ML//RAG//vector database

A vector database is a database that stores vectors paired with resources (chunked into paragraphs of about 500 tokens), and its core operation is finding the \(N\) nearest vectors to a given one, fast. The math is cosine similarity between the question (embedded on the fly) and the stored vectors.

A vector database is a database that stores vectors paired with resources (chunked into paragraphs of about 500 tokens), and its core operation is finding the NNN nearest vectors to a given one, fast. The math is cosine similarity between the question (embedded on the fly) and the stored vectors.

HNSW (Hierarchical Navigable Small World graphs) builds a graph where similar vectors are neighbors, so a search hops through about 40 nodes instead of checking all 10M vectors, from O(n)O(n)O(n) down to O(log⁡n)O(\log n)O(logn).

IVF (Inverted File Index) clusters the vectors into buckets first and searches only the closest ones: one drawer instead of the whole filing cabinet.

Pinecone, Weaviate, Chroma, Qdrant, Supabase and FAISS are implementations.