Otellix

Stop the LLM Bill Shock. Production-grade observability for Go backends — built on OpenTelemetry.

---

💸 The Problem: "Bill Shock" is Real

Most LLM observability tools tell you what you spent after the money is gone. In production, a recursive prompt loop or a single "power user" can rack up thousands in costs before your billing alerts even trigger.

Otellix is different. It brings real-time cost enforcement to the application layer.

---

⚡ Key Features

• ⛓️ LangChainGo Integration: Drop-in callbacks.Handler for LangChainGo. Automatic token, cost, and span capture for chains.
• 🔌 Zero-Config Attribution: Automatically capture user_id and project_id from Gin/Echo middleware. No more manual context plumbing.
• 🗄️ Redis Budget Store: Distributed budget enforcement for high-availability production clusters.
• 🚀 Go-Native & High-Performance: Built for Go 1.22+ using native streaming and OpenTelemetry.
• 📊 Standard-Compliant (OTel): 100% OTel compliant. Works with Jaeger, Honeycomb, Prometheus, and Grafana.
• 💎 Prompt Caching ROI: Track exactly how much you save with Anthropic/OpenAI prompt caching.

---

🚀 Quick Start

go get github.com/oluwajubelo1/otellix

🕹️ Try it Locally (Docker)

Want to see the full stack (Prometheus + Grafana) in action? Run our one-liner demo:

curl -sL https://raw.githubusercontent.com/oluwajubelo1/otellix/main/scripts/demo.sh | bash

---

1. Simple Tracing

Wrap your provider calls to get instant cost attribution and standard OTel spans.

p := openai.New()
result, err := otellix.Trace(ctx, p, params,
    otellix.WithUserID("user_456"),
)

2. LangChainGo Integration

Just add the Otellix handler to your LLM configuration.

import "github.com/oluwajubelo1/otellix/integrations/langchaingo"

handler := langchaingo.NewOtellixHandler(
	langchaingo.WithProvider("openai"), // required for accurate llm.cost_usd
	langchaingo.WithModel("gpt-4o"),
)
p, _ := openai.New(openai.WithCallback(handler))

// Automatically captures token usage, cost, and spans.
res, err := p.GenerateContent(ctx, parts)

The handler records llm.* attributes on a span in the context, or creates its own llm.call span if none exists, so telemetry is always captured. Set WithProvider/WithModel (LangChainGo doesn't expose them in its callbacks) to enable cost tracking.

3. Middleware Attribution

Stop manually passing IDs. Use our middleware to discover identity from headers/JWTs.

import otellixmw "github.com/oluwajubelo1/otellix/middleware"

r := gin.Default()
r.Use(otellixmw.GinMiddleware()) // Automatically populates context with User/Project IDs

4. Local Development

Two dev modes help you see traces while building:

shutdown := otellix.SetupDev()        // (1) Pretty summary + raw OTel span JSON
defer shutdown()

shutdown := otellix.SetupDevConsole() // (2) Pretty summary only — clean terminal
defer shutdown()

• SetupDev() configures an OpenTelemetry stdout trace exporter (human-readable summary and the raw span JSON) — best for inspecting the exact span payload.
• SetupDevConsole() prints only the one-line [otellix] summary per call and leaves your existing TracerProvider untouched — ideal for demos and presentations.
• RegisterDevPrinter() enables just the summary printer when you already have your own TracerProvider configured.

---

📊 Standardized OTel Attributes

Otellix populates your spans with industry-standard attributes:

Attribute	Description
llm.provider	The LLM provider (OpenAI, Anthropic, etc.)
llm.model	The model used (e.g. claude-sonnet-4-6)
llm.input_tokens / llm.output_tokens	Token usage for the call
llm.cache_read_tokens	Tokens served from provider cache (Savings!)
llm.cost_usd	Real-time USD cost of the call

---

🏗️ Supported Providers

Provider wrappers (drop-in Call/Stream):

• Anthropic: Claude Opus/Sonnet/Haiku families with full Prompt Caching metrics.
• OpenAI: GPT-5.x, GPT-4.x, o-series, gpt-oss with caching support.
• Google Gemini: Gemini 3.x, 2.5, 2.0 & 1.5 Pro/Flash.
• Ollama: Local models with zero-dependency streaming.

Pricing-only support (built-in cost table for llm.cost_usd; bring your own provider/API): Kimi (Moonshot), DeepSeek, Qwen, GLM, ERNIE, Hunyuan, Mistral, Grok, plus every model above — 130+ models total. See docs/providers.md.

---

📖 Deep Dives

• Architecture — How Otellix fits into the OTel ecosystem.
• Budget Guardrails — Redis & InMemory stores.
• Integrations — LangChainGo setup.
• Middleware & Identity — Zero-config attribution.
• Prompt Caching — ROI analysis and tracking.

---

🤝 Contributing

We love stars and contributions! Please see CONTRIBUTING.md. Licensed under MIT.