What is the ACORN-1 method, and why does it matter for filtered search?

ACORN-1 is a filtered ANN search technique Qdrant adopted (adapted from the broader ACORN approach) specifically to handle the case where a filter is so selective that standard filtered HNSW traversal becomes inefficient — needing to explore a large portion of the graph just to find enough points that satisfy a narrow filter condition.

Standard filtered HNSW traversal checks the filter condition as it visits each candidate node, which works well when a reasonable fraction of the graph's neighborhood satisfies the filter; but when the filter is extremely selective — matching only a tiny percentage of the total collection — the traversal can end up wandering through large parts of the graph, mostly hitting non-matching nodes, before accumulating enough valid results.

ACORN-1 addresses this by adapting the traversal strategy specifically for highly selective filter scenarios, aiming to find enough qualifying candidates more directly rather than relying on the same general-purpose traversal logic used for less restrictive filters — giving Qdrant a targeted improvement for exactly the filtered-search edge case that's historically been one of the harder problems in combining ANN search with structured filtering.