Vector Search

Relevant source files

The following files were used as context for generating this wiki page:

• src/query.rs
• examples/query_search.rs
• src/iterator.rs
• src/proto/milvus.proto.milvus.rs
• tests/aggressivehpctesting/phase_2_search_and_query.rs
• src/proto/milvus.proto.schema.rs

Vector Search is a core functionality of the Milvus Rust SDK that allows users to perform Approximate Nearest Neighbor (ANN) searches, hybrid searches combining multiple vector fields, and structured data queries. It leverages high-performance gRPC communications to interact with the Milvus server, supporting complex filtering expressions, various consistency levels, and paginated results via iterators.

Sources: src/query.rs:1-15, src/iterator.rs:242-260

Search Architecture and Data Flow

The search process begins with a client-side request that encapsulates query vectors, search parameters, and optional filters. The SDK serializes these inputs into gRPC-compatible protobuf messages, specifically SearchRequest or HybridSearchRequest, before transmitting them to the Milvus proxy.

Search Execution Flow

The following diagram illustrates the sequence of operations performed when a user initiates a vector search.

Vector Search Components

ANN Search Requests

Individual search operations are defined by the AnnSearchRequest struct. This is primarily used in hybrid search scenarios where multiple vector fields are queried simultaneously.

Field	Type	Description
data	Vec<Value>	The query vector data (Float or Binary).
anns_field	String	The specific vector field name to target.
param	Vec<KeyValuePair>	Search-specific params (e.g., nprobe, metric_type).
limit	usize	Maximum number of results for this specific request.
expr	Option<String>	Boolean expression to filter entities.

Sources: src/query.rs:65-83, src/proto/milvus.proto.milvus.rs:885-899

Search Options

The SearchOptions struct provides a builder-like interface to configure search behavior for the client.search method.

let options = SearchOptions::new()
    .limit(10)
    .output_fields(vec!["id".to_string(), "age".to_string()])
    .add_param("metric_type", "L2")
    .add_param("anns_field", "picture");

Sources: examples/query_search.rs:64-70, src/query.rs:434-450

Hybrid Search and Ranking

Hybrid search allows for multi-vector field queries, merging results through ranking algorithms. The SDK provides two primary rankers:

1. WeightedRanker: Applies user-defined weights to different search requests.
2. RrfRanker: Uses Reciprocal Rank Fusion (RRF) with a configurable k factor (defaulting to 60).

Data Query and Retrieval

The query method is used for metadata-based filtering without vector similarity. It supports boolean expressions and template-based parameterization to prevent injection and simplify complex logic.

Query Options and Consistency

The SDK supports multiple consistency levels to balance performance and data freshness:

• Strong: Ensures all operations see the latest data (guarantee_timestamp = 0).
• Bounded: Provides bounded staleness.
• Eventually: Provides eventual consistency for maximum performance.
• Session: Guarantees consistency within a single client session.

Sources: src/query.rs:31-50, src/query.rs:549-573

Advanced Search Patterns

Search Iterators

For handling large result sets that exceed gRPC message limits, the SearchIterator and QueryIterator provide a streaming interface. These iterators manage session timestamps and cursors internally to fetch data in batches.

Parameter	Default	Description
batch_size	1000	Number of entities fetched per internal gRPC call.
limit	Unlimited	Total maximum entities to return across all iterations.
iterator_cp_file	None	Optional file path to save/load progress checkpoints.

Sources: src/iterator.rs:68-85, src/iterator.rs:253-271

Iterator Sequence Diagram

Summary

Vector Search in the Milvus Rust SDK is implemented as a robust, asynchronous system capable of handling single-vector ANN searches, complex multi-vector hybrid searches, and paginated data retrieval. By utilizing protobuf schemas and structured options, the SDK ensures type safety and high performance across diverse search workloads, from low-latency interactive queries to high-throughput batch processing via iterators.

Sources: src/query.rs:1-25, tests/aggressivehpctesting/phase_2_search_and_query.rs:115-140