsheet, for developers, data scientists and product people building search or RAG features. You measure vector length and normalise vectors to length 1, score help articles against a customer question with the dot product and with cosine similarity using SUMPRODUCT and see why they disagree, retrieve the top 3 articles with RANK.EQ, LARGE and INDEX/MATCH and a similarity cut-off, compare Euclidean distance with cosine rankings, work out tokens, chunks and embedding cost with overlap, measure the recall@5 of an approximate nearest-neighbour index against exact search, blend keyword and vector scores into a hybrid ranking and evaluate a RAG retriever with hit rate@k and MRR. The first lab is free.
Lessons
- Embeddings as Vectors: Length, Normalisation & the L1 NormFree
- Dot Product vs Cosine Similarity with SUMPRODUCT Pro
- Top-k Retrieval: RANK.EQ, LARGE and INDEX/MATCH Pro
- Euclidean Distance vs Cosine: When Do the Rankings Disagree? Pro
- Chunking: Tokens per Chunk, Number of Chunks & Embedding Cost Pro
- Approximate Search: Recall@k vs Exact Search and the Speed Trade-off Pro
- Hybrid Search: Blending Keyword and Vector Scores Pro
- RAG Retrieval Evaluation: Hit Rate@k, MRR & a Chart Pro