observerai 0.0.3

observerai - Structured Observability for Generative AI

📦 observerai – Structured Observability for Generative AI

observerai is a Python library that enables runtime observability and evaluation Log-as-Metric for Generative AI applications.

It is designed to be multi-provider (OpenAI, Gemini, Claude, etc.) and multi-modal (text, image, embeddings).

All metrics are emitted as structured logs following the Log-as-Metric approach, allowing seamless integration with platforms like GCP, Datadog, New Relic, Elastic, and others.

✅ Features

• 📊 Structured metric logging for LLM usage
• 🧵 Trace & span context via contextvars (thread-safe)
• ⏱️ Latency & token usage tracking
• 🧠 Prompt/response capture
• 🚨 Exception tracing
• 🧩 Flexible decorator interface
• 🔧 Custom metadata support
• 🧱 Built on top of pydantic and structlog

⚙️ Installation

pip install observerai[openai]

Only the OpenAI dependency is included for now. Future versions will support gemini, claude, etc.

🚀 How to Use

In your application, just decorate the function that makes OpenAI requests:

import uuid
from openai import OpenAI
from observerai.openai import metric_chat_create
from observerai.context import TraceContext

client = OpenAI()
TraceContext.set_trace_id(str(uuid.uuid4()))


@metric_chat_create(metadata={"user_id": "123"})
def test_openai_with_metadata():
    return client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[{"role": "user", "content": "Qual a capital da França?"}],
    )


@metric_chat_create()
def test_openai_without_metadata():
    return client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[{"role": "user", "content": "Qual a capital da Argentina?"}],
    )


resposta = test_openai_with_metadata()
print(resposta.choices[0].message.content)

resposta = test_openai_without_metadata()
print(resposta.choices[0].message.content)

📤 Output (Structured Log)

The decorator logs all metrics as a single structured JSON object to stdout:

{
  "trace_id": "fadfd7d6-9150-4327-961f-dad5f048add1",
  "span_id": "c346b100-30d8-4eea-91e3-ddcc67d8d5e0",
  "flow_id": "d52f542f-2211-45e6-94ca-a6d55617787e", 
  "response": {
    "status_code": 200,
    "latency": {
      "time": 481,
      "unit": "ms"
    }
  },
  "exception": null,
  "version": "0.0.1",
  "metadata": {
    "user_id": "123"
  },
  "name": "gpt-4o",
  "provider": "openai",
  "endpoint": "/chat/completions",
  "conversation": {
    "question": "Qual a capital da França?",
    "answer": "A capital da França é Paris."
  },
  "token": {
    "prompt": 14,
    "completion": 9,
    "total": 23
  },
  "evaluation": null,
  "event": "observerai.openai.completion.chat_create",
  "level": "info",
  "timestamp": "2025-03-24T19:21:08.115226Z"
}

🧭 Roadmap

• OpenAI support (text completions)
• Gemini & Claude providers
• RAG evaluations (RagasX)
• Plug-and-play LLM evaluations via TruLens

👥 Contributing

PRs and discussions are welcome. Stay tuned for contribution guidelines and plugin architecture.