← Safety

LangSmithLangSmith

LangChain

云端Cloud 商业平台Commercial platform 自托管Self-hosted WebUIWebUI evaleval LangChainLangChain 多模型Multiple models partial lifecycle: active 核验Checked 2026-11-06

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.

未确认的信息Information not confirmed

主体归属：LangChain, Inc. 生态；主体细节本次未逐条核验
LLM-as-judge 与 decision model 评测器是否必须自备模型凭据
评分队列的吞吐、延迟与抽样口径
数据保留时长与 TTL 的具体数值口径
是否使用客户 trace 训练模型（privacy 页当日 404）
MCP 接入方式与对外集成面
导出与删除接口清单
备案公示检索

Whether LLM-as-judge and decision-model evaluators require your own model credentials
Queue throughput, latency and sampling rules for online evaluation
Exact retention windows and TTL values
Whether customer traces are used to train models (privacy page 404 today)
MCP support and the outward integration surface
The export and deletion interface inventory
Filing-registry search

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

LangSmith 是托管型的观测与评测平台。当日在册的官方自我定位是「agent 与 LLM 观测平台」，评测是其中一个面板。

适合与不适合When to use it, when not to

适合想要一个控制台同时管 trace、数据集、评测器和告警，又不想自运维观测栈的团队。已有 OpenTelemetry 的应用可以先接进来，之后再决定要不要同时发往别的后端。不适合把许可门槛当次要问题的团队：自托管按官方口径是企业版加购，还要自备集群与外部存储。本站只写它支持哪些评测类型，不写它测出来的分数，也不做同类排名。

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-11-06

出品属地Origin: 国外出品foreign-made　官方产品与文档域名归 LangChain（www.langchain.com 与 docs.langchain.com），按当日在册口径记 foreign；公司注册地未逐条核验。the official product and docs domains belong to LangChain (www.langchain.com and docs.langchain.com); recorded as foreign on today's evidence; place of incorporation not item-verified.
国内使用Direct access: 大陆可直连direct OK　本档 9 个在册源页面当日直抓均 200（旧域名 docs.smith.langchain.com 当日 308 跳新域名）；docs 站的 security 与 otlp 路径当日 404，数据条款面因此记未知。all nine on-record source pages fetched 200 today (the old docs.smith.langchain.com domain 308-redirected to the new one); the docs site's security and otlp paths returned 404 today, which is why the data-terms surface is recorded as unknown.
备案状态CAC filing: 备案未核验filing unknown：本次未检索备案公示名单；托管型开发者平台是否落入备案口径未核验，记入未知项清单。the filing registry was not searched today; whether a hosted developer platform falls under filing scope is unverified — listed under unknowns.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 被观测与被评测的模型都不由平台提供：官方产品页点名可接 OpenAI、Anthropic、Vercel AI SDK、LlamaIndex 与自研实现（当日在册）。评测器面含人工、代码、LLM-as-judge、decision model 与成对比较（evaluation-concepts 当日在册）。judge 是否必须自备模型凭据，该页未说明，记为未知。
Neither the observed nor the evaluated models come from the platform: the official product page names OpenAI, Anthropic, Vercel AI SDK, LlamaIndex and custom implementations as supported (fetched today). The evaluator surface covers human, code, LLM-as-judge, decision model and pairwise comparisons (evaluation-concepts fetched today). Whether judge evaluators require your own model credentials is not stated on that page — recorded as unknown.
运行位置Where it runs: 主形态是厂商托管云（SaaS），另有 BYOC 与企业版自托管：官方写「Self-hosted LangSmith 是 Enterprise 加购」「在你自己的 Kubernetes 集群里跑」，默认捆绑存储、生产建议改用外部存储服务（self-hosted 文档当日在册）。控制台在浏览器，SDK、CLI 与 API 是代码面。
The primary form is the vendor-hosted cloud (SaaS), with BYOC and enterprise self-hosting: the official line reads that self-hosted LangSmith is an add-on to the Enterprise plan and runs in your own Kubernetes cluster, storage is bundled by default and external storage services are recommended for production (self-hosted docs fetched today). The console lives in the browser; SDKs, the CLI and the API are the code surfaces.
本地文件Local files: 数据面默认在厂商侧，除非走 BYOC 或自托管；trace 由 SDK 或 OpenTelemetry 上报， OTLP 端点是 LANGSMITH_ENDPOINT 加 /otel（当日在册）。平台不读本机代码库； 导出与删除接口面本次未逐条核验。
The data plane sits on the vendor side unless you use BYOC or self-hosting; traces are reported via SDK or OpenTelemetry, with the OTLP base URL being LANGSMITH_ENDPOINT plus /otel (fetched today). The platform does not read your local codebase; the export and deletion interface surface was not itemized today.
关机后的任务Tasks after shutdown: 在线评测跑在生产 trace 上（evaluation-concepts 当日在册），离线评测用带参考输出的数据集做回归对照；官方 changelog 当日在册条目提到在线评测器的细粒度权限与自定义视图。 评分队列的吞吐、延迟与抽样口径未在册，记为未知。
Online evaluation runs on production traces (evaluation-concepts fetched today); offline evaluation uses datasets with reference outputs for regression comparison. Today's changelog entries cover granular permissions and custom views for online evaluators. Queue throughput, latency and sampling rules are not on record — recorded as unknown.
工具与扩展Tools & extensions: OpenTelemetry 摄入并可用 collector 做多目的地 fanout（当日在册）；官方 CLI 可查询与管理 traces、runs、datasets、evaluators、experiments、insights、threads（当日在册）；文档索引在册的 Terraform 面把评测器与规则当代码管。MCP 接入方式未核验到官方声明，记为未知。
OpenTelemetry ingestion with collector fanout to multiple destinations (fetched today); the official CLI queries and manages traces, runs, datasets, evaluators, experiments, insights and threads (fetched today); the docs index on record lists a Terraform surface that manages evaluators and rules as code. MCP support was not verified in official statements — recorded as unknown.
上下文与记忆Context & memory: 观测与评测对象按 workspace 与 project 组织，权限按 workspace 配内置或自定义角色（enterprise 文档当日在册）；数据集带参考输出，用于回归对照。 保留时长与 TTL 的具体口径未逐条在册，记为未知；「跨会话记忆」不是本类平台的功能面。
Observed and evaluated objects are organized by workspace and project, with built-in or custom roles defined per workspace (enterprise docs fetched today); datasets carry reference outputs used for regression comparison. Retention windows and TTL specifics were not itemized — recorded as unknown; cross-session memory is not a feature surface for this class of platform.
权限与限制Permissions & limits: 合规标记当日在册：SOC 2 Type II、HIPAA、GDPR（enterprise 文档直抓）；SAML 或 OIDC 单点登录与即时用户配置同页在册。档位与配额事实为 Developer 免费档、Plus 与按需 Enterprise（含自托管与 BYOC），价格数字按 D9 不采。是否用客户 trace 训练模型、数据保留条款正文未抓到（privacy 页当日 404），记为未知。
Compliance markers on record today: SOC 2 Type II, HIPAA and GDPR (enterprise docs direct-fetched); SAML or OIDC single sign-on with just-in-time provisioning on the same page. Tier and quota facts are a free Developer tier, Plus, and custom Enterprise including self-hosted and BYOC; price numbers are excluded per D9. Whether customer traces train models, and the data-retention clause text, were not fetched (the privacy page returned 404 today) — recorded as unknown.
适合什么任务Best for: 适合要「一个控制台管住 trace、数据集、评测器与告警」、又不想自运维观测栈的团队；已用 OpenTelemetry 的应用可先接进来再换后端。不适合把许可门槛当次要问题、或必须整机自托管却买不到企业档的团队；本站不测分也不排名。
Fits teams that want one console covering traces, datasets, evaluators and alerts without operating their own observability stack; applications already on OpenTelemetry can feed in first and decide later whether to fan out elsewhere. Not a fit where licensing thresholds are treated as secondary: self-hosting is, per official docs, an Enterprise add-on that also needs your own cluster and external storage. This site records which evaluation types are supported, not the scores they produce, and does not rank peers.

头号误解Top misconception

把「支持任意框架」当成零成本兼容——接入面靠 SDK 与 OTel，字段口径要在 collector 侧对齐（当日在册）
把自托管当成可自选部署形态——官方口径是 Enterprise 加购，默认捆绑存储、生产要换外部服务
把在线评测当实时告警——异步评分与权限面在 changelog 在册，延迟与抽样口径未逐条核验

证据来源Evidence sources

最近核验记录Latest recorded check 2026-11-06，内容更新日content updated on 2026-11-06。单项复核范围见正文或来源说明，不代表所有字段在同一天重新核验。See the text or source notes for item-specific scope; this does not mean every field was rechecked on that date.

类型Type	名称Name	链接Link
docs	LangSmith · 产品页（自述 Agent 与 LLM Observability Platform；tracing 与 monitoring 与 evaluation；可接任意 LLM 框架；managed cloud 与 BYOC 与 self-hosted，2026-11-06 直抓 200）LangSmith · product page (self-described Agent & LLM observability platform; tracing, monitoring and evaluation; works with any LLM framework; managed cloud, BYOC and self-hosted options, direct-fetched 200 on 2026-11-06)	https://www.langchain.com/langsmith
docs	LangSmith · Evaluation concepts（人工与代码与 LLM-as-judge 与 decision model 与成对评测；离线数据集回归与在线 trace 评测；标注队列，2026-11-06 直抓 200）LangSmith · Evaluation concepts (human, code, LLM-as-judge, decision-model and pairwise evaluators; offline dataset regression and online trace evaluation; annotation queues, direct-fetched 200 on 2026-11-06)	https://docs.langchain.com/langsmith/evaluation-concepts
changelog	LangSmith · Changelog（当日直抓在册，最新条目日期 2026-11-02，覆盖评测、数据集、队列与在线评测器）LangSmith · Changelog (direct-fetched today; newest entry dated 2026-11-02, covering evaluation, datasets, queues and online evaluators)	https://docs.langchain.com/langsmith/changelog
docs	LangSmith · Self-hosted（Enterprise 加购、自有 Kubernetes 集群、默认捆绑存储与生产用外部存储、license key 走销售，2026-11-06 直抓 200）LangSmith · Self-hosted (Enterprise add-on, your own Kubernetes cluster, bundled storage by default and external storage for production, license key via sales, direct-fetched 200 on 2026-11-06)	https://docs.langchain.com/langsmith/self-hosted
docs	LangSmith · Enterprise（自有基础设施托管、SAML 与 OIDC 与即时配置、按 workspace 的角色、SOC 2 Type II 与 HIPAA 与 GDPR，2026-11-06 直抓 200）LangSmith · Enterprise (hosting entirely in your own infrastructure, SAML and OIDC with just-in-time provisioning, per-workspace roles, SOC 2 Type II, HIPAA and GDPR, direct-fetched 200 on 2026-11-06)	https://docs.langchain.com/langsmith/enterprise.md
docs	LangSmith · Trace with OpenTelemetry（支持 OTel 追踪、OTLP 端点为 API 地址加 /otel、LANGSMITH_OTEL_ENABLED、collector 多目的地 fanout，2026-11-06 直抓 200）LangSmith · Trace with OpenTelemetry (OpenTelemetry-based tracing, OTLP base URL is the API URL plus /otel, LANGSMITH_OTEL_ENABLED, collector fanout, direct-fetched 200 on 2026-11-06)	https://docs.langchain.com/langsmith/trace-with-opentelemetry.md
docs	LangSmith · CLI（查询与管理 traces、runs、datasets、evaluators、experiments、insights、threads；curl 安装；OAuth 或 API key 登录，2026-11-06 直抓 200）LangSmith · CLI (query and manage traces, runs, datasets, evaluators, experiments, insights and threads; curl install; OAuth or API-key login, direct-fetched 200 on 2026-11-06)	https://docs.langchain.com/langsmith/langsmith-cli.md
docs	LangSmith · 文档索引（在册含 Terraform 管理评测器与规则即代码等条目，索引页正文 2026-11-06 直抓 200，目标页未逐条抓取）LangSmith · docs index (on record includes a Terraform surface managing evaluators and rules as code; index fetched 200 on 2026-11-06, target pages not itemized)	https://docs.langchain.com/_llms/products-and-setup/lang-smith-setup.txt
pricing	LangSmith · Pricing（Developer 免费档、Plus、按需 Enterprise 含自托管与 BYOC；价格数字不采，2026-11-06 直抓 200）LangSmith · Pricing (free Developer tier, Plus, custom Enterprise including self-hosted and BYOC; no price numbers, direct-fetched 200 on 2026-11-06)	https://www.langchain.com/pricing

官方口径与未核到项Official record and unverified items

本站尚未完成实测，结论只覆盖能力面与许可面。

实测建议（三条，都指向官方可读页没给的东西）：

judge 凭据归谁：评测器调用产生的模型费用落在厂商侧还是客户账户
在线评测的端到端延迟与漏评率：同一条 trace 从产生到出分要多久
自托管的真实运维量：Kubernetes 集群、外部存储与 license key 轮换的组合成本

This site has not completed measured runs; conclusions cover the capability and license surfaces only.

Suggested measurements (three, each pointing at what the readable official pages do not give):

Who owns the judge credentials: whether evaluator model calls bill to the vendor side or to your account
End-to-end latency and miss rate of online evaluation: how long one trace takes from ingest to score
The real operational load of self-hosting: Kubernetes cluster, external storage and license-key rotation combined

变更记录说明

官方 changelog 当日直抓在册，最新条目日期 2026-11-02，条目覆盖评测、数据集、标注队列与在线评测器的权限面。旧域名 docs.smith.langchain.com 当日 308 跳转新域名，档案一律按新域名挂源。