Space Partition Tree And Graph

• Basically uses document vectors (or equivalent) to extract feature vectors for things that will be searched on.
• Then uses some clustering algorithm so simlar documents are placed in to similar islands.
• You make a tree (bottom-up) of these clusters.
• You can now locate documents by seeing which leaf is "most like" the thing being sought, recursively, until you reach an island of candidates.
• Probably also you can do these tree searches in parallel if you choose.

Typically "L2 distances" or "cosine similarity" are used to see if two feature vectors are similar.

• Ars Technica announcement
• Microsoft AI announcement
• Github repo