Vector search is brute-force linear scan — no approximate nearest neighbor

[rohitg00]

Problem

For every search query, the vector index computes cosine similarity with every stored vector. With 10K vectors of 768 dimensions, this is ~7.7M floating-point operations per search.

Location

src/state/vector-index.ts:35-52

Suggested Fix

Use an approximate nearest neighbor algorithm (HNSW, IVF) or a WebAssembly-based library for efficient vector search.

[rohitg00]

Fixed — vector search now uses bounded min-heap with early skip (src/state/vector-index.ts:35-58). O(n) scan with top-K tracking instead of O(n log n) full sort.