Create a chat-style REST API in Spring Boot

Introduction

A chat API for our series accepts a user message (and optional session id) and returns an assistant reply string. We separate the HTTP shape from the AI call with ChatService and LlmClient-style interfaces so the next post can swap providers (OpenAI, Ollama, Azure OpenAI) by config only.

Real-world explanation

• POST /api/chat with JSON body, 201 or 200 (pick one and document; here 200 with a synthetic id is fine)
• Id in the path is separate from the “conversation” id we will use in persistence (next posts)
• Validation — max length on the message to control cost and abuse

Step-by-step: DTOs

dto/ChatRequest.java

package com.example.demoservice.dto;
 
import jakarta.validation.constraints.NotBlank;
import jakarta.validation.constraints.Size;
 
public record ChatRequest(
    @NotBlank
    @Size(max = 8000) String message
) { }

dto/ChatResponse.java

public record ChatResponse(
    String reply
) { }

Service interface (implementation in next post)

package com.example.demoservice.service;
 
public interface ChatService {
  String replyToUserMessage(String userMessage);
}

Controller

package com.example.demoservice.web;
 
import com.example.demoservice.dto.*;
import com.example.demoservice.service.ChatService;
import jakarta.validation.Valid;
import org.springframework.web.bind.annotation.*;
 
@RestController
@RequestMapping("/api/chat")
public class ChatController {
 
  private final ChatService chatService;
 
  public ChatController(ChatService chatService) {
    this.chatService = chatService;
  }
 
  @PostMapping
  public ChatResponse ask(@RequestBody @Valid ChatRequest request) {
    String text = chatService.replyToUserMessage(request.message());
    return new ChatResponse(text);
  }
}

Wiring: provide a ChatService @Bean of your implementation in the next article (a @Service that calls the LLM).

Common mistakes

• Blocking the EventLoop does not apply to Servlet stack by default—still set sensible timeouts on the outbound client
• Echoing unescaped model text in HTML in a future web view would need escaping; here it is JSON only
• No input length limit = bills and DoS risk

Best practices

• Rate limit (bucket per user or IP) before or in conjunction with this endpoint in production
• Structured logging with a request id (MDC) so support can correlate user report with one line in Kibana/CloudWatch later
• Feature flag the path in config for staged rollout

Final summary

You have a clean REST face for “send message, get reply.” The next post connects ChatService to a real model over HTTP with keys from the environment.