Embeddings and Vector Search
Embeddings, similarity search, vector databases, hybrid search with reranking, and where semantic search fails.
5 notes · 15 min total
- 1EmbeddingsHow an embedding turns text into a list of numbers where similar meanings land close together, and why one model must embed everything.3 min
- 2Similarity SearchHow cosine similarity measures how close two vectors are, why it compares direction rather than length, and how top-k search works.2 min
- 3Vector DatabasesWhy vector databases use approximate nearest neighbor search, how HNSW and IVF indexes work, and when pgvector is enough.3 min
- 4Keyword vs Semantic SearchWhere semantic and BM25 keyword search each fail, how hybrid search fuses them with RRF, and how a cross-encoder reranks the results.4 min
- 5Weak SpotsThe predictable ways semantic search fails, the fix for each, and how to decide when hybrid search and reranking are worth the cost.3 min