Go SDK for Comet ML Opik

Go SDK for Opik - an open-source LLM observability platform by Comet ML.

Current Version: v0.5.0 - See Release Notes

Installation

go get github.com/agentplexus/go-opik

Quick Start

package main

import (
    "context"
    "log"

    "github.com/agentplexus/go-opik"
)

func main() {
    // Create client (uses OPIK_API_KEY and OPIK_WORKSPACE env vars for Opik Cloud)
    client, err := opik.NewClient(
        opik.WithProjectName("My Project"),
    )
    if err != nil {
        log.Fatal(err)
    }

    ctx := context.Background()

    // Create a trace
    trace, _ := client.Trace(ctx, "my-trace",
        opik.WithTraceInput(map[string]any{"prompt": "Hello"}),
    )

    // Create a span for an LLM call
    span, _ := trace.Span(ctx, "llm-call",
        opik.WithSpanType(opik.SpanTypeLLM),
        opik.WithSpanModel("gpt-4"),
    )

    // Do your LLM call here...

    // End span with output
    span.End(ctx, opik.WithSpanOutput(map[string]any{"response": "Hi!"}))

    // End trace
    trace.End(ctx)
}

OmniObserve Integration

The SDK includes a built-in llmops subpackage that implements the OmniObserve llmops.Provider interface. This allows you to use Opik through a unified observability abstraction alongside other providers like Phoenix, Langfuse, etc.

package main

import (
    "context"

    "github.com/agentplexus/omniobserve/llmops"
    _ "github.com/agentplexus/go-opik/llmops" // Register Opik provider
)

func main() {
    // Open the Opik provider through OmniObserve
    provider, err := llmops.Open("opik",
        llmops.WithAPIKey("your-api-key"),
        llmops.WithWorkspace("your-workspace"),
        llmops.WithProjectName("my-project"),
    )
    if err != nil {
        panic(err)
    }
    defer provider.Close()

    ctx := context.Background()

    // Start a trace
    ctx, trace, _ := provider.StartTrace(ctx, "my-operation")
    defer trace.End()

    // Start a span
    ctx, span, _ := provider.StartSpan(ctx, "llm-call",
        llmops.WithSpanType(llmops.SpanTypeLLM),
    )
    span.SetModel("gpt-4")
    span.SetInput("Hello, world!")
    span.SetOutput("Hi there!")
    span.End()
}

This pattern allows you to:

• Switch between observability providers (Opik, Phoenix, Langfuse) without code changes
• Use a consistent API across different LLM observability platforms
• Build provider-agnostic observability tooling

Feature Comparison

Feature	Opik (Python)	go-opik	omniobserve/llmops	Tests	Notes
Tracing
StartTrace	✅	✅	✅	✅
StartSpan	✅	✅	✅	✅
SetInput/Output	✅	✅	✅	✅
SetModel/Provider	✅	✅	✅	✅
SetUsage (tokens)	✅	✅	✅	✅
AddFeedbackScore	✅	✅	✅	✅
TraceFromContext	✅	✅	✅	✅
SpanFromContext	✅	✅	✅	✅
Nested Spans	✅	✅	✅	✅
Span Types	✅	✅	✅	✅	general, llm, tool, guardrail
Duration/Timing	✅	✅	✅	✅
Prompts
CreatePrompt	✅	✅	✅	✅
GetPrompt	✅	✅	✅	✅	By name + optional version
ListPrompts	✅	✅	✅	✅
CreatePromptVersion	✅	✅	❌		Not in omniobserve interface
ListPromptVersions	✅	✅	❌		Not in omniobserve interface
Datasets
CreateDataset	✅	✅	✅	✅
GetDataset	✅	✅	✅		By name
AddDatasetItems	✅	✅	✅	✅
ListDatasets	✅	✅	✅	✅
DeleteDataset	✅	✅	❌		Not in omniobserve interface
Experiments
CreateExperiment	✅	✅	❌		Not in omniobserve interface
LogExperimentItem	✅	✅	❌		Not in omniobserve interface
ListExperiments	✅	✅	❌		Not in omniobserve interface
Projects
CreateProject	✅	✅	✅	✅
GetProject	✅	✅	✅
ListProjects	✅	✅	✅	✅
SetProject	✅	✅	✅	✅
Evaluation
Evaluate	✅	✅	✅	✅	Run metrics
AddFeedbackScore	✅	✅	✅	✅	Record results
Advanced
Distributed Tracing	✅	✅	❌		Not in omniobserve interface
Streaming Spans	✅	✅	❌		Not in omniobserve interface
Attachments	✅	✅	❌		Not in omniobserve interface
HTTP Middleware	❌	✅	❌		Go SDK extension
Local Recording	❌	✅	❌		Go SDK extension
Batching Client	✅	✅	❌		Not in omniobserve interface

Running omniobserve/llmops tests:

# Skip tests when no API key is set
go test -v ./llmops/

# Run tests with API key
export OPIK_API_KEY=your-api-key
export OPIK_WORKSPACE=your-workspace  # optional
go test -v ./llmops/

Configuration

Environment Variables

Variable	Description
OPIK_URL_OVERRIDE	API endpoint URL
OPIK_API_KEY	API key for Opik Cloud
OPIK_WORKSPACE	Workspace name for Opik Cloud
OPIK_PROJECT_NAME	Default project name
OPIK_TRACK_DISABLE	Set to "true" to disable tracing

Config File

Create ~/.opik.config:

[opik]
url_override = https://www.comet.com/opik/api
api_key = your-api-key
workspace = your-workspace
project_name = My Project

Programmatic Configuration

client, err := opik.NewClient(
    opik.WithURL("https://www.comet.com/opik/api"),
    opik.WithAPIKey("your-api-key"),
    opik.WithWorkspace("your-workspace"),
    opik.WithProjectName("My Project"),
)

Features

Traces and Spans

// Create a trace
trace, _ := client.Trace(ctx, "my-trace",
    opik.WithTraceInput(input),
    opik.WithTraceMetadata(map[string]any{"key": "value"}),
    opik.WithTraceTags("tag1", "tag2"),
)

// Create spans (supports nesting)
span1, _ := trace.Span(ctx, "outer-span")
span2, _ := span1.Span(ctx, "inner-span")

// End spans and traces
span2.End(ctx, opik.WithSpanOutput(output))
span1.End(ctx)
trace.End(ctx)

Span Types

// LLM spans
span, _ := trace.Span(ctx, "llm-call",
    opik.WithSpanType(opik.SpanTypeLLM),
    opik.WithSpanModel("gpt-4"),
    opik.WithSpanProvider("openai"),
)

// Tool spans
span, _ := trace.Span(ctx, "tool-call",
    opik.WithSpanType(opik.SpanTypeTool),
)

// General spans (default)
span, _ := trace.Span(ctx, "processing",
    opik.WithSpanType(opik.SpanTypeGeneral),
)

Feedback Scores

// Add feedback to traces
trace.AddFeedbackScore(ctx, "accuracy", 0.95, "High accuracy")

// Add feedback to spans
span.AddFeedbackScore(ctx, "relevance", 0.87, "Mostly relevant")

Context Propagation

// Start trace with context
ctx, trace, _ := opik.StartTrace(ctx, client, "my-trace")

// Start nested spans
ctx, span1, _ := opik.StartSpan(ctx, "span-1")
ctx, span2, _ := opik.StartSpan(ctx, "span-2") // Automatically nested under span1

// Get current trace/span from context
currentTrace := opik.TraceFromContext(ctx)
currentSpan := opik.SpanFromContext(ctx)

Listing Traces and Spans

// List recent traces
traces, _ := client.ListTraces(ctx, page, size)
for _, t := range traces {
    fmt.Printf("Trace: %s (ID: %s)\n", t.Name, t.ID)
}

// List spans for a specific trace
spans, _ := client.ListSpans(ctx, traceID, page, size)
for _, s := range spans {
    fmt.Printf("Span: %s (Type: %s, Model: %s)\n", s.Name, s.Type, s.Model)
}

Distributed Tracing

// Inject trace headers into outgoing requests
opik.InjectDistributedTraceHeaders(ctx, req)

// Extract trace headers from incoming requests
headers := opik.ExtractDistributedTraceHeaders(req)

// Continue a distributed trace
ctx, span, _ := client.ContinueTrace(ctx, headers, "handle-request")

// Use propagating HTTP client
httpClient := opik.PropagatingHTTPClient()

Streaming Support

// Start a streaming span
ctx, streamSpan, _ := opik.StartStreamingSpan(ctx, "stream-response",
    opik.WithSpanType(opik.SpanTypeLLM),
)

// Add chunks as they arrive
for chunk := range chunks {
    streamSpan.AddChunk(chunk.Content,
        opik.WithChunkTokenCount(chunk.Tokens),
    )
}

// Mark final chunk
streamSpan.AddChunk(lastChunk, opik.WithChunkFinishReason("stop"))

// End with accumulated data
streamSpan.End(ctx)

Datasets

// Create a dataset
dataset, _ := client.CreateDataset(ctx, "my-dataset",
    opik.WithDatasetDescription("Test data for evaluation"),
    opik.WithDatasetTags("test", "evaluation"),
)

// Insert items
dataset.InsertItem(ctx, map[string]any{
    "input":    "What is the capital of France?",
    "expected": "Paris",
})

// Insert multiple items
items := []map[string]any{
    {"input": "2+2", "expected": "4"},
    {"input": "3+3", "expected": "6"},
}
dataset.InsertItems(ctx, items)

// Retrieve items
items, _ := dataset.GetItems(ctx, 1, 100)

// List datasets
datasets, _ := client.ListDatasets(ctx, 1, 100)

// Get dataset by name
dataset, _ := client.GetDatasetByName(ctx, "my-dataset")

// Delete dataset
dataset.Delete(ctx)

Experiments

// Create an experiment
experiment, _ := client.CreateExperiment(ctx, "my-dataset",
    opik.WithExperimentName("gpt-4-evaluation"),
    opik.WithExperimentMetadata(map[string]any{"model": "gpt-4"}),
)

// Log experiment items
experiment.LogItem(ctx, datasetItemID, traceID,
    opik.WithExperimentItemInput(input),
    opik.WithExperimentItemOutput(output),
)

// Complete or cancel experiments
experiment.Complete(ctx)
experiment.Cancel(ctx)

// List experiments
experiments, _ := client.ListExperiments(ctx, datasetID, 1, 100)

// Delete experiment
experiment.Delete(ctx)

Prompts

// Create a prompt
prompt, _ := client.CreatePrompt(ctx, "greeting-prompt",
    opik.WithPromptDescription("Greeting template"),
    opik.WithPromptTemplate("Hello, {{name}}! Welcome to {{place}}."),
    opik.WithPromptTags("greeting", "template"),
)

// Get prompt by name
version, _ := client.GetPromptByName(ctx, "greeting-prompt", "")

// Render template with variables
rendered := version.Render(map[string]string{
    "name":  "Alice",
    "place": "Wonderland",
})
// Result: "Hello, Alice! Welcome to Wonderland."

// Extract variables from template
vars := version.ExtractVariables()
// Result: ["name", "place"]

// Create new version
newVersion, _ := prompt.CreateVersion(ctx, "Hi, {{name}}!",
    opik.WithVersionChangeDescription("Simplified greeting"),
)

// List all versions
versions, _ := prompt.GetVersions(ctx, 1, 100)

// List all prompts
prompts, _ := client.ListPrompts(ctx, 1, 100)

HTTP Middleware

import "github.com/agentplexus/go-opik/middleware"

// Wrap HTTP handlers with automatic tracing
handler := middleware.TracingMiddleware(client, "api-request")(yourHandler)

// Use tracing HTTP client for outgoing requests
httpClient := middleware.TracingHTTPClient("external-call")
resp, _ := httpClient.Get("https://api.example.com/data")

// Or wrap an existing transport
transport := middleware.NewTracingRoundTripper(http.DefaultTransport, "api-call")
httpClient := &http.Client{Transport: transport}

Local Recording (Testing)

// Record traces locally without sending to server
client := opik.RecordTracesLocally("my-project")
trace, _ := client.Trace(ctx, "test-trace")
span, _ := trace.Span(ctx, "test-span")
span.End(ctx)
trace.End(ctx)

// Access recorded data
traces := client.Recording().Traces()
spans := client.Recording().Spans()

Batching

// Create a batching client for efficient API calls
client, _ := opik.NewBatchingClient(
    opik.WithProjectName("My Project"),
)

// Operations are batched automatically
client.AddFeedbackAsync("trace", traceID, "score", 0.95, "reason")

// Flush pending operations
client.Flush(5 * time.Second)

// Close and flush on shutdown
defer client.Close(10 * time.Second)

Attachments

// Create attachment from file
attachment, _ := opik.NewAttachmentFromFile("/path/to/image.png")

// Create attachment from bytes
attachment := opik.NewAttachmentFromBytes("data.json", jsonBytes, "application/json")

// Create text attachment
attachment := opik.NewTextAttachment("notes.txt", "Some text content")

// Get data URL for embedding
dataURL := attachment.ToDataURL()

Evaluation Framework

Heuristic Metrics

import (
    "github.com/agentplexus/go-opik/evaluation"
    "github.com/agentplexus/go-opik/evaluation/heuristic"
)

// Create metrics
metrics := []evaluation.Metric{
    heuristic.NewEquals(false),           // Case-insensitive equality
    heuristic.NewContains(false),         // Substring check
    heuristic.NewIsJSON(),                // JSON validation
    heuristic.NewLevenshteinSimilarity(false), // Edit distance
    heuristic.NewBLEU(4),                 // BLEU score
    heuristic.NewROUGE(1.0),              // ROUGE-L score
    heuristic.MustRegexMatch(`\d+`),      // Regex matching
}

// Evaluate
engine := evaluation.NewEngine(metrics, evaluation.WithConcurrency(4))
input := evaluation.NewMetricInput("What is 2+2?", "The answer is 4.")
input = input.WithExpected("4")

result := engine.EvaluateOne(ctx, input)
fmt.Printf("Average score: %.2f\n", result.AverageScore())

LLM Judge Metrics

import (
    "github.com/agentplexus/go-opik/evaluation/llm"
    "github.com/agentplexus/go-opik/integrations/openai"
)

// Create LLM provider
provider := openai.NewProvider(openai.WithModel("gpt-4o"))

// Create LLM-based metrics
metrics := []evaluation.Metric{
    llm.NewAnswerRelevance(provider),
    llm.NewHallucination(provider),
    llm.NewFactuality(provider),
    llm.NewCoherence(provider),
    llm.NewHelpfulness(provider),
}

// Custom judge with custom prompt
customJudge := llm.NewCustomJudge("tone_check", `
Evaluate whether the response maintains a professional tone.

User message: {{input}}
AI response: {{output}}

Return JSON: {"score": <0.0-1.0>, "reason": "<explanation>"}
`, provider)

G-EVAL

geval := llm.NewGEval(provider, "fluency and coherence").
    WithEvaluationSteps([]string{
        "Check if the response is grammatically correct",
        "Evaluate logical flow of ideas",
        "Assess clarity of expression",
    })

score := geval.Score(ctx, input)

LLM Provider Integrations

OpenAI

import "github.com/agentplexus/go-opik/integrations/openai"

// Create provider for evaluation
provider := openai.NewProvider(
    openai.WithAPIKey("your-api-key"),
    openai.WithModel("gpt-4o"),
)

// Create tracing provider (auto-traces all calls)
tracingProvider := openai.TracingProvider(opikClient,
    openai.WithModel("gpt-4o"),
)

// Use tracing HTTP client with existing code
httpClient := openai.TracingHTTPClient(opikClient)

Anthropic

import "github.com/agentplexus/go-opik/integrations/anthropic"

// Create provider for evaluation
provider := anthropic.NewProvider(
    anthropic.WithAPIKey("your-api-key"),
    anthropic.WithModel("claude-sonnet-4-20250514"),
)

// Create tracing provider (auto-traces all calls)
tracingProvider := anthropic.TracingProvider(opikClient)

// Use tracing HTTP client with existing code
httpClient := anthropic.TracingHTTPClient(opikClient)

CLI Tool

# Install CLI
go install github.com/agentplexus/go-opik/cmd/opik@latest

# Configure
opik configure -api-key=your-key -workspace=your-workspace

# List projects
opik projects -list

# Create project
opik projects -create="New Project"

# List traces
opik traces -list -project="My Project" -limit=20

# List datasets
opik datasets -list

# Create dataset
opik datasets -create="evaluation-data"

# List experiments
opik experiments -list -dataset="my-dataset"

API Client Access

For advanced usage, access the underlying ogen-generated API client:

api := client.API()
// Use api.* methods for full API access

Error Handling

// Check for specific errors
if opik.IsNotFound(err) {
    // Handle not found
}

if opik.IsUnauthorized(err) {
    // Handle auth failure
}

if opik.IsRateLimited(err) {
    // Handle rate limiting
}

Tutorials

Agentic Observability

For integrating Opik with agent frameworks like Google ADK and Eino, see the Agentic Observability Tutorial. This tutorial covers:

• Tracing Google ADK agents with tools
• Tracing Eino workflow graphs
• Multi-agent orchestration observability
• Best practices for agent debugging and monitoring

Development

Running Tests

go test ./...

Running Linter

golangci-lint run

Regenerating API Client

The SDK uses ogen to generate a type-safe API client from the Opik OpenAPI specification. When the upstream API changes, regenerate the client:

Prerequisites:

# Install ogen
go install github.com/ogen-go/ogen/cmd/ogen@latest

Update the OpenAPI spec:

# Download latest spec from Opik repository
curl -o openapi/openapi.yaml \
  https://raw.githubusercontent.com/comet-ml/opik/main/sdks/code_generation/fern/openapi/openapi.yaml

Generate the client:

./generate.sh

This script runs ogen, applies necessary fixes, and verifies the build. For detailed documentation on the generation process and troubleshooting, see the Development Guide.

License

MIT License - see LICENSE for details.

Related

• Opik - Open-source LLM observability platform
• Opik Python SDK - Official Python SDK
• Opik Documentation - Official documentation