voiceeval-sdk 0.1.3

Enterprise-grade Observability and Evaluation SDK for Voice Agents

VoiceEval SDK (Python)

VoiceEval is an enterprise-grade observability and evaluation SDK designed specifically for Voice Agents and LLM-powered applications. It provides detailed tracing, latency breakdown, and cost analysis with zero configuration.

🚀 Key Features

• 🔎 Zero-Config Auto-Instrumentation: Automatically detects and traces calls from major LLM providers (OpenAI, Anthropic, Google Gemini) without any code changes.
• 🛡️ Secure Ingestion Proxy: All traces are sent through a secure proxy (server/), separating your application logic from downstream observability backends (like Langfuse). This ensures you maintain full control over your data and API keys.
• ⚡ High Performance: Built on top of OpenTelemetry, utilizing efficient asynchronous Batch exports (OTLP/HTTP) to ensure negligible runtime overhead.
• 🧩 Standardized Data Model: Uses standard OTel semantic conventions, making your data portable and interoperable with any OTel-compatible backend.

📦 Installation

Install the SDK via pip (or uv):

pip install voiceeval-sdk
# or
uv add voiceeval-sdk

For local development:

git clone https://github.com/voiceeval/voiceeval-sdk.git
cd voiceeval-sdk
pip install -e .

🏁 Quickstart

1. Initialize the Client

Initialize the Client at the start of your application. This single line sets up the OTel exporter and enables auto-instrumentation for all installed LLM libraries.

from voiceeval import Client

# Initialize SDK - connects to your local proxy or prod endpoint
client = Client(
    api_key="your_voiceeval_api_key",  # or set VOICE_EVAL_API_KEY env var
    base_url="http://api.voiceeval.com/v1/traces"
)

2. Run Your Agent

That's it! Any calls to supported libraries like openai or anthropic are now automatically traced.

from openai import OpenAI

# No manual wrapping needed!
client_openai = OpenAI()
response = client_openai.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello world"}]
)

3. Manual Tracing (Optional)

For functions that don't call LLMs (like your business logic or RAG pipeline), use the @observe decorator:

from voiceeval import observe

@observe(name_override="rag_retrieval")
def retrieve_documents(query: str):
    # Your complex logic here
    return docs

🔌 Supported Providers

The SDK automatically instruments the following libraries if they are found in your environment:

Provider	Library	Status
OpenAI	openai	✅ Auto-Instrumented
Anthropic	anthropic	✅ Auto-Instrumented
Google Gemini	google-generativeai	✅ Auto-Instrumented

Note: If a library is not installed, the SDK gracefully skips it.

📄 License

MIT