Vector search with filters in Neo4j v2026.01 (Preview)

Neo4j v2026.01 introduces a preview feature for vector search with filters, enhancing the functionality by allowing predicates to be applied inside the vector index at query time, which tailors search results to user-specific criteria such as language, tenant, or category. This approach minimizes latency and maintains relevance without extensive over-fetching or scans, and it is part of a broader enhancement that includes native Cypher syntax for vector search, eliminating the need for procedure calls and preparing for advanced AI capabilities. The feature offers three filtering methods: in-index filtering for consistent low-latency retrieval, post-filtering for refining results using richer graph logic, and pre-filtering for exact scoring on a candidate set. Initial testing shows that in-index filtering maintains low latency and high recall accuracy, while Cypher before search performs best with smaller candidate sets, and post-filtering requires over-fetching for high recall accuracy. These developments are intended for evaluation and feedback rather than production use, with a general availability target for Neo4j v2026.02.