marp	true
theme	vibeminds
paginate	true
style	/* Two-column layout / .columns { display: flex; gap: 40px; align-items: flex-start; } .column-left { flex: 1; } .column-right { flex: 1; } / Section divider slides */ section.section-divider { display: flex; flex-direction: column; justify-content: center; align-items: center; text-align: center; background: linear-gradient(135deg, #1a1a3e 0%, #4a3f8a 50%, #2d2d5a 100%); } section.section-divider h1 { font-size: 3.5em; margin-bottom: 0.2em; } section.section-divider h2 { font-size: 1.5em; color: #b39ddb; font-weight: 400; } section.section-divider p { font-size: 1.1em; color: #9575cd; margin-top: 1em; }

OmniObserve

Unified LLM & ML Observability for Go

A vendor-agnostic abstraction layer for AI application observability

The Problem

Observability Fragmentation

Multiple LLM observability platforms exist (Langfuse, Opik, Phoenix, etc.)
Each has its own SDK with different APIs
Switching providers requires significant code changes
Teams get locked into specific vendors
No standardized way to instrument Go AI applications

The Solution

OmniObserve

Instrument once, observe anywhere

Single unified API for all providers
Seamless provider switching without code changes
Clean Go idioms and patterns
Full feature support: tracing, evaluation, prompts, datasets

provider, _ := llmops.Open("opik", llmops.WithAPIKey("..."))
// Switch to Langfuse? Just change "opik" to "langfuse"

Key Features

Feature	Description
Unified Interface	Single API across all providers
Full Tracing	Traces, spans, token usage, costs
Evaluation	Metrics, feedback scores
Dataset Management	Create and manage test datasets
Prompt Versioning	Store and version templates
Context Propagation	Automatic via `context.Context`

Supported Providers

Provider	Description	Highlights
Comet Opik	Open-source, self-hosted	Full-featured, prompt versioning
Langfuse	Cloud & self-hosted	Batch ingestion, cost tracking
Arize Phoenix	OpenTelemetry-based	OTel integration, distributed tracing

All providers support core tracing and evaluation capabilities.

Architecture Overview

omniobserve/
├── omniobserve.go        # Main package, re-exports
├── llmops/               # LLM observability
│   ├── llmops.go         # Core interfaces
│   ├── trace.go          # Trace/Span interfaces
│   └── langfuse/         # Langfuse adapter
├── integrations/         # LLM library integrations
│   └── omnillm/          # OmniLLM hook
├── mlops/                # ML operations (future)
└── sdk/                  # Provider SDKs

# Provider adapters in standalone SDKs:
# go-opik/llmops, go-phoenix/llmops

Core Interfaces

type Provider interface {
    Tracer            // Trace/span operations
    Evaluator         // Evaluation and feedback
    PromptManager     // Prompt template management
    DatasetManager    // Test dataset management
    ProjectManager    // Project/workspace management
    AnnotationManager // Span/trace annotations
    io.Closer

    Name() string
}

Composed of specialized interfaces for maximum flexibility.

Tracer Interface

type Tracer interface {
    // Start a new trace
    StartTrace(ctx context.Context, name string, opts ...TraceOption)
        (context.Context, Trace, error)

    // Start a span within current trace
    StartSpan(ctx context.Context, name string, opts ...SpanOption)
        (context.Context, Span, error)

    // Retrieve from context
    TraceFromContext(ctx context.Context) (Trace, bool)
    SpanFromContext(ctx context.Context) (Span, bool)
}

Span Types

Categorize your operations for better observability:

const (
    SpanTypeGeneral   // General operations
    SpanTypeLLM       // LLM API calls
    SpanTypeTool      // Tool/function calls
    SpanTypeRetrieval // Vector search, RAG retrieval
    SpanTypeAgent     // Agent orchestration
    SpanTypeChain     // Chain operations
    SpanTypeGuardrail // Safety checks
)

Quick Start

import (
    "github.com/plexusone/omniobserve/llmops"
    _ "github.com/agentplexus/go-opik/llmops"  // Register provider
)

func main() {
    // Open provider
    provider, _ := llmops.Open("opik",
        llmops.WithAPIKey("your-api-key"),
        llmops.WithProjectName("my-project"),
    )
    defer provider.Close()

    // Start tracing...
}

Basic Tracing Example

ctx := context.Background()

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

// Start a span for the LLM call
ctx, span, _ := provider.StartSpan(ctx, "gpt-4-call",
    llmops.WithSpanType(llmops.SpanTypeLLM),
    llmops.WithModel("gpt-4"),
)

span.SetInput(messages)
// ... call LLM ...
span.SetOutput(response)
span.SetUsage(llmops.TokenUsage{TotalTokens: 150})
span.End()

RAG Pipeline Example

ctx, trace, _ := provider.StartTrace(ctx, "rag-pipeline")

// Retrieval span
ctx, retrieval, _ := provider.StartSpan(ctx, "vector-search",
    llmops.WithSpanType(llmops.SpanTypeRetrieval))
retrieval.SetOutput(documents)
retrieval.End()

// Generation span
ctx, generation, _ := provider.StartSpan(ctx, "generate",
    llmops.WithSpanType(llmops.SpanTypeLLM),
    llmops.WithModel("gpt-4"))
generation.SetOutput(response)
generation.End()

trace.End()

Adding Feedback Scores

// Add quality score to a span
span.AddFeedbackScore(ctx, "relevance", 0.95,
    llmops.WithFeedbackReason("Directly addressed the query"),
    llmops.WithFeedbackCategory("quality"),
)

// Add user satisfaction to a trace
trace.AddFeedbackScore(ctx, "user_satisfaction", 0.8)

// Programmatic evaluation
provider.AddFeedbackScore(ctx, llmops.FeedbackScoreOpts{
    TraceID: trace.ID(),
    Name:    "accuracy",
    Score:   0.92,
})

Dataset Management

// Create a dataset
dataset, _ := provider.CreateDataset(ctx, "qa-test-cases",
    llmops.WithDatasetDescription("Q&A evaluation dataset"),
)

// Add test items
provider.AddDatasetItems(ctx, "qa-test-cases", []llmops.DatasetItem{
    {
        Input:    map[string]any{"query": "What is Go?"},
        Expected: map[string]any{"answer": "A programming language..."},
    },
    {
        Input:    map[string]any{"query": "Who created Go?"},
        Expected: map[string]any{"answer": "Google engineers..."},
    },
})

Prompt Management

// Create a versioned prompt template
prompt, _ := provider.CreatePrompt(ctx, "chat-system",
    `You are {{.role}}. Answer questions about {{.topic}}.`,
    llmops.WithPromptDescription("Main chat template"),
)

// Retrieve latest version
prompt, _ := provider.GetPrompt(ctx, "chat-system")

// Render with variables
rendered := prompt.Render(map[string]any{
    "role":  "a helpful assistant",
    "topic": "programming",
})
// "You are a helpful assistant. Answer questions about programming."

Switching Providers

// Just change the import and provider name!

// Opik
import _ "github.com/agentplexus/go-opik/llmops"
provider, _ := llmops.Open("opik", llmops.WithAPIKey("..."))

// Langfuse
import _ "github.com/plexusone/omniobserve/llmops/langfuse"
provider, _ := llmops.Open("langfuse", llmops.WithAPIKey("..."))

// Phoenix
import _ "github.com/agentplexus/go-phoenix/llmops"
provider, _ := llmops.Open("phoenix", llmops.WithEndpoint("..."))

Your tracing code stays exactly the same!

OmniLLM Integration

Automatically instrument LLM calls via OmniLLM:

import (
    "github.com/plexusone/omnillm"
    omnillmhook "github.com/plexusone/omniobserve/integrations/omnillm"
)

// Create hook with any OmniObserve provider
hook := omnillmhook.NewHook(provider)

// Attach to OmniLLM client
client := omnillm.NewClient(
    omnillm.WithObservabilityHook(hook),
)
// All LLM calls are now automatically traced!

Captures: model, provider, input/output, tokens, streaming, errors

Configuration Options

Client Options

llmops.WithAPIKey("...")
llmops.WithEndpoint("https://...")
llmops.WithProjectName("my-project")
llmops.WithTimeout(30 * time.Second)
llmops.WithDebug(true)

Span Options

llmops.WithSpanType(llmops.SpanTypeLLM)
llmops.WithModel("gpt-4")
llmops.WithProvider("openai")
llmops.WithTokenUsage(usage)

Provider Capabilities Matrix

Feature	Opik	Langfuse	Phoenix
Tracing	✅ Yes	✅ Yes	✅ Yes
Evaluation	✅ Yes	✅ Yes	✅ Yes
Prompts	✅ Yes	🟡 Partial	❌ No
Datasets	✅ Yes	✅ Yes	🟡 Partial
Cost Tracking	✅ Yes	✅ Yes	❌ No
OpenTelemetry	❌ No	❌ No	✅ Yes

Design Patterns Used

Adapter Pattern: Provider implementations wrap native SDKs
Factory Pattern: Dynamic provider registration and instantiation
Functional Options: Clean, extensible configuration
Composition: Provider interface composes specialized interfaces
Context Propagation: Trace/span context via context.Context

Error Handling

// Sentinel errors
if errors.Is(err, llmops.ErrMissingAPIKey) {
    log.Fatal("API key required")
}

// Error classification
if llmops.IsNotFound(err) {
    // Handle 404
}
if llmops.IsRateLimited(err) {
    // Implement backoff
}
if llmops.IsNotImplemented(err) {
    // Feature not available for this provider
}

Roadmap

Current (v0.1.0)

LLM observability with 3 providers
Full tracing, evaluation, datasets
OmniLLM integration for automatic instrumentation

Planned

MLOps interfaces (experiments, model registry)
Additional providers (Lunary, MLflow, W&B)
OTLP export support
Async batch processing

Getting Started

Install

go get github.com/plexusone/omniobserve
go get github.com/agentplexus/go-opik  # For Opik provider

Import

import (
    "github.com/plexusone/omniobserve/llmops"
    _ "github.com/agentplexus/go-opik/llmops"
)

Documentation

GitHub: github.com/plexusone/omniobserve
Go Docs: pkg.go.dev/github.com/plexusone/omniobserve

Summary

OmniObserve provides:

Unified API for LLM observability
Provider flexibility without vendor lock-in
Clean Go patterns (options, interfaces, context)
Comprehensive features (tracing, evaluation, datasets, prompts)
Easy adoption with minimal code changes

Instrument once. Observe anywhere.

Thank You

Questions?

GitHub: github.com/plexusone/omniobserve

provider, _ := llmops.Open("your-choice", ...)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

OmniObserve

Unified LLM & ML Observability for Go

The Problem

Observability Fragmentation

The Solution

OmniObserve

Key Features

Supported Providers

Architecture Overview

Core Interfaces

Tracer Interface

Span Types

Quick Start

Basic Tracing Example

RAG Pipeline Example

Adding Feedback Scores

Dataset Management

Prompt Management

Switching Providers

OmniLLM Integration

Configuration Options

Client Options

Span Options

Provider Capabilities Matrix

Design Patterns Used

Error Handling

Roadmap

Current (v0.1.0)

Planned

Getting Started

Install

Import

Documentation

Summary

OmniObserve provides:

Thank You

Questions?

FilesExpand file tree

PRESENTATION.md

Latest commit

History

PRESENTATION.md

File metadata and controls

OmniObserve

Unified LLM & ML Observability for Go

The Problem

Observability Fragmentation

The Solution

OmniObserve

Key Features

Supported Providers

Architecture Overview

Core Interfaces

Tracer Interface

Span Types

Quick Start

Basic Tracing Example

RAG Pipeline Example

Adding Feedback Scores

Dataset Management

Prompt Management

Switching Providers

OmniLLM Integration

Configuration Options

Client Options

Span Options

Provider Capabilities Matrix

Design Patterns Used

Error Handling

Roadmap

Current (v0.1.0)

Planned

Getting Started

Install

Import

Documentation

Summary

OmniObserve provides:

Thank You

Questions?