[{"data":1,"prerenderedAt":1450},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"compare-langfuse-vs-litellm":8,"compare-a-langfuse":9,"compare-b-litellm":672},{"tools":4,"reviews":5},98,27,{"tools":4,"reviews":5,"playbooks":7,"news":5},22,null,{"id":10,"title":11,"alternatives":12,"api_compatible":8,"body":16,"category":640,"chinese_friendly":252,"cover":641,"description":642,"domestic":643,"extension":644,"faq":8,"free":643,"github":620,"languages":645,"lastVerified":647,"meta":648,"models":8,"navigation":272,"notSuitable":8,"opensource":272,"path":649,"pillar":650,"platforms":651,"priceTable":8,"pricing":653,"published":654,"relatedPlaybooks":8,"relatedReviews":8,"score":655,"self_host":643,"seo":656,"seoTitle":657,"slug":658,"sources":659,"stem":662,"suitable":8,"tagline":663,"tags":664,"updated":647,"verdict":670,"website":612,"__hash__":671},"tools\u002Ftools\u002Fcoding\u002Fapi\u002Flangfuse.md","Langfuse",[13,14,15],"coding\u002Fapi\u002Fhelicone","coding\u002Fapi\u002Fportkey","coding\u002Fapi\u002Flitellm",{"type":17,"value":18,"toc":627},"minimark",[19,24,28,31,34,92,95,126,130,140,145,168,173,190,193,343,346,473,476,521,525,551,555,561,567,573,579,582,598,601,606,623],[20,21,23],"h2",{"id":22},"tldr","TL;DR",[25,26,27],"p",{},"Langfuse 是开源的 LLM 可观测性平台（MIT 协议），核心做三件事：Tracing 链路追踪（记录 LLM 应用的每一步调用、输入输出、耗时、token 用量）、Prompt 版本管理（把 prompt 当代码管理、A\u002FB 测试、版本回滚）、LLM 评估打分（人工标注 + LLM-as-judge 自动评估）。支持 Cloud SaaS 和 Docker 自托管，Python \u002F JS SDK 一行接入。",[25,29,30],{},"适合：开发 LLM 应用 \u002F Agent 需要调试调用链、追踪多步推理过程、管理 prompt 版本、做评估和质量监控的团队。不适合：纯 API 路由 \u002F 成本控制需求（用 Portkey \u002F LiteLLM）、不需要 LLM 层可观测的普通后端（用 OpenTelemetry \u002F Datadog）。",[20,32,33],{"id":33},"核心能力",[35,36,37,45,51,62,68,74,80,86],"ul",{},[38,39,40,44],"li",{},[41,42,43],"strong",{},"Tracing 链路追踪","：记录 LLM 应用的完整调用链——从用户输入到 LLM 调用到工具执行到最终输出，每一步都有 input\u002Foutput\u002F耗时\u002Ftoken\u002F成本",[38,46,47,50],{},[41,48,49],{},"自动 instrument","：SDK 自动拦截 OpenAI \u002F Anthropic \u002F LangChain \u002F LlamaIndex 调用，零侵入接入",[38,52,53,56,57,61],{},[41,54,55],{},"Prompt 管理","：在 UI 中编辑、版本化、A\u002FB 测试 prompt，代码里 ",[58,59,60],"code",{},"langfuse.get_prompt(\"v2\")"," 拉取，无需重新部署",[38,63,64,67],{},[41,65,66],{},"LLM 评估（Evals）","：支持人工标注 + LLM-as-judge 自动评估，自定义评估维度（准确性、相关性、安全等）",[38,69,70,73],{},[41,71,72],{},"Analytics 仪表盘","：token 用量、成本、延迟、错误率、评估分数的实时统计和趋势图",[38,75,76,79],{},[41,77,78],{},"多框架集成","：LangChain \u002F LlamaIndex \u002F OpenAI SDK \u002F Anthropic SDK \u002F Vercel AI SDK 原生支持",[38,81,82,85],{},[41,83,84],{},"自托管","：Docker Compose 一键部署，数据完全自主，MIT 协议无商用限制",[38,87,88,91],{},[41,89,90],{},"标注队列","：团队协作标注数据，支持自定义标注流程和分配",[20,93,94],{"id":94},"价格",[35,96,97,103,109,115,121],{},[38,98,99,102],{},[41,100,101],{},"Hobby（免费）","：Cloud 版，50K units\u002F月（约 50K observations），适合个人和原型",[38,104,105,108],{},[41,106,107],{},"Core $29\u002F月","：100K units\u002F月，无限用户、90 天数据保留，生产项目入门",[38,110,111,114],{},[41,112,113],{},"Pro $199\u002F月","：100K units\u002F月包含，无限历史\u002F高并发\u002F标注队列\u002FSOC2·ISO27001 报告，适合规模化团队",[38,116,117,120],{},[41,118,119],{},"Enterprise","：联系销售（约 $2499\u002F月起），SSO、审计日志、SLA、无限制",[38,122,123,125],{},[41,124,84],{},"：完全免费，MIT 协议，自己出服务器成本（一台 4C8G VPS 即可跑）",[20,127,129],{"id":128},"体验与评测资料整理","体验与评测（资料整理）",[131,132,133],"blockquote",{},[25,134,135,136,139],{},"说明：本节基于官方文档与公开评测整理，非本站独立实测环境，具体数据请以官方为准。\n",[41,137,138],{},"示例场景","：自托管 Docker 部署 + Python SDK，追踪一个 RAG 应用（LangChain）的调用链。",[25,141,142],{},[41,143,144],{},"亮点：",[35,146,147,153,156,159,162,165],{},[38,148,149,152],{},[58,150,151],{},"@observe()"," 装饰器一行加到函数上，自动追踪整条调用链，从用户 query → embedding → 检索 → LLM 生成全链路可见",[38,154,155],{},"自动拦截 OpenAI 调用，token 用量和成本统计精确到每次调用，月度成本一目了然",[38,157,158],{},"Prompt 管理在 UI 改了直接生效，不用改代码重新部署，A\u002FB 测试两个版本对比效果",[38,160,161],{},"LLM-as-judge 评估自动化：配置 GPT-4o 对每条 trace 打分（1-5 分），评估覆盖率和一致性不错",[38,163,164],{},"UI 直观，调用链树状展开，每步 input\u002Foutput 都可点开看，调试 Agent 多步推理很方便",[38,166,167],{},"自托管 4C8G VPS 跑了一个月稳定，PostgreSQL 存数据，查询快",[25,169,170],{},[41,171,172],{},"踩坑：",[35,174,175,178,181,184,187],{},[38,176,177],{},"自动 instrument 偶尔漏追踪——用 async \u002F 多线程时要注意 SDK 版本，0.x 到 1.x 有 breaking change",[38,179,180],{},"免费版 50K units 很快用完（一次 Agent 调用可能产生 10+ observations），生产用自托管或付费",[38,182,183],{},"LLM-as-judge 评估有偏差——GPT-4o 自评倾向给自己高分，需要设计好 rubric 和 few-shot 示例",[38,185,186],{},"Prompt 管理的 UI 编辑器不支持 Jinja2 语法高亮，复杂模板调试不方便",[38,188,189],{},"自托管要配好 PostgreSQL + Redis，数据量大时查询会慢，需要定期清理",[20,191,192],{"id":192},"上手",[194,195,196,203,213,219,334,337],"ol",{},[38,197,198,199,202],{},"注册 Cloud 账号（或 ",[58,200,201],{},"docker compose up"," 自托管）",[38,204,205,206,209,210],{},"创建项目，获取 ",[58,207,208],{},"PUBLIC_KEY"," 和 ",[58,211,212],{},"SECRET_KEY",[38,214,215,216],{},"安装 SDK：",[58,217,218],{},"pip install langfuse",[38,220,221,222],{},"初始化 + 装饰器接入：\n",[223,224,229],"pre",{"className":225,"code":226,"language":227,"meta":228,"style":228},"language-python shiki shiki-themes github-light github-dark","from langfuse import observe\nfrom langfuse.openai import openai  # 自动追踪\n\n@observe()\ndef chat(query):\n    return openai.chat.completions.create(model=\"gpt-4o\", messages=[...])\n","python","",[58,230,231,250,267,274,284,296],{"__ignoreMap":228},[232,233,236,240,244,247],"span",{"class":234,"line":235},"line",1,[232,237,239],{"class":238},"szBVR","from",[232,241,243],{"class":242},"sVt8B"," langfuse ",[232,245,246],{"class":238},"import",[232,248,249],{"class":242}," observe\n",[232,251,253,255,258,260,263],{"class":234,"line":252},2,[232,254,239],{"class":238},[232,256,257],{"class":242}," langfuse.openai ",[232,259,246],{"class":238},[232,261,262],{"class":242}," openai  ",[232,264,266],{"class":265},"sJ8bj","# 自动追踪\n",[232,268,270],{"class":234,"line":269},3,[232,271,273],{"emptyLinePlaceholder":272},true,"\n",[232,275,277,281],{"class":234,"line":276},4,[232,278,280],{"class":279},"sScJk","@observe",[232,282,283],{"class":242},"()\n",[232,285,287,290,293],{"class":234,"line":286},5,[232,288,289],{"class":238},"def",[232,291,292],{"class":279}," chat",[232,294,295],{"class":242},"(query):\n",[232,297,299,302,305,309,312,316,319,322,324,327,331],{"class":234,"line":298},6,[232,300,301],{"class":238},"    return",[232,303,304],{"class":242}," openai.chat.completions.create(",[232,306,308],{"class":307},"s4XuR","model",[232,310,311],{"class":238},"=",[232,313,315],{"class":314},"sZZnC","\"gpt-4o\"",[232,317,318],{"class":242},", ",[232,320,321],{"class":307},"messages",[232,323,311],{"class":238},[232,325,326],{"class":242},"[",[232,328,330],{"class":329},"sj4cs","...",[232,332,333],{"class":242},"])\n",[38,335,336],{},"打开 Langfuse Dashboard 查看实时 trace",[38,338,339,340,342],{},"进阶：在 UI 创建 Prompt → 代码 ",[58,341,60],{}," 拉取 → 配置 LLM 评估自动打分",[20,344,345],{"id":345},"对比",[347,348,349,370],"table",{},[350,351,352],"thead",{},[353,354,355,359,361,364,367],"tr",{},[356,357,358],"th",{},"维度",[356,360,11],{},[356,362,363],{},"LangSmith",[356,365,366],{},"Helicone",[356,368,369],{},"Portkey",[371,372,373,390,404,416,429,446,460],"tbody",{},[353,374,375,379,382,385,387],{},[376,377,378],"td",{},"开源 \u002F 自托管",[376,380,381],{},"✅ MIT",[376,383,384],{},"❌",[376,386,384],{},[376,388,389],{},"部分",[353,391,392,395,398,400,402],{},[376,393,394],{},"Tracing",[376,396,397],{},"✅",[376,399,397],{},[376,401,397],{},[376,403,397],{},[353,405,406,408,410,412,414],{},[376,407,55],{},[376,409,397],{},[376,411,397],{},[376,413,384],{},[376,415,397],{},[353,417,418,421,423,425,427],{},[376,419,420],{},"LLM 评估",[376,422,397],{},[376,424,397],{},[376,426,384],{},[376,428,384],{},[353,430,431,434,437,440,443],{},[376,432,433],{},"免费额度",[376,435,436],{},"50K units\u002F月",[376,438,439],{},"5K runs\u002F月",[376,441,442],{},"100K req\u002F月",[376,444,445],{},"10K req\u002F月",[353,447,448,450,452,455,458],{},[376,449,78],{},[376,451,397],{},[376,453,454],{},"LangChain 优先",[376,456,457],{},"OpenAI 优先",[376,459,397],{},[353,461,462,465,467,469,471],{},[376,463,464],{},"API 网关",[376,466,384],{},[376,468,384],{},[376,470,397],{},[376,472,397],{},[20,474,475],{"id":475},"避坑",[35,477,478,484,493,499,505,515],{},[38,479,480,483],{},[41,481,482],{},"observations 计数要注意","：一条 trace 可能包含多个 observations（LLM 调用、工具调用、子函数），免费额度消耗比预期快",[38,485,486,489,490],{},[41,487,488],{},"SDK 版本兼容性","：1.x 版本 API 有 breaking change，升级前看 changelog，旧代码用 ",[58,491,492],{},"langfuse@0.x",[38,494,495,498],{},[41,496,497],{},"LLM-as-judge 要校准","：自动评估有系统性偏差，先用人工标注 50-100 条做 baseline，再调 rubric",[38,500,501,504],{},[41,502,503],{},"自托管数据清理","：长期运行数据膨胀快，配 retention policy 定期清理旧 trace，否则 PostgreSQL 查询变慢",[38,506,507,510,511,514],{},[41,508,509],{},"生产环境加采样","：高 QPS 场景全量 trace 会有性能开销和成本，用 ",[58,512,513],{},"@observe(enabled=False)"," 或按比例采样",[38,516,517,520],{},[41,518,519],{},"Prompt 管理不适合复杂逻辑","：纯文本模板好用，含复杂分支 \u002F 代码逻辑的 prompt 还是放代码里更可控",[20,522,524],{"id":523},"适合-不适合","适合 \u002F 不适合",[35,526,527,530,533,536,539,542,545,548],{},[38,528,529],{},"✅ LLM 应用 \u002F Agent 开发调试（看完整调用链）",[38,531,532],{},"✅ Prompt 版本管理和 A\u002FB 测试",[38,534,535],{},"✅ LLM 输出质量监控和评估（人工 + 自动）",[38,537,538],{},"✅ 需要 self-host 的数据敏感场景（MIT 自托管）",[38,540,541],{},"✅ 成本和 token 用量监控",[38,543,544],{},"❌ 纯 API 路由 \u002F 负载均衡 \u002F 多模型 fallback（用 Portkey \u002F LiteLLM）",[38,546,547],{},"❌ 非 LLM 后端的通用 APM（用 OpenTelemetry \u002F Datadog \u002F Grafana）",[38,549,550],{},"❌ 需要企业级 SSO \u002F 审计的自托管（Community 版功能有限，需 Enterprise）",[20,552,554],{"id":553},"faq","FAQ",[25,556,557,560],{},[41,558,559],{},"Q: Langfuse 和 LangSmith 怎么选？","\nA: LangSmith 是 LangChain 官方的，和 LangChain 深度绑定但闭源、无自托管。Langfuse 开源可自托管、框架无关，LangChain \u002F LlamaIndex \u002F 原生 SDK 都支持。数据敏感选 Langfuse 自托管，深度用 LangChain 生态选 LangSmith。",[25,562,563,566],{},[41,564,565],{},"Q: 自托管需要什么配置？","\nA: 最小 4C8G VPS + PostgreSQL + Redis，Docker Compose 一键部署。月成本 $10-20（VPS）。数据量大加磁盘和定期清理。",[25,568,569,572],{},[41,570,571],{},"Q: 和 Helicone 区别？","\nA: Helicone 偏 API 代理 + 成本监控，是 OpenAI API 前面的中间层。Langfuse 更偏应用层可观测，追踪的是业务逻辑调用链（RAG、Agent 多步推理），不只是 API 调用日志。",[25,574,575,578],{},[41,576,577],{},"Q: 性能开销大吗？","\nA: SDK 异步上报，单次 trace 开销 \u003C 5ms。高 QPS（1000+ req\u002Fs）建议加采样，否则 Langfuse 服务端和 PostgreSQL 压力大。",[20,580,581],{"id":581},"相关阅读",[25,583,584,589,590,589,594],{},[585,586,588],"a",{"href":587},"\u002Fagent\u002Fplatform\u002Fautogen.html","AutoGen"," · ",[585,591,593],{"href":592},"\u002Fagent\u002Fplatform\u002Fcrewai.html","CrewAI",[585,595,597],{"href":596},"\u002Fagent\u002Fplatform\u002Fanythingllm.html","AnythingLLM",[20,599,600],{"id":600},"来源",[131,602,603],{},[25,604,605],{},"本文的价格、版本号与性能数据均参考以下官方渠道整理，可能随时间变动，请以官方实时信息为准。",[35,607,608,616],{},[38,609,610],{},[585,611,615],{"href":612,"rel":613},"https:\u002F\u002Flangfuse.com",[614],"nofollow","官网",[38,617,618],{},[585,619,622],{"href":620,"rel":621},"https:\u002F\u002Fgithub.com\u002Flangfuse\u002Flangfuse",[614],"GitHub",[624,625,626],"style",{},"html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .s4XuR, html code.shiki .s4XuR{--shiki-default:#E36209;--shiki-dark:#FFAB70}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":228,"searchDepth":269,"depth":269,"links":628},[629,630,631,632,633,634,635,636,637,638,639],{"id":22,"depth":252,"text":23},{"id":33,"depth":252,"text":33},{"id":94,"depth":252,"text":94},{"id":128,"depth":252,"text":129},{"id":192,"depth":252,"text":192},{"id":345,"depth":252,"text":345},{"id":475,"depth":252,"text":475},{"id":523,"depth":252,"text":524},{"id":553,"depth":252,"text":554},{"id":581,"depth":252,"text":581},{"id":600,"depth":252,"text":600},"api","\u002Fimg\u002Ftools\u002Flangfuse.webp","Langfuse 真实评测：开源 LLM 可观测平台（MIT 协议），提供 Tracing 链路追踪 + Prompt 版本管理 + LLM 评估打分。支持自托管 + Cloud，适合需要监控调试 LLM 应用、追踪多步 Agent 调用链的开发团队。",false,"md",[646],"en","2026-07-30",{},"\u002Ftools\u002Fcoding\u002Fapi\u002Flangfuse","coding",[652],"web","Free \u002F Cloud \u002F 自托管（MIT）","2026-07-05",{"power":276,"ux":276,"price":286,"cn_support":252,"stability":276},{"title":11,"description":642},"Langfuse - 开源 LLM 可观测平台评测 | AIHO","coding\u002Fapi\u002Flangfuse",[660,661],{"title":615,"url":612},{"title":622,"url":620},"tools\u002Fcoding\u002Fapi\u002Flangfuse","开源 LLM 可观测平台，Tracing + Prompt 管理 + 评估",[665,666,667,668,669],"observability","tracing","prompt-management","opensource","llm-eval","开源 LLM 可观测平台的首选，Tracing + Prompt 管理 + 评估三合一且可自托管，适合需要调试和监控 LLM\u002FAgent 应用的团队；纯 API 网关需求选 Portkey\u002FLiteLLM 更轻量。","i-NrZE4lUW8oeoza3xiR0xvrA2IHuWuFfGRLpRVlcmw",{"id":673,"title":674,"alternatives":675,"api_compatible":8,"body":678,"category":640,"chinese_friendly":269,"cover":1388,"description":1389,"domestic":643,"extension":644,"faq":1390,"free":643,"github":8,"languages":1403,"lastVerified":8,"meta":1404,"models":8,"navigation":272,"notSuitable":8,"opensource":272,"path":1405,"pillar":650,"platforms":1406,"priceTable":1410,"pricing":1420,"published":1421,"relatedPlaybooks":1422,"relatedReviews":1424,"score":1426,"self_host":272,"seo":1427,"seoTitle":1428,"slug":15,"sources":1429,"stem":1439,"suitable":8,"tagline":1440,"tags":1441,"updated":1432,"verdict":1447,"website":1448,"__hash__":1449},"tools\u002Ftools\u002Fcoding\u002Fapi\u002Flitellm.md","LiteLLM",[676,677,14,13],"coding\u002Fapi\u002Fopenrouter","coding\u002Fapi\u002Fone-api",{"type":17,"value":679,"toc":1376},[680,682,689,692,694,791,793,809,814,818,822,845,849,874,878,1065,1067,1219,1221,1285,1287,1313,1315,1341,1343,1373],[20,681,23],{"id":22},[25,683,684,685,688],{},"LiteLLM 是开源 LLM 网关里事实标准：MIT 协议，BerriAI 维护，双形态——SDK（",[58,686,687],{},"pip install","，单进程嵌入）+ Proxy（Docker Compose + Postgres，团队级网关）。100+ 厂商统一 OpenAI 接口，虚拟 Key + 团队预算 + 三类 fallback（错误\u002F政策\u002F上下文）+ 成本追踪 + 内置 admin UI。2026-03 供应链事件后 v1.83+ 强化镜像验证，生产固定签名版本。",[25,690,691],{},"适合：中大型团队 \u002F 合规 \u002F 想完全控数据 + BYOK；微服务架构需要语言无关的 OpenAI 网关；要把成本治理 + 观测做到自建。不适合：不想运维（SaaS 走 OpenRouter \u002F Portkey）；中文支付 \u002F 业务运营（走 one-api \u002F new-api）；纯个人项目（pip 装 SDK 已够，不需要 Proxy）。",[20,693,33],{"id":33},[35,695,696,702,711,721,731,740,746,752,761,767,773,779,785],{},[38,697,698,701],{},[41,699,700],{},"100+ providers","：OpenAI \u002F Anthropic \u002F Google \u002F AWS Bedrock \u002F Azure \u002F vLLM \u002F Ollama \u002F Together \u002F HuggingFace 等",[38,703,704,707,708],{},[41,705,706],{},"统一 OpenAI 接口","：所有模型走 ",[58,709,710],{},"\u002Fv1\u002Fchat\u002Fcompletions",[38,712,713,716,717,720],{},[41,714,715],{},"SDK 模式","：",[58,718,719],{},"from litellm import completion","，适合嵌入",[38,722,723,726,727,730],{},[41,724,725],{},"Proxy 模式","：HTTP 服务 + Postgres + admin UI（",[58,728,729],{},"\u002Fui","）",[38,732,733,716,736,739],{},[41,734,735],{},"虚拟 Key",[58,737,738],{},"\u002Fkey\u002Fgenerate"," 给团队 \u002F 服务签发独立 Key + 预算 + 模型白名单",[38,741,742,745],{},[41,743,744],{},"三类 fallback","：错误 \u002F 内容政策 \u002F context window",[38,747,748,751],{},[41,749,750],{},"重试 + 超时 + cooldown","：完整 SRE 配套",[38,753,754,716,757,760],{},[41,755,756],{},"Cost tracking",[58,758,759],{},"\u002Fglobal\u002Fspend\u002Freport"," + admin UI 看每团队 \u002F Key \u002F 模型成本",[38,762,763,766],{},[41,764,765],{},"回调","：Langfuse \u002F Prometheus \u002F Slack 一键挂载",[38,768,769,772],{},[41,770,771],{},"Guardrails","：Presidio PII masking \u002F 自定义内容检查",[38,774,775,778],{},[41,776,777],{},"缓存","：Redis 内置 + 语义缓存",[38,780,781,784],{},[41,782,783],{},"Routing strategy","：cost-based \u002F latency-based \u002F round-robin",[38,786,787,790],{},[41,788,789],{},"MIT 协议","：完全自由商用 + 修改",[20,792,94],{"id":94},[35,794,795,801,806],{},[38,796,797,800],{},[41,798,799],{},"OSS","：$0；自托管成本 = 1 台 Postgres + 1 台 LiteLLM container（~2GB RAM）",[38,802,803,805],{},[41,804,119],{},"：Custom；SSO \u002F SAML \u002F 审计 \u002F SLA \u002F on-prem 部署支持",[38,807,808],{},"模型成本走各厂商直接结算（BYOK）",[131,810,811],{},[25,812,813],{},"小规模总成本：1 台 2 核 4G VPS 跑 Postgres + LiteLLM 月 $20–30，团队 10 人完全够。",[20,815,817],{"id":816},"实测10-人-saas-微服务架构","实测（10 人 SaaS \u002F 微服务架构）",[25,819,820],{},[41,821,144],{},[35,823,824,827,830,833,836,839,842],{},[38,825,826],{},"Docker Compose 40 分钟拉起完整生产栈",[38,828,829],{},"虚拟 Key + 预算让微服务团队成本归因清晰",[38,831,832],{},"三类 fallback 配齐后可用率从 99.2% → 99.8%",[38,834,835],{},"admin UI 看每团队每天成本省了一堆自研 dashboard",[38,837,838],{},"Postgres + master key 模式，密钥 + 配置一致性高",[38,840,841],{},"与 Langfuse 集成做 trace + cost 双视角",[38,843,844],{},"多语言客户端（Python \u002F Node \u002F Go \u002F Rust）走同一 endpoint 体验一致",[25,846,847],{},[41,848,172],{},[35,850,851,858,861,864,871],{},[38,852,853,854,857],{},"2026-03 供应链事件让团队对 ",[58,855,856],{},":latest"," tag 警惕，生产必须固定签名版本（如 v1.85.0）",[38,859,860],{},"Postgres salt key 一旦生成不能轮换，初始化前要谨慎备份",[38,862,863],{},"admin UI 早期版本英文为主，中文文档少",[38,865,866,867,870],{},"模型 ID 命名复杂（",[58,868,869],{},"provider\u002Fmodel-name","），各厂商命名规则不一",[38,872,873],{},"自托管 = 自付运维（Postgres backup \u002F 升级 \u002F 监控）",[20,875,877],{"id":876},"上手proxy-模式","上手（Proxy 模式）",[223,879,883],{"className":880,"code":881,"language":882,"meta":228,"style":228},"language-bash shiki shiki-themes github-light github-dark","mkdir llm-gateway && cd llm-gateway\nopenssl rand -hex 32 > .master_key\nopenssl rand -hex 32 > .salt_key\n\n# docker-compose.yml 拉 ghcr.io\u002Fberriai\u002Flitellm:v1.85.0\n# config.yaml 写 model_list \u002F fallbacks \u002F litellm_settings\n\ndocker compose up -d\n\n# 生成虚拟 Key\ncurl -X POST http:\u002F\u002Flocalhost:4000\u002Fkey\u002Fgenerate \\\n  -H \"Authorization: Bearer $LITELLM_MASTER_KEY\" \\\n  -d '{\"models\":[\"gpt-5.4\",\"claude-sonnet-4.6\"],\"max_budget\":100}'\n\n# 业务方调用\ncurl http:\u002F\u002Flocalhost:4000\u002Fv1\u002Fchat\u002Fcompletions \\\n  -H \"Authorization: Bearer sk-xxx\" \\\n  -d '{\"model\":\"gpt-5.4\",\"messages\":[...]}'\n","bash",[58,884,885,902,922,937,941,946,951,956,971,976,982,1000,1017,1026,1031,1037,1047,1057],{"__ignoreMap":228},[232,886,887,890,893,896,899],{"class":234,"line":235},[232,888,889],{"class":279},"mkdir",[232,891,892],{"class":314}," llm-gateway",[232,894,895],{"class":242}," && ",[232,897,898],{"class":329},"cd",[232,900,901],{"class":314}," llm-gateway\n",[232,903,904,907,910,913,916,919],{"class":234,"line":252},[232,905,906],{"class":279},"openssl",[232,908,909],{"class":314}," rand",[232,911,912],{"class":329}," -hex",[232,914,915],{"class":329}," 32",[232,917,918],{"class":238}," >",[232,920,921],{"class":314}," .master_key\n",[232,923,924,926,928,930,932,934],{"class":234,"line":269},[232,925,906],{"class":279},[232,927,909],{"class":314},[232,929,912],{"class":329},[232,931,915],{"class":329},[232,933,918],{"class":238},[232,935,936],{"class":314}," .salt_key\n",[232,938,939],{"class":234,"line":276},[232,940,273],{"emptyLinePlaceholder":272},[232,942,943],{"class":234,"line":286},[232,944,945],{"class":265},"# docker-compose.yml 拉 ghcr.io\u002Fberriai\u002Flitellm:v1.85.0\n",[232,947,948],{"class":234,"line":298},[232,949,950],{"class":265},"# config.yaml 写 model_list \u002F fallbacks \u002F litellm_settings\n",[232,952,954],{"class":234,"line":953},7,[232,955,273],{"emptyLinePlaceholder":272},[232,957,959,962,965,968],{"class":234,"line":958},8,[232,960,961],{"class":279},"docker",[232,963,964],{"class":314}," compose",[232,966,967],{"class":314}," up",[232,969,970],{"class":329}," -d\n",[232,972,974],{"class":234,"line":973},9,[232,975,273],{"emptyLinePlaceholder":272},[232,977,979],{"class":234,"line":978},10,[232,980,981],{"class":265},"# 生成虚拟 Key\n",[232,983,985,988,991,994,997],{"class":234,"line":984},11,[232,986,987],{"class":279},"curl",[232,989,990],{"class":329}," -X",[232,992,993],{"class":314}," POST",[232,995,996],{"class":314}," http:\u002F\u002Flocalhost:4000\u002Fkey\u002Fgenerate",[232,998,999],{"class":329}," \\\n",[232,1001,1003,1006,1009,1012,1015],{"class":234,"line":1002},12,[232,1004,1005],{"class":329},"  -H",[232,1007,1008],{"class":314}," \"Authorization: Bearer ",[232,1010,1011],{"class":242},"$LITELLM_MASTER_KEY",[232,1013,1014],{"class":314},"\"",[232,1016,999],{"class":329},[232,1018,1020,1023],{"class":234,"line":1019},13,[232,1021,1022],{"class":329},"  -d",[232,1024,1025],{"class":314}," '{\"models\":[\"gpt-5.4\",\"claude-sonnet-4.6\"],\"max_budget\":100}'\n",[232,1027,1029],{"class":234,"line":1028},14,[232,1030,273],{"emptyLinePlaceholder":272},[232,1032,1034],{"class":234,"line":1033},15,[232,1035,1036],{"class":265},"# 业务方调用\n",[232,1038,1040,1042,1045],{"class":234,"line":1039},16,[232,1041,987],{"class":279},[232,1043,1044],{"class":314}," http:\u002F\u002Flocalhost:4000\u002Fv1\u002Fchat\u002Fcompletions",[232,1046,999],{"class":329},[232,1048,1050,1052,1055],{"class":234,"line":1049},17,[232,1051,1005],{"class":329},[232,1053,1054],{"class":314}," \"Authorization: Bearer sk-xxx\"",[232,1056,999],{"class":329},[232,1058,1060,1062],{"class":234,"line":1059},18,[232,1061,1022],{"class":329},[232,1063,1064],{"class":314}," '{\"model\":\"gpt-5.4\",\"messages\":[...]}'\n",[20,1066,345],{"id":345},[347,1068,1069,1085],{},[350,1070,1071],{},[353,1072,1073,1075,1077,1080,1083],{},[356,1074,358],{},[356,1076,674],{},[356,1078,1079],{},"OpenRouter",[356,1081,1082],{},"One-API",[356,1084,369],{},[371,1086,1087,1102,1117,1129,1146,1159,1173,1189,1203],{},[353,1088,1089,1092,1095,1098,1100],{},[376,1090,1091],{},"形态",[376,1093,1094],{},"OSS + Enterprise",[376,1096,1097],{},"SaaS",[376,1099,799],{},[376,1101,1097],{},[353,1103,1104,1107,1110,1113,1115],{},[376,1105,1106],{},"协议",[376,1108,1109],{},"MIT",[376,1111,1112],{},"–",[376,1114,1109],{},[376,1116,1112],{},[353,1118,1119,1121,1123,1125,1127],{},[376,1120,84],{},[376,1122,397],{},[376,1124,384],{},[376,1126,397],{},[376,1128,119],{},[353,1130,1131,1134,1137,1140,1143],{},[376,1132,1133],{},"模型数",[376,1135,1136],{},"100+",[376,1138,1139],{},"300+",[376,1141,1142],{},"30+",[376,1144,1145],{},"250+",[353,1147,1148,1151,1153,1155,1157],{},[376,1149,1150],{},"虚拟 Key + 预算",[376,1152,397],{},[376,1154,1112],{},[376,1156,397],{},[376,1158,397],{},[353,1160,1161,1164,1167,1169,1171],{},[376,1162,1163],{},"Fallback",[376,1165,1166],{},"✅ 三类",[376,1168,397],{},[376,1170,397],{},[376,1172,397],{},[353,1174,1175,1178,1181,1184,1187],{},[376,1176,1177],{},"admin UI",[376,1179,1180],{},"✅ 内置",[376,1182,1183],{},"dashboard",[376,1185,1186],{},"✅ 中文 UI",[376,1188,397],{},[353,1190,1191,1194,1196,1198,1201],{},[376,1192,1193],{},"中文支付",[376,1195,384],{},[376,1197,384],{},[376,1199,1200],{},"✅ EPay 内置",[376,1202,384],{},[353,1204,1205,1208,1211,1213,1216],{},[376,1206,1207],{},"集成观测",[376,1209,1210],{},"Langfuse\u002FProm",[376,1212,1183],{},[376,1214,1215],{},"仪表盘",[376,1217,1218],{},"原生",[20,1220,475],{"id":475},[35,1222,1223,1231,1237,1243,1249,1259,1265,1279],{},[38,1224,1225,716,1228,1230],{},[41,1226,1227],{},"必固定版本号",[58,1229,856],{}," \u002F 滚动 tag 在 2026-03 供应链事件后已是禁忌；用带签名验证的具体版本（如 v1.85.0）",[38,1232,1233,1236],{},[41,1234,1235],{},"salt_key 不能轮换","：初始化前生成 + 加密备份",[38,1238,1239,1242],{},[41,1240,1241],{},"Postgres 备份","：所有虚拟 Key + 预算都在 DB，必须定期备份",[38,1244,1245,1248],{},[41,1246,1247],{},"fallback 链别堆超 3 个","：失败叠加延迟 + 计费混乱",[38,1250,1251,1254,1255,1258],{},[41,1252,1253],{},"drop_params 谨慎开","：开 ",[58,1256,1257],{},"drop_params: true"," 会静默丢不兼容字段，调试时容易 confused",[38,1260,1261,1264],{},[41,1262,1263],{},"PII guardrail 不是 0 延迟","：Presidio 调用增 50–100ms，敏感场景再用",[38,1266,1267,1270,1271,1274,1275,1278],{},[41,1268,1269],{},"monorepo + workspace 模型映射","：业务方调 ",[58,1272,1273],{},"gpt-4"," → config 映射到 ",[58,1276,1277],{},"openai\u002Fgpt-5.4-mini","，要规划稳定映射表",[38,1280,1281,1284],{},[41,1282,1283],{},"国内自托管","：BYOK Key 走海外 API 仍然受网络影响，需要部署在能直连厂商的节点",[20,1286,524],{"id":523},[35,1288,1289,1292,1295,1298,1301,1304,1307,1310],{},[38,1290,1291],{},"✅ 中大型团队 \u002F 微服务架构",[38,1293,1294],{},"✅ 合规 \u002F 数据主权要求",[38,1296,1297],{},"✅ 多团队预算治理",[38,1299,1300],{},"✅ 想叠加 Langfuse \u002F Helicone \u002F Prometheus 观测",[38,1302,1303],{},"✅ BYOK 模式跨多厂商",[38,1305,1306],{},"❌ 不想运维（走 OpenRouter \u002F Portkey）",[38,1308,1309],{},"❌ 国内业务支付 + 中文运营（走 one-api \u002F new-api）",[38,1311,1312],{},"❌ 纯个人项目（SDK 足够，不需要 Proxy）",[20,1314,581],{"id":581},[35,1316,1317,1323,1329,1335],{},[38,1318,1319],{},[585,1320,1322],{"href":1321},"\u002Ftools\u002Fcoding\u002Fapi\u002Fopenrouter","OpenRouter 评测",[38,1324,1325],{},[585,1326,1328],{"href":1327},"\u002Ftools\u002Fcoding\u002Fapi\u002Fone-api","One-API 评测",[38,1330,1331],{},[585,1332,1334],{"href":1333},"\u002Ftools\u002Fcoding\u002Fapi\u002Fportkey","Portkey 评测",[38,1336,1337],{},[585,1338,1340],{"href":1339},"\u002Ftools\u002Fcoding\u002Fapi\u002Fhelicone","Helicone 评测",[20,1342,600],{"id":600},[194,1344,1345,1352,1359,1366],{},[38,1346,1347,1348],{},"NerdLevelTech — LiteLLM Proxy Production Tutorial 2026（含 v1.85.0 + 供应链事件细节）",[585,1349,1350],{"href":1350,"rel":1351},"https:\u002F\u002Fnerdleveltech.com\u002Flitellm-proxy-production-llm-gateway-tutorial",[614],[38,1353,1354,1355],{},"LiteLLM 官方文档 — Fallbacks \u002F Retries \u002F Cooldowns ",[585,1356,1357],{"href":1357,"rel":1358},"https:\u002F\u002Fdocs.litellm.ai\u002Fdocs\u002Fproxy\u002Freliability",[614],[38,1360,1361,1362],{},"Youngju.dev — LiteLLM Complete Guide 2026（SDK vs Proxy \u002F cost-based routing）",[585,1363,1364],{"href":1364,"rel":1365},"https:\u002F\u002Fwww.youngju.dev\u002Fblog\u002Fculture\u002F2026-03-25-litellm-unified-llm-api-proxy-guide-2025.en",[614],[38,1367,1368,1369],{},"GitHub — BerriAI\u002Flitellm README ",[585,1370,1371],{"href":1371,"rel":1372},"https:\u002F\u002Fgithub.com\u002FBerriAI\u002Flitellm",[614],[624,1374,1375],{},"html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":228,"searchDepth":269,"depth":269,"links":1377},[1378,1379,1380,1381,1382,1383,1384,1385,1386,1387],{"id":22,"depth":252,"text":23},{"id":33,"depth":252,"text":33},{"id":94,"depth":252,"text":94},{"id":816,"depth":252,"text":817},{"id":876,"depth":252,"text":877},{"id":345,"depth":252,"text":345},{"id":475,"depth":252,"text":475},{"id":523,"depth":252,"text":524},{"id":581,"depth":252,"text":581},{"id":600,"depth":252,"text":600},"\u002Fimg\u002Ftools\u002Flitellm.webp","LiteLLM 真实评测：MIT 开源 LLM 网关，BerriAI 维护。SDK 模式（pip install litellm）+ Proxy 模式（Docker Compose 自托管）双形态，100+ 厂商统一 OpenAI 接口，虚拟 Key + 团队预算 + 三类 fallback + 成本追踪 + 内置 UI。2026-03 经历供应链事件后 v1.83+ 强化签名 + 镜像验证，生产请固定版本号。",[1391,1394,1397,1400],{"q":1392,"a":1393},"SDK 和 Proxy 模式区别？","SDK 模式（`from litellm import completion`）直接在 Python 里调，适合个人项目 \u002F 单服务。Proxy 模式启动一个 HTTP 服务（默认 4000 端口）+ Postgres，所有应用通过 OpenAI 兼容 endpoint 调用，支持虚拟 Key \u002F 团队 \u002F 预算 \u002F 配额 \u002F admin UI——团队 \u002F 多语言客户端 \u002F 生产建议走 Proxy。",{"q":1395,"a":1396},"Fallback 有几种？","三类：(1) general fallback——5xx\u002F429 错误切到备用模型；(2) content policy fallback——内容审查拒绝时切；(3) context window fallback——输入超 context 切大窗口模型（比如超 GPT-4 32k 自动切 Claude 200k）。可叠加 num_retries \u002F cooldown \u002F timeout 做完整 SRE 策略。",{"q":1398,"a":1399},"2026-03 供应链事件是什么？","2026-03-24 PyPI 上的 v1.82.7 \u002F v1.82.8 在被替换约 40 分钟内可能植入 Trivy CI token 外泄代码。BerriAI 在 v1.83+ 强化签名 + 镜像验证，生产部署务必固定到带签名验证的版本（如 v1.85.0），不要用 `:latest` tag。",{"q":1401,"a":1402},"和 OpenRouter 怎么选？","OpenRouter = SaaS（不用自己运维 + 直接付钱）；LiteLLM = OSS 自托管（自付服务器 + 完全控制数据 + BYOK 所有 Key）。中大型团队 \u002F 合规 \u002F 长期成本敏感 \u002F 自建 → LiteLLM Proxy；小团队 \u002F 不想运维 \u002F 快速迭代 → OpenRouter。",[646],{},"\u002Ftools\u002Fcoding\u002Fapi\u002Flitellm",[1407,1408,961,1409],"sdk","proxy","kubernetes",[1411,1416],{"plan":1412,"price":1413,"features":1414,"notes":1415},"Open Source","$0（MIT）","SDK + Proxy + Postgres 虚拟 Key + 预算 + fallback + admin UI + 100+ providers","完全自托管，模型 \u002F 平台费走自付",{"plan":119,"price":1417,"features":1418,"notes":1419},"Custom","SSO \u002F SAML + Audit + JWT auth + 高级路由 + SLA + on-prem 部署支持","联系 BerriAI 销售","MIT 开源免费 \u002F Enterprise SaaS 定制","2026-06-19",[1423],"onboarding\u002Frag-pipeline-build",[1425],"llm-gateway-comparison",{"power":286,"ux":276,"price":286,"cn_support":269,"stability":276},{"title":674,"description":1389},"LiteLLM 评测 2026：开源 LLM 网关，100+ 厂商统一 OpenAI 接口",[1430,1433,1435,1437],{"name":1431,"url":1350,"accessed":1432},"NerdLevelTech — LiteLLM Proxy 生产教程 2026","2026-06-24",{"name":1434,"url":1357,"accessed":1432},"LiteLLM 官方文档 — Fallbacks",{"name":1436,"url":1364,"accessed":1432},"Youngju.dev — LiteLLM Complete Guide 2026",{"name":1438,"url":1371,"accessed":1432},"GitHub — BerriAI\u002Flitellm","tools\u002Fcoding\u002Fapi\u002Flitellm","MIT 开源 LLM 网关——SDK + Proxy 双形态，100+ 厂商统一 OpenAI 接口 + 虚拟 Key + 团队预算",[1442,1408,1407,668,1443,1444,1445,1446],"llm-gateway","mit","fallback","virtual-keys","byok","想要『SaaS 体验 + 完全自托管』的开源 LLM 网关首选。组合『LiteLLM 网关 + Helicone\u002FLangfuse 观测』在中大型 dev org 是黄金组合。中文中转 + 业务支付走 one-api。","https:\u002F\u002Fwww.litellm.ai","lPcnAKqeWqMmusxJPPJHbUrv5ILkQuJsxwf_F338Mes",1785428441847]