Filing in this dataset is recent and fragmented rather than concentrated in one dominant company. Several assignees that appear in the historical record — including large technology firms and specialist hardware makers — show zero filings in the most recent year, while smaller or newer entrants show sharp year-over-year gains from a low base. This pattern suggests the field has not yet settled around a small number of clear leaders, and rankings from a year ago may not hold going forward.

Graph-based methods, exemplified by the hierarchical navigable small world approach that is the most-cited prior art in this dataset, build a navigable graph over the vector space and traverse it to find near neighbors. Cluster-based methods, such as the newest representative filing in this dataset, first route a query to a nearby cluster and then search within it, sometimes using a graph between clusters rather than between individual vectors. Cluster-first approaches tend to target lower memory footprint and faster search-start latency, while pure graph methods are typically claimed for higher recall at larger scale.

China is the single largest receiving office in this dataset, ahead of the United States, and several of the most-cited records — including a multimodal retrieval-augmented generation method and a data retrieval method and system — originate from Chinese filers. This concentration lines up with heavy commercial deployment of large-model retrieval pipelines in China, where vector search is being claimed as an infrastructure layer underneath generative AI products rather than as a standalone database technology.

US20260154293A1 claims a specific sequence — identifying a search-start cluster, traversing an inter-cluster graph to select further target clusters, then comparing vectors within each selected cluster. It does not automatically block flat graph-only search or cluster-based search that skips the inter-cluster graph step, since those sit closer to earlier, already-cited prior art. Any team building a clustered index should check specifically whether cluster selection in their system uses a graph traversal step, since that is the feature this filing turns on.

Both terms appear directly in the search criteria used to build this dataset, but neither shows up as a large, distinct claim cluster compared to core index-structure filings, which sit overwhelmingly in the G06F16 classification. That gap suggests filtered search and incremental (online) index updates are still relatively open claim territory, even though they are operationally critical for any production vector database that needs to delete, update, or apply metadata filters without full reindexing.

This dataset covers 72 published records, all of which are also counted as 72 distinct patent families in the assignee ranking, meaning there is little duplicate or continuation filing within the set as gathered. Filing activity is heavily weighted toward the most recent years, with 2025 as the peak year so far at 38 filings, so the total family count is likely to keep growing as more recent applications publish.