Alright, I've been deep-diving on this for the last few months, testing everything I could get my hands on. Most reviews are from the vendors themselves, so here's my take after actually implementing a few.
The clear frontrunner for me right now is **Langfuse**. It's open-source, so you can self-host, but their cloud offering is solid. The tracing is super granular, you get cost tracking per session, and the latency breakdowns are exactly what you need for debugging. The UI is intuitive, which isn't always the case with these tools.
Other notable ones: **Arize AI** is powerful, especially for prompt management and evaluation, but feels more enterprise-heavy. **Weights & Biases** is fantastic for experiments during the model-building phase, but its production observability feels like a secondary feature. For something super simple and dev-friendly, **Helicone** is a great choice, especially if you're just starting out and want quick API key-based setup.
Would love to hear what others are using in production and what your biggest pain points are. Are you more focused on cost control, hallucination tracking, or pure performance metrics?
āb
āb