Overview
Phoenix is an LLM tooling project that helps you evaluate, observe, or optimize large language model applications. It addresses the engineering challenges of working with LLMsโnon-determinism, prompt sensitivity, cost, latency, and observabilityโso you can ship reliable LLM features rather than brittle demos.
Getting Started
Install Phoenix and integrate it with your LLM call site (wrap your model client or point it at your trace endpoint). Run an evaluation suite against your prompts to establish a baseline, then iterate on prompts and models while tracking metrics. Use the observability dashboard to inspect real production traces.
# Install
pip install phoenix
# Evaluate a prompt against a dataset
phoenix eval --prompt ./prompts/summarize.txt --dataset ./data/eval.jsonl
# Observe production calls
phoenix trace --serve :6006Use Cases
Use cases for Phoenix: (1) regression-testing prompts after a model upgrade; (2) A/B testing different prompt templates on real traffic; (3) tracing and debugging failed LLM calls in production; (4) cost and latency monitoring with per-prompt breakdowns; (5) automatic prompt optimization via feedback-driven search.
Tips
Tips for Phoenix: version your prompts like codeโstore them in Git with a changelog. Build an eval set from real (anonymized) production examples, not synthetic ones. Watch for "eval hacking"โoverfitting prompts to your test set. Track token cost alongside quality; a 1% quality gain at 3x cost is rarely worth it.
Ready to try Phoenix?
Visit the official site for the latest version and full documentation.
Visit Phoenix