Jiwar

Jiwar (جوار) — Arabic for "neighbor"

Vector similarity search algorithms, built from scratch.

The goal is to understand how nearest-neighbor search works under the hood — starting from brute force and progressing to graph-based approximate methods.

How it works

A small chess knowledge base (openings, tactics, endgames) is chunked, embedded with all-MiniLM-L6-V2, and used as the test dataset.

Algorithms implemented:

Module	Algorithm	Approach
brute.py	Brute Force	Compares query against every vector. O(n) per query — exact but slow.
nsw.py	Navigable Small World	Builds a graph where each node knows its m closest neighbors, then searches by hopping through edges. Approximate but fast.