LangChain4J Testing Strategies

Overview

Patterns for unit testing with mocks, integration testing with Testcontainers, and end-to-end validation of RAG systems, AI Services, and tool execution.

When to Use

• Unit testing AI services: When you need fast, isolated tests for services using LangChain4j AiServices
• Integration testing LangChain4j components: When testing real ChatModel, EmbeddingModel, or RAG pipelines with Testcontainers
• Mocking AI models: When you need deterministic responses without calling external APIs
• Testing LLM-based Java applications: When validating RAG workflows, tool execution, or retrieval chains

Instructions

1. Unit Testing with Mocks

Use mock models for fast, isolated testing. See references/unit-testing.md.

ChatModel mockModel = mock(ChatModel.class);
when(mockModel.generate(any(String.class)))
    .thenReturn(Response.from(AiMessage.from("Mocked response")));

var service = AiServices.builder(AiService.class)
        .chatModel(mockModel)
        .build();

2. Configure Testing Dependencies

Setup Maven/Gradle dependencies. See references/testing-dependencies.md.

• langchain4j-test - Guardrail assertions
• testcontainers - Containerized testing
• mockito - Mock external dependencies
• assertj - Fluent assertions

3. Integration Testing with Testcontainers

Test with real services. See references/integration-testing.md.

@Testcontainers
class OllamaIntegrationTest {
    @Container
    static GenericContainer<?> ollama = new GenericContainer<>(
        DockerImageName.parse("ollama/ollama:0.5.4")
    ).withExposedPorts(11434);

    @Test
    void shouldGenerateResponse() {
        // Verify container is healthy
        assertTrue(ollama.isRunning());
        await().atMost(30, TimeUnit.SECONDS)
            .until(() -> ollama.getLogs().contains("API server listening"));

        ChatModel model = OllamaChatModel.builder()
                .baseUrl(ollama.getEndpoint())
                .build();

        // Verify model responds before running tests
        assertDoesNotThrow(() -> model.generate("ping"));

        String response = model.generate("Test query");
        assertNotNull(response);
    }
}

4. Advanced Features

Streaming, memory, error handling patterns in references/advanced-testing.md.

5. Testing Workflow

Follow the testing pyramid from references/workflow-patterns.md:

• 70% Unit Tests: Fast, isolated with mocks
• 20% Integration Tests: Real services with health checks
• 10% End-to-End Tests: Complete workflows

70% Unit Tests ─ Mock ChatModel, guardrails, edge cases
20% Integration Tests ─ Testcontainers, vector stores, RAG
10% End-to-End Tests ─ Complete user journeys

Troubleshooting

• Container fails to start: Check Docker daemon is running, verify image exists, increase timeout
• Model not responding: Verify baseUrl is correct, check container logs, ensure model is loaded
• Test timeout: Increase @Timeout duration for slow models, check container resource limits
• Flaky tests: Add retry logic or health checks before assertions

Examples

Unit Test

@Test
void shouldProcessQueryWithMock() {
    ChatModel mockModel = mock(ChatModel.class);
    when(mockModel.generate(any(String.class)))
        .thenReturn(Response.from(AiMessage.from("Test response")));

    var service = AiServices.builder(AiService.class)
            .chatModel(mockModel)
            .build();

    String result = service.chat("What is Java?");
    assertEquals("Test response", result);
}

Integration Test with Testcontainers

@Testcontainers
class RAGIntegrationTest {
    @Container
    static GenericContainer<?> ollama = new GenericContainer<>(
        DockerImageName.parse("ollama/ollama:0.5.4")
    );

    @BeforeAll
    static void waitForContainerReady() {
        await().atMost(60, TimeUnit.SECONDS)
            .until(() -> ollama.getLogs().contains("API server listening"));
    }

    @Test
    void shouldCompleteRAGWorkflow() {
        assertTrue(ollama.isRunning());

        var chatModel = OllamaChatModel.builder()
                .baseUrl(ollama.getEndpoint())
                .build();

        var embeddingModel = OllamaEmbeddingModel.builder()
                .baseUrl(ollama.getEndpoint())
                .build();

        var store = new InMemoryEmbeddingStore<>();
        var retriever = EmbeddingStoreContentRetriever.builder()
                .chatModel(chatModel)
                .embeddingStore(store)
                .embeddingModel(embeddingModel)
                .build();

        var assistant = AiServices.builder(RagAssistant.class)
                .chatLanguageModel(chatModel)
                .contentRetriever(retriever)
                .build();

        String response = assistant.chat("What is Spring Boot?");
        assertNotNull(response);
        assertTrue(response.contains("Spring"));
    }
}

Best Practices

• Use @BeforeEach/@AfterEach for test isolation
• Never call real APIs in unit tests; use mocks
• Include @Timeout for external service calls
• Test both success and error handling scenarios
• Validate response coherence and edge cases

Common Patterns

Mock Strategy

ChatModel mockModel = mock(ChatModel.class);
when(mockModel.generate(anyString())).thenReturn(Response.from(AiMessage.from("Mocked")));
when(mockModel.generate(eq("Hello"))).thenReturn(Response.from(AiMessage.from("Hi")));
when(mockModel.generate(contains("Java"))).thenReturn(Response.from(AiMessage.from("Java")));

Assertion Helpers

assertThat(response).isNotNull().isNotEmpty();
assertThat(response).containsAll(expectedKeywords);
assertThat(response).doesNotContain("error");

Reference Documentation

• Testing Dependencies - Maven/Gradle configuration
• Unit Testing - Mock models, guardrails
• Integration Testing - Testcontainers, real services
• Advanced Testing - Streaming, memory, error handling
• Workflow Patterns - Test pyramid, best practices

Constraints and Warnings

• AI responses are non-deterministic; use mocks for reliable unit tests
• Avoid real API calls in tests to prevent costs and rate limiting
• Integration tests require Docker; use container health checks
• RAG tests need properly seeded embedding stores
• Mock-based tests cannot guarantee actual LLM behavior; supplement with integration tests
• Use test-specific configuration profiles; never affect production data