Vector demo, exact kNN, live
You are looking at Bitcoin address behavioural vectors15-float behavioural vectors, one per address in one Semurg substrate. Ask it to run exact kNN over every vector and it runs the real query live on this box, on data you can download and re-run yourself ▸.
Click an example in the agent rail, or ask recall at 128 bits for the density tradeoff. Every run is measured end to end, on this box, on the real dataset. Nothing is pre-computed.
Full-precision float-cosine scan over every stored vector: it recovers 100% of the true neighbours (recall 1.0), no index to build and no recall cliff. At this corpus size exact brute-force beats approximate HNSW on both recall and speed. Measured live.
the vector-DB leader
RECALL CLIFF. An approximate HNSW index trades recall for speed: it returns the wrong neighbours some fraction of the time, and you must build the graph index before the first query.
semurg, one engine
100% recall@10, zero index. The exact float-cosine topK over every stored vector IS the ground truth, so no neighbour is ever missed, and there is nothing to build first.
It beats the approximate index on recall AND speed, with no index at all.
Bitcoin address behavioural vectors, derived from the live value-flow. Measured live on this box once the deriver has materialized them.
ask the agent to "search", the exact-scan recall and QPS render here
Honest framing: exact brute-force wins on both recall and speed at this corpus size, so that is the headline. It is not a universal law. As the corpus grows past the exact-scan budget, the binary-density code above becomes the right lever (denser storage, a bounded recall tradeoff), and it runs in the same engine that also serves graph, search and key-value.