← Safety

核验标记表示资料完整度，不是产品评分。下列未确认项可能影响选型；请按对应来源逐项确认。The confidence mark describes evidence completeness, not a product score. Unconfirmed items below may affect your choice; check them against the relevant sources.

English summaries and interface labels are available. Detailed source notes below are currently in Chinese.

模型与开放条件Model & openness

平台不绑模型厂商，评测面分三层：LLM-as-judge、自定义 Python 与 TypeScript 评估器、 确定性检查（server-evals 文档当日直抓在册）。内置确定性评估器 contains、exact_match、regex、 levenshtein_distance、json_distance 不需要 LLM 与 API key（同日直抓在册）。 judge 侧支持的厂商适配器全清单本次未逐条核验，记为未知。

The platform is not tied to model vendors; the evaluation surface has three layers — LLM-as-judge, custom Python and TypeScript evaluators, and deterministic checks (server-evals doc direct-fetched today). The built-in deterministic evaluators contains, exact_match, regex, levenshtein_distance and json_distance need no LLM and no API key (fetched the same day). The full list of judge-side vendor adapters supported was not verified item by item — recorded as unknown.

运行位置Where it runs

本机起服务是主路线：pip install arize-phoenix 或 Docker（仓库 README 与 PyPI 当日在册， requires-python ≥3.11），界面是本机起的 Web UI；仓库另有 helm 与 kustomize 目录供自有集群部署。 厂商侧托管线是 Arize AX，不是 Phoenix 本体（README 当日在册）。

Running the service on your own machine is the main route: pip install arize-phoenix or Docker (repo README and PyPI fetched today; requires-python ≥3.11), with a web UI served locally; the repository also ships helm and kustomize directories for your own clusters. The vendor-hosted line is Arize AX, not Phoenix itself (README fetched today).

本地文件Local files

自托管时 trace 与评测数据留在自有环境，经 OpenTelemetry 上报；平台不读本机代码库或工作区文件。 self-hosting 文档分区存在 deployment-options 与 configuration 两个子目录（当日目录结构在册）， 但正文本次未抓到，外部数据库与对象存储的配置面记为未知。

Self-hosted, traces and evaluation data stay in your environment and arrive over OpenTelemetry; the platform does not read your local codebase or workspace files. The self-hosting docs area contains deployment-options and configuration subfolders (directory structure fetched today), but their bodies were not fetched, so external database and object-storage configuration is recorded as unknown.

关机后的任务Tasks after shutdown

服务端评估器可挂在数据集上，随实验自动运行、无需本地代码（server-evals 文档当日在册）； 批量对照靠 experiments 跑数据集。关机续跑不是这类平台的功能面。 评估任务的调度、抽样与重试口径未逐条在册，记为未知。

Server-side evaluators can be attached to datasets and run automatically with experiments, with no local code (server-evals doc fetched today); batch comparison runs on datasets through experiments. Resume-after-shutdown is not a feature surface for this kind of platform. Scheduling, sampling and retry semantics are not on record item by item — unknown.

工具与扩展Tools & extensions

OpenTelemetry 追踪 + Python/JS SDK（Releases 当日在册含 @arizeai/phoenix-cli）； 仓库文档分区含 tracing、prompt-engineering、datasets-and-experiments、evaluation（当日结构在册）。 MCP 接入方式与 API 面本次未核验到官方声明，记为未知。

OpenTelemetry tracing plus Python/JS SDKs (Releases fetched today include @arizeai/phoenix-cli); the in-repo docs areas cover tracing, prompt-engineering, datasets-and-experiments and evaluation (structure fetched today). MCP support and the outward API surface were not verified in official statements — recorded as unknown.

上下文与记忆Context & memory

观测对象是 trace 与 span，评测对象是数据集与实验的对照（datasets-and-experiments 分区含 concepts、how-to、quickstart 文件，当日结构在册）。 数据集版本、split 与实验比对口径未读到正文，记为未知。

What is observed is traces and spans; what is evaluated is datasets against experiments (the datasets-and-experiments area holds concepts, how-to and quickstart files, structure fetched today). Dataset versioning, splits and experiment-comparison semantics were not read — recorded as unknown.

权限与限制Permissions & limits

许可边界最要紧：ELv2 禁止把软件作为托管或管理服务对外提供，企业功能靠 license key （LICENSE 当日直抓在册）。自托管本体无账号体系。 官方文档站与产品页当日对本站抓取通道返回 403，商业条款、数据用途与训练开关未读到，记为未知。

The license boundary matters most: ELv2 forbids providing the software to third parties as a hosted or managed service, and enterprise functionality is gated by a license key (LICENSE fetched today). Self-hosting has no account system. The official docs site and product page returned 403 to this site's fetch channel today, so commercial terms, data usage and training opt-out were not read — recorded as unknown.

适合什么任务Best for

适合已把 OpenTelemetry 当观测底座、要在 notebook 或自有集群里跑评测、需要「评估器挂数据集」 这种批量对照的团队。不适合只要托管控制台、打算把 Phoenix 包一层对外售卖（ELv2 限制这条）， 或要现成红队与合规审计面的场景；本站不测分也不排名。

Fits teams already standardizing on OpenTelemetry who want batch comparisons in a notebook or on their own cluster: traces come in first, become datasets, evaluators run with the experiment, and each score links back to its source call. Not a fit for those who only want a hosted console, or who intend to resell Phoenix wrapped as a service (ELv2 forbids that explicitly), or who need a ready-made red-team or compliance-audit surface. This site does not measure scores and does not rank.

本页由This page is generated from the ai-handbook 数据层生成（CC BY 4.0）。data layer (CC BY 4.0). 方法论与坐标系定义见仓库内For the methodology and coordinate definitions, see METHODOLOGY.md.