aviaraeye 0.6.0

Ergonomic LLM observability wrapper around Langfuse

AviаraEye

Ergonomic LLM observability wrapper around Langfuse. Provides tracing, prompt management, scoring, and auto-instrumentation for OpenAI, Azure OpenAI, Google Gemini, and Anthropic (Claude).

Installation

pip install aviaraeye[all]

# Or with uv
uv pip install aviaraeye[all]

# Individual providers
pip install aviaraeye[openai]     # OpenAI
pip install aviaraeye[azure]      # Azure OpenAI
pip install aviaraeye[gemini]     # Google Gemini
pip install aviaraeye[anthropic]  # Anthropic (Claude)

Configuration

Set environment variables:

export LANGFUSE_PUBLIC_KEY="pk-lf-..."
export LANGFUSE_SECRET_KEY="sk-lf-..."
export LANGFUSE_HOST="https://aviaraeye.aviaralabs.com"

Quick Start

from aviaraeye import AviaraEye, traced, observe, get_prompt
from aviaraeye.providers.openai import OpenAI

# Initialize
eye = AviaraEye()

# Auto-traced OpenAI calls
client = OpenAI()

with traced("my-pipeline", user_id="user-1", tags=["demo"]) as t:
    response = client.chat.completions.create(
        model="gpt-4",
        messages=[{"role": "user", "content": "Hello!"}],
    )
    t.set_output(response.choices[0].message.content)
    t.score("quality", 1.0)

# Decorator-based tracing
@observe
def process_document(doc: bytes) -> dict:
    return extract(doc)

# Prompt management
prompt = get_prompt("extraction-v2", label="production")
compiled = prompt.compile(schema=my_schema)

# Flush before shutdown
eye.flush()

Features

Feature	Description
Tracing	traced(), span(), event() context managers
Multi-agent	agent_span(), tool_span(), chain_span(), retriever_span(), generation_span()
Concurrency	bind_trace_context() keeps thread-pool work in one trace; W3C carrier for Celery
Decorators	@observe, @generation for auto-tracing functions
Providers	Drop-in OpenAI and Azure OpenAI; Gemini and Claude auto-instrumentation
Prompts	Versioned prompt fetch, compile, and link to traces
Prompt sync	Publish, promote a label, and sync a codebase to Langfuse — idempotent, with a read-only drift check for CI
Scoring	Numeric, boolean, categorical, and text scores
Sessions	session_context() for user/session propagation
Metadata	Metadata, Tags builders for standardised annotation

Provider Usage

OpenAI / Azure OpenAI

from aviaraeye.providers.openai import OpenAI
from aviaraeye.providers.azure_openai import AzureOpenAI

# Drop-in — all calls auto-traced
client = OpenAI()
response = client.chat.completions.create(model="gpt-4", messages=[...])

Google Gemini

from aviaraeye.providers.gemini import instrument_gemini
from google import genai

instrument_gemini()  # One-time setup (auto-called by AviaraEye())

client = genai.Client(api_key="...")
response = client.models.generate_content(model="gemini-2.5-flash", contents="...")

Documentation

• Getting Started
• Tracing Guide
• Provider Integration
• Prompt Management
• Scoring & Evaluation
• User & Session Tracking

Development

# Install in editable mode with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Lint
ruff check src/

# Type check
mypy src/

License

MIT - Aviara Labs