Support for AI Framework Instrumentation (e.g., Spring AI Tool Calls)

[AYue-94]

With the rapid adoption of AI capabilities in modern applications, frameworks like Spring AI are becoming increasingly common. These frameworks often involve complex operations such as LLM (Large Language Model) calls, tool/function executions, and retrieval-augmented generation (RAG) steps.

1. Tool Call Instrumentation

Local span should be created, for local tool executions like:

@Tool(description = "Get the current date and time in the user's timezone")
String getCurrentDateTime() {
    return LocalDateTime.now().atZone(LocaleContextHolder.getTimeZone().toZoneId()).toString();
}

2. LLM Instrumentation

metrics or tags should include:

• Token usage (prompt, completion, total)
• Model provider and model name

For LLM executions like:

@RestController
public class ChatController {
    @Autowired
    private ChatModel chatModel;
    @GetMapping("/chatWithTool")
    public Object chatWithTool() {
        String input = "What day is tomorrow?";
        String output = ChatClient.create(chatModel)
                .prompt(input)
                .tools(new DateTimeTools())
                .call()
                .content();
        return output;
    }
}

[wu-sheng]

We are open to AI framework instructions. Meanwhile, if we need metrics from those AI relative tracing, we need more works on the backend.