TechPrufer
Get started
Milvus
Milvus logoMilvus · Feature trail

StructArray for Entity- and Element-Level Search

1 post · Oct 2025 – Aug 2026

What this trail covers

Milvus 3.0 introduces StructArray for entity- and element-level search, enabling complex data modeling and filtering. This post details the addition of the Array of Structs field type and the MAX_SIM metric, allowing all embeddings for a single entity to be stored in a single record and enabling Milvus to score and return the entity holistically. This eliminates the need for application-layer deduplication, grouping, bucketing, and reranking logic for multi-vector entities. Examples include Wiki. This post elaborates on the practical application of Array of Structs and MAX_SIM for entity-level retrieval, demonstrating how to store multiple vectors and associated metadata within a single record and how the MAX_SIM metric holistically scores entities based on the best match across all their constituent vectors. It provides concrete examples for e-commerce product search and knowledge base retrieval, highlighting the elimination of application-layer deduplication and reranking.

All Milvus trails
  1. One Entity, Many Vectors: Entity- and Element-Level Search with Milvus 3.0 StructArrayThis post details the addition of the StructArray feature to Milvus, enabling a native data model for entities containing ordered arrays of Struct elements. Each element can carry scalar metadata, vector embeddings, or both. The post explains the data model, schema design with PyMilvus, indexing strategies for different vector subfields (e.g., MAX_SIM_COSINE for EmbeddingList search, COSINE for element-level search), and the insert path. It highlights the difference between StructArray's same-element filtering (using MATCH_* family) and parallel arrays, ensuring correct correlation across scalar subfields within the same element. The post also covers how to define and index vector subfields separately for different search modes and the API boundary where StructArray remains an array of structured objects.
  2. Will Gemini Embedding 2 Kill Multi-Vector Search in Vector Databases?This post analyzes the impact of Google's Gemini Embedding 2 on multi-vector search in vector databases like Milvus. It explains that while Gemini Embedding 2 simplifies multimodal data by mapping text, images, video, audio, and documents into a single vector space, it does not replace the need for Milvus's multi-vector search capabilities. Multi-vector search is crucial for scenarios where different modalities represent distinct semantic views (e.g., biometrics, agentic tools) or when combining different retrieval goals (e.g., keyword precision with semantic recall, structured attributes with visual similarity). The post highlights that Milvus's hybrid search and multi-vector capabilities can be used in conjunction with Gemini Embedding 2, allowing for a unified multimodal column alongside additional sparse or attribute-specific vector columns within the same Milvus collection.
  3. Unlocking True Entity-Level Retrieval: New Array-of-Structs and MAX_SIM Capabilities in MilvusThis post details the practical implementation and benefits of the Array of Structs field type and the MAX_SIM metric for entity-level retrieval. It explains how Array of Structs allows bundling multiple vectors and scalar fields within a single record, aligning storage with real-world entities. The MAX_SIM metric is explained as a scoring strategy that compares a query against all vectors within an Array of Structs and uses the maximum similarity as the entity's final score, thus eliminating the need for application-layer deduplication, grouping, and reranking. Concrete examples for e-commerce product search and knowledge base retrieval are provided to illustrate these capabilities.
  4. vLLM Semantic Router + Milvus: How Semantic Routing and Caching Build Scalable AI Systems the Smart WayThis post details the integration of Milvus with the vLLM Semantic Router to create a scalable AI system. It explains how Milvus acts as a semantic cache layer, storing embeddings of queries and their corresponding responses. When a new query arrives, Milvus is used to find semantically similar past queries. If a match is found above a certain similarity threshold, the cached response is returned, avoiding redundant computation and model inference. The post provides a practical guide on setting up Milvus and the Semantic Router locally for testing this caching mechanism, including configuration modifications for Milvus backend and testing with curl requests to demonstrate cache hits and misses.