Skip to main content

Hazelcast

Hazelcast is a distributed in-memory data grid and computing platform. LangChain4j integrates with Hazelcast through two modules, split so the open-source path stays free of any Enterprise/licensing requirement:

  • langchain4j-community-hazelcast (open source) — provides HazelcastChatMemoryStore, a ChatMemoryStore backed by a Hazelcast IMap. Runs against the open-source Community Edition with no license.
  • langchain4j-community-hazelcast-enterprise (requires Hazelcast Enterprise) — provides HazelcastEmbeddingStore (vector search via VectorCollection) and HazelcastCPMapChatMemoryStore (a strongly-consistent, CP-Subsystem-backed chat memory store). This module re-exports langchain4j-community-hazelcast, so Enterprise consumers also get the IMap-based store from the single dependency.

Maven Dependency​

Open source (Chat Memory Store)​

<dependency>
    <groupId>dev.langchain4j</groupId>
    <artifactId>langchain4j-community-hazelcast</artifactId>
    <version>${latest version here}</version>
</dependency>

The Hazelcast dependency is provided, so add the edition you run against. By default this is the open-source Community Edition:

<dependency>
    <groupId>com.hazelcast</groupId>
    <artifactId>hazelcast</artifactId>
    <version>5.7.0</version>
</dependency>

Hazelcast Enterprise (Embedding Store + CP chat memory)​

<dependency>
    <groupId>dev.langchain4j</groupId>
    <artifactId>langchain4j-community-hazelcast-enterprise</artifactId>
    <version>${latest version here}</version>
</dependency>

Or use the BOM to manage versions consistently:

<dependencyManagement>
    <dependencies>
        <dependency>
            <groupId>dev.langchain4j</groupId>
            <artifactId>langchain4j-community-bom</artifactId>
            <version>${latest version here}</version>
            <type>pom</type>
            <scope>import</scope>
        </dependency>
    </dependencies>
</dependencyManagement>
note

HazelcastEmbeddingStore and HazelcastCPMapChatMemoryStore require Hazelcast Enterprise. com.hazelcast:hazelcast-enterprise is not on Maven Central — it is pulled in transitively by langchain4j-community-hazelcast-enterprise, but you must add the Hazelcast release repository and a valid Enterprise license key:

<repositories>
    <repository>
        <id>hazelcast-release</id>
        <name>Hazelcast Release Repository</name>
        <url>https://repository.hazelcast.com/release/</url>
    </repository>
</repositories>

Provide the license key via config.setLicenseKey(...) or the HZ_LICENSEKEY environment variable. HazelcastCPMapChatMemoryStore additionally requires the CP Subsystem to be enabled; it fails fast otherwise.

Chat Memory Store​

HazelcastChatMemoryStore (open source) stores each chat memory as a JSON-serialised list of ChatMessages in a Hazelcast IMap. The HazelcastInstance you supply may be an embedded member or a thin client — the builder does not distinguish.

// Embedded member
HazelcastInstance hz = Hazelcast.newHazelcastInstance(new Config());

ChatMemoryStore store = HazelcastChatMemoryStore.builder()
        .hazelcastInstance(hz)
        .name("chatMemory") // optional, defaults to "chatMemory"
        .build();
// Client connecting to an external cluster
ClientConfig clientConfig = new ClientConfig();
clientConfig.getNetworkConfig().addAddress("hazelcast-host:5701");
HazelcastInstance hzClient = HazelcastClient.newHazelcastClient(clientConfig);

ChatMemoryStore store = HazelcastChatMemoryStore.builder()
        .hazelcastInstance(hzClient)
        .build();

You can also wrap a pre-configured IMap directly with HazelcastChatMemoryStore.create(IMap<String, String>).

Strongly-consistent variant (Enterprise)​

HazelcastCPMapChatMemoryStore (Enterprise) is an alternative backed by a CPMap in the CP Subsystem. It is linearizable (strongly consistent, Raft-backed), avoiding lost updates when the same memoryId is updated concurrently. The CP Subsystem must be enabled on the instance.

Config config = new Config();
config.getCPSubsystemConfig().setCPMemberCount(3); // CP Subsystem must be enabled
HazelcastInstance hz = Hazelcast.newHazelcastInstance(config);

ChatMemoryStore store = HazelcastCPMapChatMemoryStore.builder()
        .hazelcastInstance(hz)
        .name("chatMemory")
        .build();

Trade-offs versus the IMap store: a CPMap is not partitioned (it must fit within each CP member's RAM, default 100 MB total) and has no TTL/eviction. It suits many small conversations; prefer the IMap-based store for unbounded history across a very large user population.

Embedding Store​

HazelcastEmbeddingStore (Enterprise) is backed by a Hazelcast VectorCollection. The vector index dimension must match the embedding model in use; the metric defaults to COSINE.

HazelcastInstance hz = Hazelcast.newHazelcastInstance(new Config());

EmbeddingStore<TextSegment> store = HazelcastEmbeddingStore.builder()
        .hazelcastInstance(hz)
        .collectionName("embeddings")   // optional, defaults to "embeddings"
        .dimension(384)                 // required, must match the embedding model
        .metric(Metric.COSINE)          // optional, defaults to COSINE
        .build();

You can also wrap a pre-configured VectorCollection with HazelcastEmbeddingStore.create(VectorCollection<String, TextSegmentDocument>). The relevance score returned by search(...) is Hazelcast's already-normalised COSINE score, used directly.

Limitations​

  • removeAll(Filter) is not supported — VectorCollection has no server-side predicate delete; it throws UnsupportedFeatureException. Removal by id, removeAll(Collection<String>) and removeAll() are supported.
  • Metadata filtering during search is not performed server-side. If EmbeddingSearchRequest carries a filter, it is applied client-side after retrieval (a warning is logged), which may return fewer than maxResults matches.

APIs​

  • HazelcastChatMemoryStore — IMap-based (AP), open source
  • HazelcastCPMapChatMemoryStore — CPMap-based (CP, linearizable), Enterprise
  • HazelcastEmbeddingStore — VectorCollection-based vector store, Enterprise

Examples​