langchain4j-tool-function-calling-patterns

$2d

1. Annotate Methods with @Tool

Define a tool class with methods annotated @Tool. Provide a description as the first parameter. Use @P for each parameter description.

public class WeatherTools {

    private final WeatherService weatherService;

    public WeatherTools(WeatherService weatherService) {

        this.weatherService = weatherService;

    }

    @Tool("Get current weather for a city")

    public String getWeather(

            @P("City name") String city,

            @P("Temperature unit: celsius or fahrenheit") String unit) {

        return weatherService.getWeather(city, unit);

    }

}

Validate: Create an instance and confirm the class loads without errors.

2. Register Tools with AiServices

Use AiServices.builder() to register tool instances with the chat model.

MathAssistant assistant = AiServices.builder(MathAssistant.class)

    .chatModel(chatModel)

    .tools(new Calculator(), new WeatherTools(weatherService))

    .build();

Validate: Call assistant.chat("What is 2 + 2?") and verify the LLM responds without throwing.

3. Test Tool Invocation End-to-End

Send a prompt that triggers tool usage and verify the tool executes and its result is incorporated.

String response = assistant.chat("What is the weather in Rome?");

System.out.println(response);

Validate: Check logs for tool invocation and confirm the response uses the tool output.

4. Handle Tool Execution Errors

Add error handlers to gracefully manage failures without exposing stack traces.

AiServices.builder(Assistant.class)

    .chatModel(chatModel)

    .tools(new ExternalServiceTools())

    .toolExecutionErrorHandler((request, exception) -> {

        logger.error("Tool '{}' failed: {}", request.name(), exception.getMessage());

        return "An error occurred while processing your request";

    })

    .hallucinatedToolNameStrategy(request ->

        ToolExecutionResultMessage.from(request,

            "Error: tool '" + request.name() + "' does not exist"))

    .toolArgumentsErrorHandler((error, context) ->

        ToolErrorHandlerResult.text("Invalid arguments: " + error.getMessage()))

    .build();

Validate: Trigger an error condition and confirm the LLM receives a safe error message.

5. Optimize for Performance and Scale

Enable concurrent tool execution and set timeouts for long-running tools.

AiServices.builder(Assistant.class)

    .chatModel(chatModel)

    .tools(new DbTools(), new HttpTools())

    .executeToolsConcurrently(Executors.newFixedThreadPool(5))

    .toolExecutionTimeout(Duration.ofSeconds(30))

    .build();

Validate: Run concurrent requests and confirm no thread contention or deadlocks.

Examples

Calculator Tool with Full Class

public class Calculator {

    @Tool("Perform basic arithmetic")

    public double calculate(

            @P("Expression like 2+2 or 10*5") String expression) {

        // Parse and evaluate expression

        return eval(expression);

    }

}

Assistant assistant = AiServices.builder(Assistant.class)

    .chatModel(ChatModel.builder()

        .apiKey(System.getenv("API_KEY"))

        .model("gpt-4o")

        .build())

    .tools(new Calculator())

    .build();

Immediate Return Tool (No LLM Response)

@Tool(value = "Send email notification", returnBehavior = ReturnBehavior.IMMEDIATELY)

public void sendEmail(@P("Recipient email address") String to,

                     @P("Email subject") String subject,

                     @P("Email body") String body) {

    emailService.send(to, subject, body);

}

Dynamic Tool Provider

ToolProvider provider = request -> {

    if (request.userContext().contains("admin")) {

        return List.of(new AdminTools());

    }

    return List.of(new UserTools());

};

AiServices.builder(Assistant.class)

    .chatModel(chatModel)

    .toolProvider(provider)

    .build();

Best Practices

• **Descriptive @Tool names**: Use imperative verbs ("Get", "Send", "Calculate") with clear scope

• **Precise @P descriptions**: Include format, constraints, and valid values — vague descriptions cause incorrect LLM calls

• Safe error handling: Never expose stack traces; return user-friendly error strings

• Timeout configuration: Always set .toolExecutionTimeout() for external service calls

• Concurrent execution: Enable .executeToolsConcurrently() when tools are independent

• Input validation: Validate parameters inside the tool method; return descriptive errors

• Permission checks: Perform authorization inside the tool, not at the AI service level

• Audit logging: Log tool name, parameters, and execution result for debugging and compliance

Common Issues and Solutions

Issue

Solution

LLM calls non-existent tool

Add .hallucinatedToolNameStrategy() returning a safe error message

Tools receive wrong parameters

Refine @P descriptions; add .toolArgumentsErrorHandler()

Tool execution hangs

Set .toolExecutionTimeout(Duration.ofSeconds(N))

Rate limit errors from external API

Add retry logic or rate limiter inside the tool method

LLM ignores tool output

Ensure the tool returns a string the LLM can interpret

See references/error-handling.md for resilience patterns and references/core-patterns.md for parameter and return type details.

Quick Reference

Annotation / API

Purpose

@Tool

Marks a method as a callable tool

@P

Describes a tool parameter for the LLM

@ToolMemoryId

Injects conversation/user ID into the tool

AiServices.builder()

Creates AI service with registered tools

ReturnBehavior.IMMEDIATELY

Execute tool without waiting for LLM response

ToolProvider

Dynamic tool provisioning based on context

executeToolsConcurrently()

Run independent tool calls in parallel

toolExecutionTimeout()

Timeout for individual tool calls

Constraints and Warnings

• Sensitive data: Never pass API keys, passwords, or credentials in @Tool or @P descriptions

• Side effects: Tools that modify data should warn in their description; AI models may call them multiple times

• Large tool sets: Excessive tools confuse LLM models — use ToolProvider for conditional registration

• Blocking operations: Tools should not perform long synchronous I/O without timeout configuration

• Stack trace exposure: Always route exceptions through error handlers that return safe strings

• Parameter precision: Vague @P descriptions directly cause incorrect tool calls — be specific about formats and constraints

• Concurrent safety: Ensure tool classes are stateless or thread-safe when using executeToolsConcurrently()

Related Skills

• langchain4j-ai-services-patterns — High-level AI service configuration

• langchain4j-rag-implementation-patterns — RAG retrieval with tool integration

• langchain4j-spring-boot-integration — Tool registration in Spring Boot applications

References

• references/setup-configuration.md — Maven setup, chat model configuration, first tool registration

• references/core-patterns.md — Basic tool definition, complex parameters, return types

• references/advanced-features.md — Memory context, dynamic tool providers, streaming, immediate return

• references/error-handling.md — Error handlers, retry logic, monitoring

• references/integration-examples.md — Database, REST API, and context-aware tool examples