Weaviate adds PathSeer, an adaptive filtered vector search, in version 1.40

Why filtered search is a hard problem

Weaviate argues that most vector search use cases need filtering, and that it is probably the most popular query type. A filtered vector search pairs a semantic query, such as books about architecture, with structured criteria like published_date > 2000 or language = 'en'.

The company describes two families of approaches. Filter-aware methods require filters to be defined upfront. They work well on low-cardinality static data, but struggle when filters or even the schema change over time. Filter-agnostic methods, such as ACORN, instead explore the approximate nearest neighbour graph in different ways.

Weaviate says it was one of the first vector databases to ship ACORN as an HNSW filter strategy in late 2024, and that it became the default in 1.34. ACORN handles negatively correlated filters well, but loses recall on very sparse filters. Above 40% selectivity its two-hop expansion costs more than it saves, so Weaviate restricted its use based on selectivity.

How PathSeer works

PathSeer comes from the SIGMOD 2026 paper PathSeer: Adaptive Neighbor Handling for Efficient Filtered ANN Search by Zhiyue Li, Guangyan Zhang and colleagues at Tsinghua University. Zhiyue Li also added PathSeer to Weaviate as part of the research.

The paper frames a core choice during graph search. With distance first, the system computes the distance to every neighbor and checks the filter later. This matches Weaviate's sweeping strategy. It keeps the graph connected, but at low selectivity many distance computations are wasted. With filter first, the filter is checked on each neighbor and distances are computed only for those that pass. This is essentially ACORN, which avoids wasted distance calculations but can disconnect the graph and lower recall.

The paper measured both on six workloads and found the best choice varied by dataset. PathSeer decides per node instead. In the authors' Faiss implementation, it gave 1.2 to 13.5 times the throughput of plain HNSW and 1.6 to 10.6 times the throughput of ACORN at 90% recall@10, for selectivities from 100% down to 0.1%.

Weaviate changed the approach so that separate neighbor lists are no longer needed. PathSeer is therefore compatible with normal HNSW indices, and filter strategies can be changed on an existing index without a rebuild. The modified version has two phases:

• Two-hop burst: while the heap holds fewer than ef entries, each popped node also adds its two-hop neighbors that match the filter. This stops once the heap is full.
• Skipping distances: once the heap is full, if a popped node does not match the filter, its non-matching neighbors are skipped with no distance computed. If it matches, all its neighbors get distances. Non-matching nodes next to matching ones stay in the search as bridges.

Weaviate notes that it resolves filters into a roaring bitmap before the search starts, making a filter check a lookup of around 17 ns, while a distance computation costs roughly 50 to 110 ns depending on quantization and vector size.

Benchmark results

The tests used a Weaviate 1.40 build on a GCP c4-standard-16 instance with 16 vCPUs and 60 GB of memory. The flat search cutoff was disabled so that filtered ANN performance could be measured even where a brute force search would normally trigger. Datasets included BEIR Cohere embeddings, Sphere DPR with 1M 768-dimension embeddings, and wiki-dpr 10M with 27 generated filters.

On the BEIR Cohere topical filters, Weaviate says PathSeer matches or beats both sweeping and ACORN at every selectivity range. At 20% selectivity, where the query vector lies inside the filter region and sweeping previously outperformed ACORN, PathSeer outperforms both. On wiki-dpr 10M, PathSeer clearly outperforms ACORN and also exceeds sweeping on topical filters. At 0.1% selectivity, Weaviate notes brute force search would be the better choice under the default flatSearchCutoff of 40,000 objects.

Conjunction filters, which Weaviate calls the closest to production use, combine a topical filter with a range on object id. Across six such filters spanning 55k to 911k passages, PathSeer was the fastest at 0.90 recall on all six, by 1.1 to 1.4 times over the better of sweeping and ACORN. ACORN reached 0.90 on only three. On a 238k-passage conjunction, PathSeer was 1.36 times sweeping at 0.90 recall and reached 0.98 at ef 512, while ACORN topped out at 0.86. On a 532k-passage conjunction at 5.3% selectivity, PathSeer was 1.3 times sweeping and 1.7 times ACORN at 0.90, with a recall ceiling of 0.98 against 0.93 for ACORN.

Random filters, which have zero correlation with the query, are the case Weaviate says most exposes ACORN, since the search cannot move toward a region with more matching vectors.

What it means for Weaviate users

PathSeer is available in Weaviate 1.40. Because it works with existing HNSW indices, users can switch filter strategy without rebuilding. For workloads with changing filters or combined semantic and range conditions, Weaviate's benchmarks point to higher throughput at the same recall than its two existing strategies. These results come from the company's own tests.