Describe what Arize Phoenix is and where it diverges from Arize AX

Flashcard·Easy·4.0 · 0·~30s·Asked atAirbnb Cresta Wandb

Attempt it

Phoenix is the open-source, self-hostable observability and evaluation tool from Arize, OTel-native, with strong RAG-focused features (retrieval-quality views, embedding-drift visualization). Arize AX is the hosted enterprise platform with the same conceptual model but adds production-scale storage, ML monitoring for non-LLM models, SLAs, and team features. Phoenix is where you start for free; AX is where you graduate to once volume, retention, and compliance requirements outgrow self-hosting.

Your answer

Arize is the company. They make two products that share the same core model and the same OpenInference attribute conventions. **Phoenix** is the open-source, self-hostable tracing and evaluation tool. You can run it locally with `pip install arize-phoenix` and `phoenix.launch_app()` or as a container in your infrastructure. It is OTel-native (it speaks OTLP) and has strong features for LLM-specific work: visualizing the call tree of an agent, retrieval-quality views for RAG (which chunks were retrieved, with what scores), embedding-drift visualization, and LLM-as-judge evaluators. It is free. **Arize AX** is the hosted enterprise platform. Same conceptual model, same OpenInference shape, same kind of tracing UI. AX layers on top: production-scale storage, longer retention windows, team features (RBAC, shared dashboards), ML monitoring for traditional non-LLM models (tabular model drift, fairness metrics), SLAs, and enterprise support. The relationship: Phoenix is the on-ramp, where most teams start. AX is where they graduate when volume, retention, compliance, or team-collaboration requirements outgrow self-hosting.

Arize Phoenix and Arize AX are two products from one company, structured as an open source plus hosted enterprise pair, sharing data model and instrumentation conventions but diverging on scale, deployment, and breadth. **Phoenix is the OSS tool.** Apache 2.0 licensed, installable as a Python package or a Docker container. It speaks OTLP and the OpenInference semantic conventions (Arize's LLM-specific attribute spec, broadly compatible with OTel gen_ai semconv). Core features are LLM-tracing-first: an agent and RAG-aware UI that visualizes the call tree, retrieval-quality views (per-query top-k with scores and content previews), embedding-drift visualization (UMAP plots of how query and corpus embeddings cluster over time), and a built-in LLM-as-judge evaluator framework. Phoenix can run on a developer laptop, in a self-hosted container, or as a sidecar in your Kubernetes cluster. **Arize AX is the hosted enterprise platform.** Same data model: traces, evals, datasets, scores. Same instrumentation: send OpenInference-shaped OTLP spans, AX displays them with the same views. The differences are operational: - Production-scale storage and longer retention (months to years). - ML monitoring for non-LLM models (classical tabular models, drift detection, fairness metrics, feature attribution). - Team features: RBAC, audit logs, shared dashboards across projects, alerts integrated with PagerDuty and Slack. - SLAs and enterprise support. - Compliance certifications (SOC 2, HIPAA-eligible deployments). **The migration path.** Phoenix is the on-ramp. Teams start with Phoenix self-hosted to validate their instrumentation and use cases. When trace volume exceeds what self-hosting comfortably handles, retention requirements grow, or the team needs broader observability (ML models alongside LLM agents), they move to AX. Migration is low-friction because both consume the same OpenInference-shaped spans; the instrumentation code does not change.

TL;DR

Phoenix is the OSS, self-hostable, OTel-native (OpenInference) tracing and eval tool from Arize with strong RAG focus; AX is the hosted enterprise platform with production-scale storage, ML monitoring, SLAs, and team

Memory aid

Easy to grasp

Imagine a bicycle shop that sells two things from the same showroom. The first is a fully assembled bike you take home for free, tune in your garage, and ride wherever you like. The second is a premium membership where the shop stores your bike, services it on a schedule, lets your whole family share it, and gives your business a contract with a phone number to call. Phoenix is the free bike: you download it, run it on your laptop or in a small container at work, and use all the core features without paying. Arize AX is the membership: the same bike, but kept and looked after at scale by the company, with longer history, more riders, support guarantees, and extra perks for bigger teams. Phoenix is what you start with. AX is where you go when the family gets larger.

Key concepts

Observability and tracing

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Arize is one of three companies that defined the LLM-observability category between 2023 and 2025 (alongside Langfuse and the LangChain-LangSmith pair). Its two-product structure is worth understanding in detail because it represents a deliberate go to market choice that affects how teams adopt the platform.

This section covers what Phoenix is, what AX adds, how OpenInference fits into the picture, and the practical question of when a team should migrate from one to the other.

Phoenix: the open-source on-ramp

Phoenix is an Apache 2.0-licensed Python package and Docker image that runs locally or in your infrastructure. It is OTel-native (accepts OTLP/HTTP spans) and supports two semantic conventions: OpenInference (Arize-led) and OTel gen_ai semconv (OTel-blessed).

Core capabilities

LLM tracing. Visualize the call tree of any LLM application: agent steps, tool calls, model calls, retrieval calls. Drill into spans to see attributes, inputs, outputs.
RAG-specific views. A per-query retrieval panel shows the top-k chunks, their similarity scores, content previews, and corpus version. This is the view Phoenix is best-known for.
Embedding drift. UMAP-based 2D plots of query and corpus embeddings over time. Cluster-shift detection surfaces when corpora go stale or queries drift.
LLM-as-judge evals. Built-in evaluators (hallucination, retrieval relevance, toxicity) plus a framework for custom judges, runnable against datasets defined in Phoenix.
Datasets and experiments. Define labeled test sets in Phoenix, run them through different prompts or models, compare results.

How teams typically run it

Local development: pip install, launch_app in a notebook. Self-hosted staging: Docker container with a Postgres backend, accessible to the dev team. Production self-hosting: container in Kubernetes, OTel Collector routing spans to it. Phoenix can comfortably handle tens of millions of traces per month on modest hardware.

Arize AX: the hosted enterprise platform

OpenInference: the semantic convention behind both

When to migrate from Phoenix to AX

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Capability	Phoenix (OSS)	Arize AX (hosted)
License and deployment	Apache 2.0, self-host	Commercial, hosted
LLM tracing and RAG views	Yes (full)	Yes (full)
OpenInference and OTLP ingest	Yes	Yes
LLM-as-judge evals	Yes	Yes
Non-LLM ML monitoring	No	Yes
Retention	Limited by your storage	Months to years
Team RBAC and audit logs	Minimal	Yes
SLA and enterprise support	No	Yes

Real products, models, and research that use this idea.

Arize Phoenix is distributed via pip (arize-phoenix) and a Docker image; teams run it on laptops, in dev clusters, and in production for medium volumes.
OpenInference, originally Arize-led, is now broadly supported across Phoenix, AX, Langfuse, Datadog LLM Observability, and OTel itself.

What an interviewer would ask next. Try answering before peeking at the approach.

QWhen does a team outgrow Phoenix and need AX?

Three triggers: trace volume above what self-hosting comfortably handles (rough threshold: tens of millions per month, depending on hardware); retention requirements exceeding a few months; compliance or team requirements (SOC 2, RBAC, audit logs).

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Treating Phoenix and Arize AX as competitors; they are tiers in the same product family with the same data model, just different deployment and scale targets.

60 second bullets to scan on the way to the call.

How would you describe Phoenix's licensing and deployment model in one line?
Which RAG-specific features make Phoenix differentiated?

Primary sources. Browse if you want the original framing.