June 7, 2026
Quit While You’re Ahead
KNN early termination in Manticore Search
Search results got faster, and the comments instantly turned into a speed-vs-accuracy cage match
TLDR: Manticore made its search engine faster by teaching it to stop once results have basically settled, cutting a lot of wasted work on big searches. Commenters loved the speed boost but instantly split over the eternal drama: smart efficiency win, or dangerous flirtation with “close enough”?
Manticore Search just rolled out a clever way to stop searching sooner when the answer is already obvious, and yes, the community immediately turned that into a full-blown comment-section showdown. The basic idea is simple: when the system is hunting for similar results, it often keeps doing extra work long after the best matches are already clear. Manticore says its new "early termination" trick cuts that wasted effort hard — on a 1 million-item test, the work dropped to about 65% at smaller result sizes, and as low as 20% when asking for huge piles of results. In plain English: less grinding, more speed.
And the reactions? Very online. One camp basically cheered, calling it the most satisfying kind of engineering win: making things faster by noticing when to quit. Another camp immediately went into suspicious-parent mode: "Cool, but what exactly are we losing when it stops early?" Manticore says the quality hit should stay tiny — around 2-4% compared with doing the full search — but that was enough to spark the classic argument: is a small drop acceptable if users get results faster?
The jokes wrote themselves. Commenters compared it to leaving a party once the vibe peaks, ending a meeting when nobody has new ideas, and knowing when to stop scrolling because the algorithm is just showing you leftovers. The hottest mood in the room was a mix of admiration and side-eye: smart optimization or just a fancy new way to say "good enough"?
Key Points
- •Manticore Search uses HNSW for native vector search across millions of documents and introduced early termination to stop searches once results have effectively converged.
- •The article says HNSW continues exploring even after improvements become rare, creating unnecessary distance computations late in the search.
- •The benefit of early termination increases with larger k values because retrieving more neighbors requires more graph exploration and more post-convergence work.
- •Vector quantization and default 3x oversampling increase candidate exploration and latency, which makes early termination more valuable.
- •On a 1 million-vector benchmark, early termination reduced distance computations to about 65% of full search at k=60, 30% at k=1000, and 20% at k=10000, while targeting only 2-4% precision loss.