Skip to main content
The openinference-instrumentation package provides core utilities for instrumenting Java applications with OpenInference semantic conventions.

​
Installation

implementation 'com.arize:openinference-instrumentation:0.1.1'

​
OITracer

The OITracer class is a wrapper around OpenTelemetry’s Tracer that provides convenience methods for creating spans with OpenInference semantic conventions.

​
Package

com.arize.instrumentation.OITracer

​
Basic Usage

import io.opentelemetry.api.GlobalOpenTelemetry;
import io.opentelemetry.api.trace.Tracer;
import com.arize.instrumentation.OITracer;

// Get OpenTelemetry tracer
Tracer otelTracer = GlobalOpenTelemetry.getTracer("my-application");

// Create OITracer
OITracer tracer = new OITracer(otelTracer);

​
With TraceConfig

import com.arize.instrumentation.OITracer;
import com.arize.instrumentation.TraceConfig;

// Create tracer with custom configuration
TraceConfig config = TraceConfig.builder()
    .hideInputMessages(true)
    .hideOutputMessages(false)
    .build();

OITracer tracer = new OITracer(otelTracer, config);

​
Creating Spans

import io.opentelemetry.api.trace.Span;
import io.opentelemetry.api.trace.SpanKind;
import com.arize.semconv.trace.SemanticConventions;

// Create a span
Span span = tracer.spanBuilder("chat")
    .setSpanKind(SpanKind.CLIENT)
    .startSpan();

try {
    // Set OpenInference attributes
    span.setAttribute(
        SemanticConventions.OPENINFERENCE_SPAN_KIND,
        SemanticConventions.OpenInferenceSpanKind.LLM.getValue()
    );
    span.setAttribute(SemanticConventions.LLM_MODEL_NAME, "gpt-4");
    span.setAttribute(SemanticConventions.LLM_PROVIDER, "openai");
    
    // Your LLM call here
    String response = callLLM(prompt);
    
    // Set response attributes
    span.setAttribute(SemanticConventions.LLM_TOKEN_COUNT_PROMPT, 150L);
    span.setAttribute(SemanticConventions.LLM_TOKEN_COUNT_COMPLETION, 75L);
} finally {
    span.end();
}

​
TraceConfig

The TraceConfig class allows you to configure what data is captured or hidden in traces, useful for privacy and compliance requirements.

​
Package

com.arize.instrumentation.TraceConfig

​
Configuration Options

import com.arize.instrumentation.TraceConfig;

TraceConfig config = TraceConfig.builder()
    // Hide all inputs
    .hideInputs(false)
    
    // Hide all outputs
    .hideOutputs(false)
    
    // Hide input messages
    .hideInputMessages(false)
    
    // Hide output messages
    .hideOutputMessages(false)
    
    // Hide input images
    .hideInputImages(false)
    
    // Hide output images
    .hideOutputImages(false)
    
    // Hide input text
    .hideInputText(false)
    
    // Hide output text
    .hideOutputText(false)
    
    // Hide input audio
    .hideInputAudio(false)
    
    // Hide output audio
    .hideOutputAudio(false)
    
    // Hide input embeddings
    .hideInputEmbeddings(false)
    
    // Hide output embeddings
    .hideOutputEmbeddings(false)
    
    // Hide prompt templates
    .hidePromptTemplate(false)
    
    // Hide prompt template variables
    .hidePromptTemplateVariables(false)
    
    // Hide prompt template version
    .hidePromptTemplateVersion(false)
    
    // Hide tool parameters
    .hideToolParameters(false)
    
    // Base64 image max length ("unlimited" or numeric string)
    .base64ImageMaxLength("unlimited")
    
    .build();

​
Default Configuration

// Get default configuration (nothing hidden)
TraceConfig defaultConfig = TraceConfig.getDefault();

​
Example: Privacy-Focused Configuration

// Configuration for maximum privacy
TraceConfig privateConfig = TraceConfig.builder()
    .hideInputMessages(true)
    .hideOutputMessages(true)
    .hideInputImages(true)
    .hideOutputImages(true)
    .hidePromptTemplateVariables(true)
    .build();

OITracer tracer = new OITracer(otelTracer, privateConfig);

​
Example: Hide Only User Inputs

// Hide user inputs but keep outputs for debugging
TraceConfig config = TraceConfig.builder()
    .hideInputMessages(true)
    .hideInputText(true)
    .hideInputImages(true)
    .hideInputAudio(true)
    .build();

​
Manual Instrumentation Example

Here’s a complete example of manually instrumenting an LLM call: 
import io.opentelemetry.api.GlobalOpenTelemetry;
import io.opentelemetry.api.trace.Span;
import io.opentelemetry.api.trace.SpanKind;
import io.opentelemetry.api.trace.StatusCode;
import io.opentelemetry.api.trace.Tracer;
import io.opentelemetry.context.Scope;
import com.arize.instrumentation.OITracer;
import com.arize.instrumentation.TraceConfig;
import com.arize.semconv.trace.SemanticConventions;
import com.fasterxml.jackson.databind.ObjectMapper;
import java.util.*;

public class LLMClient {
    private final OITracer tracer;
    private final ObjectMapper mapper = new ObjectMapper();
    
    public LLMClient() {
        Tracer otelTracer = GlobalOpenTelemetry.getTracer("llm-client");
        this.tracer = new OITracer(otelTracer);
    }
    
    public String chat(String userMessage) throws Exception {
        Span span = tracer.spanBuilder("chat")
            .setSpanKind(SpanKind.CLIENT)
            .startSpan();
        
        try (Scope scope = span.makeCurrent()) {
            // Set span kind
            span.setAttribute(
                SemanticConventions.OPENINFERENCE_SPAN_KIND,
                SemanticConventions.OpenInferenceSpanKind.LLM.getValue()
            );
            
            // Set model information
            span.setAttribute(SemanticConventions.LLM_MODEL_NAME, "gpt-4");
            span.setAttribute(SemanticConventions.LLM_PROVIDER, "openai");
            span.setAttribute(SemanticConventions.LLM_SYSTEM, "openai");
            
            // Set input messages
            List<Map<String, String>> messages = List.of(
                Map.of("role", "user", "content", userMessage)
            );
            span.setAttribute(
                SemanticConventions.INPUT_VALUE,
                mapper.writeValueAsString(messages)
            );
            span.setAttribute(
                SemanticConventions.INPUT_MIME_TYPE,
                "application/json"
            );
            
            // Set invocation parameters
            Map<String, Object> params = Map.of(
                "temperature", 0.7,
                "max_tokens", 1000
            );
            span.setAttribute(
                SemanticConventions.LLM_INVOCATION_PARAMETERS,
                mapper.writeValueAsString(params)
            );
            
            // Call LLM (replace with actual implementation)
            String response = callOpenAI(userMessage);
            
            // Set output
            List<Map<String, String>> responseMessages = List.of(
                Map.of("role", "assistant", "content", response)
            );
            span.setAttribute(
                SemanticConventions.OUTPUT_VALUE,
                mapper.writeValueAsString(responseMessages)
            );
            span.setAttribute(
                SemanticConventions.OUTPUT_MIME_TYPE,
                "application/json"
            );
            
            // Set token counts
            span.setAttribute(SemanticConventions.LLM_TOKEN_COUNT_PROMPT, 150L);
            span.setAttribute(SemanticConventions.LLM_TOKEN_COUNT_COMPLETION, 75L);
            span.setAttribute(SemanticConventions.LLM_TOKEN_COUNT_TOTAL, 225L);
            
            span.setStatus(StatusCode.OK);
            return response;
            
        } catch (Exception e) {
            span.recordException(e);
            span.setStatus(StatusCode.ERROR, e.getMessage());
            throw e;
        } finally {
            span.end();
        }
    }
    
    private String callOpenAI(String message) {
        // Implement actual OpenAI API call
        return "This is a response from GPT-4";
    }
}

​
Integration with OpenTelemetry

OITracer works seamlessly with OpenTelemetry’s context propagation: 
import io.opentelemetry.api.trace.Span;
import io.opentelemetry.context.Context;

// Parent span
Span parentSpan = tracer.spanBuilder("parent-operation")
    .startSpan();

try (Scope scope = parentSpan.makeCurrent()) {
    // Child span automatically inherits context
    Span childSpan = tracer.spanBuilder("child-operation")
        .startSpan();
    
    try {
        // Child operation
    } finally {
        childSpan.end();
    }
} finally {
    parentSpan.end();
}

​
Dependencies

The base instrumentation package depends on:
  • OpenTelemetry API (1.49.0)
  • OpenTelemetry SDK (1.49.0)
  • OpenTelemetry Instrumentation API
  • OpenInference Semantic Conventions
  • Jackson for JSON handling
  • SLF4J for logging

​
Next Steps