// Software stack
LLM app monitoring & evaluation
Gate quality with regression tests before release, then record traces, cost and latency in production to catch silent degradation.
5 pieces4 setup stepsscenario: Evaluate and monitor AI quality
What goes in it
- 1 ★ 26k
promptfoo Evaluation, red-teaming & guardrails Open source
Regression tests and red-teaming for prompts and agents
Test your prompts, agents, and RAGs. Red teaming/pentesting/vulnerability scanning for AI. Compare performance of GPT, Claude, Gemini, DeepSeek, and more. Simple declarative configs with command line and CI/CD integration. Used by OpenAI and Anthropic.
- 2 ★ 36k
langfuse Observability & tracing Open source
Traces, cost and scores
🪢 Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.
- 3 ★ 12k
phoenix optional Observability & tracing Open source
OpenTelemetry-based trace analysis
AI Observability & Evaluation
- 4
OpenTelemetry GenAI semantic conventions Protocols, specs & agent payments Open spec
Common trace conventions for GenAI calls
- 5 ★ 19k
deepeval optional Evaluation, red-teaming & guardrails Open source
Unit-test framework for LLM outputs
The LLM Evaluation Framework
How to set it up
- 01Collect 30–100 real user questions with expected answers as a gold set
- 02Turn the gold set into tests with promptfoo or DeepEval and run them before any prompt or model change
- 03Record traces, cost and latency in production with Langfuse (or Phoenix) using OpenTelemetry GenAI conventions
- 04Feed bad conversations back into the gold set and rerun weekly
✓ Good fit
- An LLM feature is live or about to launch
✗ Look elsewhere if
- Still prototyping: compare a few outputs by hand first