Milvus

Milvus is an open-source vector database. LangChain4j can use it as an EmbeddingStore to persist embeddings and run similarity search over them.

Two Milvus modules

There are currently two Milvus integration modules. They are independent and can be used side by side.

Module	Milvus Java SDK	Capabilities	Status
langchain4j-milvus	v1 (MilvusServiceClient)	Dense vector search	Legacy. Built on the deprecated v1 SDK.
langchain4j-milvus-v2	v2 (MilvusClientV2)	Dense + sparse + hybrid search (incl. built-in BM25)	Current. Recommended for new projects.

Transitional naming — this will change in LangChain4j 2.0

The -v2 suffix refers to the Milvus SDK version, not to a version of this module. It is temporary.

In LangChain4j 2.0 the legacy langchain4j-milvus module will be removed, and langchain4j-milvus-v2 will be renamed to langchain4j-milvus (the Maven artifact, the ...milvus.v2 package, and the MilvusV2* class names will all drop the v2).

If you adopt langchain4j-milvus-v2 now, be aware that upgrading to 2.0 will require updating your Maven coordinates, imports, and class names. A migration guide will be provided with the 2.0 release.

New projects should use langchain4j-milvus-v2.

---

langchain4j-milvus-v2 (recommended)

Built on the current Milvus Java SDK v2. Supports dense vector search, sparse vector search, and hybrid (dense + sparse) search — including Milvus' built-in BM25 full-text sparse vectors.

Maven Dependency

<dependency>
    <groupId>dev.langchain4j</groupId>
    <artifactId>langchain4j-milvus-v2</artifactId>
    <version>1.19.0-beta29</version>
</dependency>

API

• MilvusV2EmbeddingStore

Basic usage — dense vector search

This is the standard EmbeddingStore usage. The search mode defaults to VECTOR (dense only).

MilvusV2EmbeddingStore store = MilvusV2EmbeddingStore.builder()
        .uri("http://localhost:19530")        // or .host("localhost").port(19530)
        .collectionName("my_collection")
        .dimension(384)                        // required when a new collection is created
        .build();

store.add(embedding, textSegment);

EmbeddingSearchResult<TextSegment> result = store.search(
        EmbeddingSearchRequest.builder()
                .queryEmbedding(queryEmbedding)
                .maxResults(5)
                .build());

Hybrid search (dense + sparse)

Hybrid search combines a dense vector search with a sparse vector search and merges the results with a re-ranker (Reciprocal Rank Fusion by default). Enable it with searchMode(HYBRID).

There are two ways to produce the sparse vectors, selected via sparseMode:

• BM25 (default) — Milvus computes the sparse vector from your text automatically. You provide the text.
• CUSTOM — you provide the sparse vectors yourself (e.g. from a BGE-M3 model).

Option A — built-in BM25 (text → sparse, computed by Milvus)

MilvusV2EmbeddingStore store = MilvusV2EmbeddingStore.builder()
        .uri("http://localhost:19530")
        .collectionName("bm25_collection")
        .dimension(384)
        .searchMode(MilvusV2EmbeddingStore.SearchMode.HYBRID)
        .sparseMode(MilvusV2EmbeddingStore.MilvusSparseMode.BM25)   // default
        .build();

// Insert: only dense embeddings are needed; Milvus builds the BM25 sparse index from the text.
store.addAll(ids, denseEmbeddings, textSegments);

// Search: provide the dense query embedding and the query text (used for BM25).
MilvusV2EmbeddingSearchRequest request = MilvusV2EmbeddingSearchRequest.milvusBuilder()
        .queryEmbedding(queryDenseEmbedding)
        .query("full-text keywords here")
        .maxResults(10)
        .build();

EmbeddingSearchResult<TextSegment> result = store.search(request);

Option B — custom sparse vectors (e.g. BGE-M3)

MilvusV2EmbeddingStore store = MilvusV2EmbeddingStore.builder()
        .uri("http://localhost:19530")
        .collectionName("hybrid_collection")
        .dimension(384)
        .searchMode(MilvusV2EmbeddingStore.SearchMode.HYBRID)
        .sparseMode(MilvusV2EmbeddingStore.MilvusSparseMode.CUSTOM) // you provide sparse vectors
        .build();

// Insert both dense and sparse embeddings.
List<SparseEmbedding> sparseEmbeddings = List.of(
        new SparseEmbedding(new long[]{1L, 42L, 300L}, new float[]{0.8f, 0.5f, 0.3f}),
        new SparseEmbedding(new long[]{7L, 99L},        new float[]{0.6f, 0.4f}));
store.addAllHybrid(ids, denseEmbeddings, sparseEmbeddings, textSegments);

// Search with a dense query embedding and a sparse query embedding.
MilvusV2EmbeddingSearchRequest request = MilvusV2EmbeddingSearchRequest.milvusBuilder()
        .queryEmbedding(queryDenseEmbedding)
        .sparseEmbedding(querySparseEmbedding)
        .maxResults(10)
        .build();

EmbeddingSearchResult<TextSegment> result = store.search(request);

note

The search mode is a property of the collection schema, set once when the store (and collection) is created. A HYBRID collection has both a dense and a sparse vector field; a VECTOR collection has only a dense field. The two are not interchangeable — switching modes requires a new collection.

Connecting to Zilliz Cloud

MilvusV2EmbeddingStore store = MilvusV2EmbeddingStore.builder()
        .uri("https://xxx.api.gcp-us-west1.zillizcloud.com")
        .token("your-api-key")
        .collectionName("my_collection")
        .dimension(384)
        .build();

You can also pass your own MilvusClientV2 instance via .milvusClient(client).

Compatibility

• Milvus Server 2.5.x or later is recommended (BM25 full-text search requires 2.5+; hybrid search requires 2.4+).
• Java 17+

---

langchain4j-milvus (legacy)

Built on the Milvus Java SDK v1. Supports dense vector search only. It will be removed in LangChain4j 2.0 (see the note above); new projects should prefer langchain4j-milvus-v2.

Maven Dependency

<dependency>
    <groupId>dev.langchain4j</groupId>
    <artifactId>langchain4j-milvus</artifactId>
    <version>1.18.0-beta28</version>
</dependency>

API

• MilvusEmbeddingStore

Creation

There are 2 ways to create a MilvusEmbeddingStore:

1. Let the store create a MilvusServiceClient internally from the host, port, and authentication details:

MilvusEmbeddingStore store = MilvusEmbeddingStore.builder()
    .host("localhost")                         // Host for Milvus instance
    .port(19530)                               // Port for Milvus instance
    .collectionName("example_collection")      // Name of the collection
    .dimension(128)                            // Dimension of vectors
    .indexType(IndexType.FLAT)                 // Index type
    .metricType(MetricType.COSINE)             // Metric type
    .username("username")                      // Username for Milvus
    .password("password")                      // Password for Milvus
    .consistencyLevel(ConsistencyLevelEnum.EVENTUALLY)  // Consistency level
    .autoFlushOnInsert(true)                   // Auto flush after insert
    .idFieldName("id")                         // ID field name
    .textFieldName("text")                     // Text field name
    .metadataFieldName("metadata")             // Metadata field name
    .vectorFieldName("vector")                 // Vector field name
    .build();

2. Pass an existing MilvusServiceClient:

// Set up a custom MilvusServiceClient
MilvusServiceClient customMilvusClient = new MilvusServiceClient(
    ConnectParam.newBuilder()
        .withHost("localhost")
        .withPort(19530)
        .build()
);

// Use the custom client in the builder
MilvusEmbeddingStore store = MilvusEmbeddingStore.builder()
    .milvusClient(customMilvusClient)          // Use an existing Milvus client
    .collectionName("example_collection")      // Name of the collection
    .dimension(128)                            // Dimension of vectors
    .indexType(IndexType.FLAT)                 // Index type
    .metricType(MetricType.COSINE)             // Metric type
    .consistencyLevel(ConsistencyLevelEnum.EVENTUALLY)  // Consistency level
    .autoFlushOnInsert(true)                   // Auto flush after insert
    .idFieldName("id")                         // ID field name
    .textFieldName("text")                     // Text field name
    .metadataFieldName("metadata")             // Metadata field name
    .vectorFieldName("vector")                 // Vector field name
    .build();

---

Examples

• MilvusEmbeddingStoreExample