choose-observability-stacklisted

# Choose an LLM observability stack There's no single best tool - the right choice depends on constraints. Gather them, then map to a recommendation. Base recommendations on this repo's curated list (verified tools + licenses), not on hype. ## Ask these constraints first 1. **Deployment**: SaaS OK, or must self-host / on-prem (data residency, regulated industry)? 2. **Primary need**: tracing/cost, *evaluation* (quality testing), or both? Prompt management too? 3. **Existing stack**: already on Datadog/Grafana/OTel? On LangChain? Using a gateway? 4. **Budget/licensing**: need a permissive OSS license (MIT/Apache), or is a commercial tier fine? (Note AGPL/Elastic-license implications for embedding.) 5. **Code-change tolerance**: want zero-code (proxy) or fine to add an SDK? 6. **Team**: engineers, or also non-technical PMs who need a UI? ## Map constraints → recommendation - **Must self-host, permissive license, want everything** → **Langfuse** (MIT core: tracing + evals + prompts) or **Comet Opik** (Apache-2.0). For eval-heavy local work, **Arize Phoenix**. - **Zero code changes, just want cost + logs** → a **gateway/proxy**: **Helicone** (change base URL), **LiteLLM** or **Portkey** (also routing). - **Already on OTel / want vendor-neutral, future-proof** → emit **OpenTelemetry GenAI semantic conventions** via **OpenLLMetry** or **OpenInference**; export to your existing backend. - **Deep in the LangChain ecosystem** → **LangSmith** (tightest integration; SDK OSS, backen