Vector Search / Vector DB
Inverted Index
Partition vectors into smaller groups using a clustering algorithm and then build an inverted index that maps cluster centroids to records
- Preprocess / quantize vectors to reduce dimensionality.
- Some implementations support incremental updates.
- Examples: IVFFlat, pg Vector
To find a match, use same clustering algorithm to map into a group, then scan that group’s vectors.
Also check nearby groups to improve accuracy.
Graph
- HNSW
- DiskANN
To find a match for a given vector, enter the graph and then greedily choose the next edge that moves closer to that vector.