trulens

Why we included this project

Shipping LLM features usually comes down to a harder question than prompt writing: did the change actually help? TruLens answers that by instrumenting any app with a decorator, recording latency, inputs, outputs, tokens, and cost for every step in a structured trace. Its LLM judges score each step and explain themselves, and the judges themselves are graded against human annotations out of the box, so you are not taking the scoring on faith. Tracing is OpenTelemetry-native, which keeps traces portable to any OTLP backend rather than tying you to a proprietary system. The comparison views are the part most useful for teams running several agent versions: score, latency, and cost sit side by side, making it easy to see where a cheaper model genuinely holds up. If you want to move agent quality from guesswork to measurable evidence without sacrificing observability portability, this is a solid place to start.