Researched and Edited by Rajat Gupta
Last updated: · How we review
Editor's Summary · LLM Observability & Evaluation Software
These tools answer two different questions and most teams need both. Tracing tells you what your application did — which prompts ran, what the retriever returned, where latency and cost went. Evaluation tells you whether the output was any good. LangSmith, Langfuse, Helicone and Traceloop lead on the first; Confident AI, Patronus, TruLens and Arize Phoenix lead on the second.
Self-hosting is the fastest way to shorten the list. Prompts and completions routinely contain customer data, and many teams cannot send them to a hosted service. Langfuse, Phoenix, Opik, Lunary, Athina and LangWatch all run in your own environment.
If you already run an observability stack, Traceloop's OpenLLMetry sends traces into Datadog, Grafana or Honeycomb rather than asking you to adopt another dashboard.
Portkey is the odd one out and worth knowing about: it sits in the request path as a gateway, so it adds routing, fallback and caching alongside observability, and captures traces without any application instrumentation.
For agents specifically, single-response evaluation misses the failure mode that matters — an agent can answer every individual turn well and still fail the conversation. Maxim and LangWatch test at the scenario level.
Quick picks for LLM Observability & Evaluation Software
- Best open-source all-rounder — Langfuse
- Best for RAG evaluation — Arize Phoenix
- Best gateway plus observability — Portkey
Who gets the most from LLM Observability & Evaluation Software
- 1Engineers debugging why a RAG pipeline returns poor answers
- 2Teams gating deployments on evaluation results in CI
- 3Platform owners tracking model cost and latency across providers
How to choose LLM Observability & Evaluation Software
Separate tracing from evaluation before shortlisting, because tools lead on one and follow on the other. If your prompts carry customer data, filter for self-hosting first and accept the shorter list. If you already run Datadog or Grafana, an OpenTelemetry-native tool avoids a second pane of glass. And if you are building agents rather than single-turn features, insist on scenario-level testing — per-response scoring will look healthy while the conversation fails.
Showing 81-22 out of 22