Milvus 3.0 StructArray: Multi-Vector Search Within One Entity

About the Webinar

A video contains clips, a product has multiple images, and a document spans passages—each with its own embeddings and metadata. Representing the entire entity with one vector can lose important local details, while storing every element as a separate row creates extra work to preserve relationships, filter correctly, and assemble results.

In this webinar, we’ll explore how StructArray in Milvus 3.0 keeps structured elements inside their parent entity while supporting both entity-level and element-level search. Using a video-search example, we’ll walk through schema design, same-element filtering, multi-vector scoring, and hybrid search. We’ll also examine indexing tradeoffs and explain when results should identify an entire video versus the specific clip that matched.

What You’ll Learn

• StructArray Data Modeling — Learn how to store multiple embeddings and aligned metadata within one entity, and design schemas and indexes for video clips, product images, document passages, and other multi-vector data.
• Same-Element Filtering — Understand how MATCH_* operators and element_filter preserve relationships between fields, ensuring that scalar conditions apply to the same clip, image, or passage.
• Entity- and Element-Level Search — Explore EmbeddingList search with MAX_SIM* scoring and element-level search that returns matching offsets, including the separate vector subfields and indexes needed to support both modes.
• Indexing and Hybrid Search Tradeoffs — Compare TokenANN, MUVERA, and LEMUR, and learn how hybrid search preserves element identity or collapses element hits into entity-level scores.