• EnglishEspañol日本語한국어Português
  • Log inStart now

AI monitoring APIs

When you've instrumented your app with AI monitoring, you can access some APIs for collecting token count and user feedback. To use the AI monitoring APIs, check that your Java agent is updated to version 8.12.0 or higher.

This doc provides procedures for updating your code to access the token count and user feedback APIs.

Record token count

If you've disabled the agent with ai_monitoring.record_content.enabled=false, you can use the setLlmTokenCountCallback(LlmTokenCountCallback llmTokenCountCallback) API to calculate token count attributes. This calculates token counts for events related to LLM embedding and completion processes without recording message contents themselves. If you want to collect token counts, follow these steps:

  1. Implement the LlmTokenCountCallback so it overrides the calculateLlmTokenCount(String model, String content) method. This calculates a token count based on a given LLM model name and the LLM message content or prompt:

    class MyTokenCountCallback implements LlmTokenCountCallback {
    public int calculateLlmTokenCount(String model, String content) {
    // Implement custom token calculating logic here based on the LLM model and content.
    // Return an integer representing the calculated token count.
    return 0;
  2. Create an instance of the LlmTokenCountCallback implementation to register the callback, then pass it to the setLlmTokenCountCallback API. For example:

    LlmTokenCountCallback myTokenCountCallback = new MyTokenCountCallback();
    // The callback needs to be registered at some point before invoking the LLM model

To use the callback, implement LlmTokenCountCallback so that it returns an integer that represents the number of tokens for a particular prompt, completion message, or embedding. If the values are less than or equal to 0, LlmTokenCountCallbacks won't attach to an event. Keep in mind that you should only call this API once. Calling this API multiple times will replace each previous callback.

Record user feedback

AI monitoring can correlate trace IDs between a generated message from your LLM models and an end user's feedback. The recordLlmFeedbackEvent API creates an argument with a map of the LlmFeedbackEventAttributes.Builder class. If you want to record user feedback, follow these steps:

  1. Use the TraceMetadata.getTraceId() API to acquire the trace ID for transactions as they execute:

    String traceId = NewRelic.getAgent().getTraceMetadata().getTraceId();
  2. Add the recordLlmFeedbackEvent(Map<String, Object> llmFeedbackEventAttributes) to correlate the trace ID with a feedback event. Here is an example of how you might record an LLM feedback event:

    String traceId = ... // acquired directly from New Relic API or retrieved from some propagation mechanism
    Map<String, String> metadata = new HashMap<>();
    metadata.put("interestingKey", "interestingVal");
    LlmFeedbackEventAttributes.Builder llmFeedbackEvenAttrBuilder = new LlmFeedbackEventAttributes.Builder(traceId, ratingString);
    Map<String, Object> llmFeedbackEventAttributes = llmFeedbackEvenAttrBuilder

If the user feedback records a different thread or different service from where the LLM prompt or response occurred, you need to acquire the trace ID from the originating thread or service. Once you've acquired the trace ID, propagate it to where the user feedback event will be recorded.

To view the parameters that the LlmFeedbackEventAttributes.Builder class takes, review the method details in our AI monitoring API doc.

Add custom LLM attributes

You can adjust your agent to collect custom LLM attributes:

  • Any custom attributes added with the NewRelic.addCustomParameter(...) API can be prefixed with llm. This automatically copies those attributes to LlmEvents
  • If you're adding custom attributes to LlmEvents with the addCustomParameters API , make sure the API call occurs before invoking the Bedrock SDK.
  • One optional custom attribute with special meaning is llm.conversation_id. You can use this to group LLM messages into specific conversations in APM.
Copyright © 2024 New Relic Inc.

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.