[{"data":1,"prerenderedAt":1538},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"compare-litellm-vs-raga":9,"compare-a-litellm":10,"compare-b-raga":850},{"tools":4,"reviews":5},78,26,{"tools":4,"reviews":5,"playbooks":7,"news":8},22,21,null,{"id":11,"title":12,"alternatives":13,"api_compatible":9,"body":18,"category":782,"chinese_friendly":290,"cover":783,"description":784,"domestic":785,"extension":786,"faq":787,"free":785,"github":9,"languages":800,"lastVerified":9,"meta":802,"models":9,"navigation":309,"notSuitable":9,"opensource":309,"path":803,"pillar":804,"platforms":805,"priceTable":809,"pricing":819,"published":820,"relatedPlaybooks":821,"relatedReviews":823,"score":825,"self_host":309,"seo":826,"seoTitle":9,"slug":827,"sources":828,"stem":838,"suitable":9,"tagline":839,"tags":840,"updated":831,"verdict":847,"website":848,"__hash__":849},"tools\u002Ftools\u002Fcoding\u002Fapi\u002Flitellm.md","LiteLLM",[14,15,16,17],"coding\u002Fapi\u002Fopenrouter","coding\u002Fapi\u002Fone-api","coding\u002Fapi\u002Fportkey","coding\u002Fapi\u002Fhelicone",{"type":19,"value":20,"toc":770},"minimark",[21,26,35,38,41,141,144,161,167,171,176,199,204,229,233,438,441,604,607,671,675,701,704,731,734,766],[22,23,25],"h2",{"id":24},"tldr","TL;DR",[27,28,29,30,34],"p",{},"LiteLLM 是开源 LLM 网关里事实标准：MIT 协议，BerriAI 维护，双形态——SDK（",[31,32,33],"code",{},"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。",[27,36,37],{},"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）。",[22,39,40],{"id":40},"核心能力",[42,43,44,52,61,71,81,90,96,102,111,117,123,129,135],"ul",{},[45,46,47,51],"li",{},[48,49,50],"strong",{},"100+ providers","：OpenAI \u002F Anthropic \u002F Google \u002F AWS Bedrock \u002F Azure \u002F vLLM \u002F Ollama \u002F Together \u002F HuggingFace 等",[45,53,54,57,58],{},[48,55,56],{},"统一 OpenAI 接口","：所有模型走 ",[31,59,60],{},"\u002Fv1\u002Fchat\u002Fcompletions",[45,62,63,66,67,70],{},[48,64,65],{},"SDK 模式","：",[31,68,69],{},"from litellm import completion","，适合嵌入",[45,72,73,76,77,80],{},[48,74,75],{},"Proxy 模式","：HTTP 服务 + Postgres + admin UI（",[31,78,79],{},"\u002Fui","）",[45,82,83,66,86,89],{},[48,84,85],{},"虚拟 Key",[31,87,88],{},"\u002Fkey\u002Fgenerate"," 给团队 \u002F 服务签发独立 Key + 预算 + 模型白名单",[45,91,92,95],{},[48,93,94],{},"三类 fallback","：错误 \u002F 内容政策 \u002F context window",[45,97,98,101],{},[48,99,100],{},"重试 + 超时 + cooldown","：完整 SRE 配套",[45,103,104,66,107,110],{},[48,105,106],{},"Cost tracking",[31,108,109],{},"\u002Fglobal\u002Fspend\u002Freport"," + admin UI 看每团队 \u002F Key \u002F 模型成本",[45,112,113,116],{},[48,114,115],{},"回调","：Langfuse \u002F Prometheus \u002F Slack 一键挂载",[45,118,119,122],{},[48,120,121],{},"Guardrails","：Presidio PII masking \u002F 自定义内容检查",[45,124,125,128],{},[48,126,127],{},"缓存","：Redis 内置 + 语义缓存",[45,130,131,134],{},[48,132,133],{},"Routing strategy","：cost-based \u002F latency-based \u002F round-robin",[45,136,137,140],{},[48,138,139],{},"MIT 协议","：完全自由商用 + 修改",[22,142,143],{"id":143},"价格",[42,145,146,152,158],{},[45,147,148,151],{},[48,149,150],{},"OSS","：$0；自托管成本 = 1 台 Postgres + 1 台 LiteLLM container（~2GB RAM）",[45,153,154,157],{},[48,155,156],{},"Enterprise","：Custom；SSO \u002F SAML \u002F 审计 \u002F SLA \u002F on-prem 部署支持",[45,159,160],{},"模型成本走各厂商直接结算（BYOK）",[162,163,164],"blockquote",{},[27,165,166],{},"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。",[22,168,170],{"id":169},"实测10-人-saas-微服务架构","实测（10 人 SaaS \u002F 微服务架构）",[27,172,173],{},[48,174,175],{},"亮点：",[42,177,178,181,184,187,190,193,196],{},[45,179,180],{},"Docker Compose 40 分钟拉起完整生产栈",[45,182,183],{},"虚拟 Key + 预算让微服务团队成本归因清晰",[45,185,186],{},"三类 fallback 配齐后可用率从 99.2% → 99.8%",[45,188,189],{},"admin UI 看每团队每天成本省了一堆自研 dashboard",[45,191,192],{},"Postgres + master key 模式，密钥 + 配置一致性高",[45,194,195],{},"与 Langfuse 集成做 trace + cost 双视角",[45,197,198],{},"多语言客户端（Python \u002F Node \u002F Go \u002F Rust）走同一 endpoint 体验一致",[27,200,201],{},[48,202,203],{},"踩坑：",[42,205,206,213,216,219,226],{},[45,207,208,209,212],{},"2026-03 供应链事件让团队对 ",[31,210,211],{},":latest"," tag 警惕，生产必须固定签名版本（如 v1.85.0）",[45,214,215],{},"Postgres salt key 一旦生成不能轮换，初始化前要谨慎备份",[45,217,218],{},"admin UI 早期版本英文为主，中文文档少",[45,220,221,222,225],{},"模型 ID 命名复杂（",[31,223,224],{},"provider\u002Fmodel-name","），各厂商命名规则不一",[45,227,228],{},"自托管 = 自付运维（Postgres backup \u002F 升级 \u002F 监控）",[22,230,232],{"id":231},"上手proxy-模式","上手（Proxy 模式）",[234,235,240],"pre",{"className":236,"code":237,"language":238,"meta":239,"style":239},"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","",[31,241,242,266,288,304,311,318,324,329,344,349,355,373,390,399,404,410,420,430],{"__ignoreMap":239},[243,244,247,251,255,259,263],"span",{"class":245,"line":246},"line",1,[243,248,250],{"class":249},"sScJk","mkdir",[243,252,254],{"class":253},"sZZnC"," llm-gateway",[243,256,258],{"class":257},"sVt8B"," && ",[243,260,262],{"class":261},"sj4cs","cd",[243,264,265],{"class":253}," llm-gateway\n",[243,267,269,272,275,278,281,285],{"class":245,"line":268},2,[243,270,271],{"class":249},"openssl",[243,273,274],{"class":253}," rand",[243,276,277],{"class":261}," -hex",[243,279,280],{"class":261}," 32",[243,282,284],{"class":283},"szBVR"," >",[243,286,287],{"class":253}," .master_key\n",[243,289,291,293,295,297,299,301],{"class":245,"line":290},3,[243,292,271],{"class":249},[243,294,274],{"class":253},[243,296,277],{"class":261},[243,298,280],{"class":261},[243,300,284],{"class":283},[243,302,303],{"class":253}," .salt_key\n",[243,305,307],{"class":245,"line":306},4,[243,308,310],{"emptyLinePlaceholder":309},true,"\n",[243,312,314],{"class":245,"line":313},5,[243,315,317],{"class":316},"sJ8bj","# docker-compose.yml 拉 ghcr.io\u002Fberriai\u002Flitellm:v1.85.0\n",[243,319,321],{"class":245,"line":320},6,[243,322,323],{"class":316},"# config.yaml 写 model_list \u002F fallbacks \u002F litellm_settings\n",[243,325,327],{"class":245,"line":326},7,[243,328,310],{"emptyLinePlaceholder":309},[243,330,332,335,338,341],{"class":245,"line":331},8,[243,333,334],{"class":249},"docker",[243,336,337],{"class":253}," compose",[243,339,340],{"class":253}," up",[243,342,343],{"class":261}," -d\n",[243,345,347],{"class":245,"line":346},9,[243,348,310],{"emptyLinePlaceholder":309},[243,350,352],{"class":245,"line":351},10,[243,353,354],{"class":316},"# 生成虚拟 Key\n",[243,356,358,361,364,367,370],{"class":245,"line":357},11,[243,359,360],{"class":249},"curl",[243,362,363],{"class":261}," -X",[243,365,366],{"class":253}," POST",[243,368,369],{"class":253}," http:\u002F\u002Flocalhost:4000\u002Fkey\u002Fgenerate",[243,371,372],{"class":261}," \\\n",[243,374,376,379,382,385,388],{"class":245,"line":375},12,[243,377,378],{"class":261},"  -H",[243,380,381],{"class":253}," \"Authorization: Bearer ",[243,383,384],{"class":257},"$LITELLM_MASTER_KEY",[243,386,387],{"class":253},"\"",[243,389,372],{"class":261},[243,391,393,396],{"class":245,"line":392},13,[243,394,395],{"class":261},"  -d",[243,397,398],{"class":253}," '{\"models\":[\"gpt-5.4\",\"claude-sonnet-4.6\"],\"max_budget\":100}'\n",[243,400,402],{"class":245,"line":401},14,[243,403,310],{"emptyLinePlaceholder":309},[243,405,407],{"class":245,"line":406},15,[243,408,409],{"class":316},"# 业务方调用\n",[243,411,413,415,418],{"class":245,"line":412},16,[243,414,360],{"class":249},[243,416,417],{"class":253}," http:\u002F\u002Flocalhost:4000\u002Fv1\u002Fchat\u002Fcompletions",[243,419,372],{"class":261},[243,421,423,425,428],{"class":245,"line":422},17,[243,424,378],{"class":261},[243,426,427],{"class":253}," \"Authorization: Bearer sk-xxx\"",[243,429,372],{"class":261},[243,431,433,435],{"class":245,"line":432},18,[243,434,395],{"class":261},[243,436,437],{"class":253}," '{\"model\":\"gpt-5.4\",\"messages\":[...]}'\n",[22,439,440],{"id":440},"对比",[442,443,444,465],"table",{},[445,446,447],"thead",{},[448,449,450,454,456,459,462],"tr",{},[451,452,453],"th",{},"维度",[451,455,12],{},[451,457,458],{},"OpenRouter",[451,460,461],{},"One-API",[451,463,464],{},"Portkey",[466,467,468,484,499,514,531,544,558,574,588],"tbody",{},[448,469,470,474,477,480,482],{},[471,472,473],"td",{},"形态",[471,475,476],{},"OSS + Enterprise",[471,478,479],{},"SaaS",[471,481,150],{},[471,483,479],{},[448,485,486,489,492,495,497],{},[471,487,488],{},"协议",[471,490,491],{},"MIT",[471,493,494],{},"–",[471,496,491],{},[471,498,494],{},[448,500,501,504,507,510,512],{},[471,502,503],{},"自托管",[471,505,506],{},"✅",[471,508,509],{},"❌",[471,511,506],{},[471,513,156],{},[448,515,516,519,522,525,528],{},[471,517,518],{},"模型数",[471,520,521],{},"100+",[471,523,524],{},"300+",[471,526,527],{},"30+",[471,529,530],{},"250+",[448,532,533,536,538,540,542],{},[471,534,535],{},"虚拟 Key + 预算",[471,537,506],{},[471,539,494],{},[471,541,506],{},[471,543,506],{},[448,545,546,549,552,554,556],{},[471,547,548],{},"Fallback",[471,550,551],{},"✅ 三类",[471,553,506],{},[471,555,506],{},[471,557,506],{},[448,559,560,563,566,569,572],{},[471,561,562],{},"admin UI",[471,564,565],{},"✅ 内置",[471,567,568],{},"dashboard",[471,570,571],{},"✅ 中文 UI",[471,573,506],{},[448,575,576,579,581,583,586],{},[471,577,578],{},"中文支付",[471,580,509],{},[471,582,509],{},[471,584,585],{},"✅ EPay 内置",[471,587,509],{},[448,589,590,593,596,598,601],{},[471,591,592],{},"集成观测",[471,594,595],{},"Langfuse\u002FProm",[471,597,568],{},[471,599,600],{},"仪表盘",[471,602,603],{},"原生",[22,605,606],{"id":606},"避坑",[42,608,609,617,623,629,635,645,651,665],{},[45,610,611,66,614,616],{},[48,612,613],{},"必固定版本号",[31,615,211],{}," \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）",[45,618,619,622],{},[48,620,621],{},"salt_key 不能轮换","：初始化前生成 + 加密备份",[45,624,625,628],{},[48,626,627],{},"Postgres 备份","：所有虚拟 Key + 预算都在 DB，必须定期备份",[45,630,631,634],{},[48,632,633],{},"fallback 链别堆超 3 个","：失败叠加延迟 + 计费混乱",[45,636,637,640,641,644],{},[48,638,639],{},"drop_params 谨慎开","：开 ",[31,642,643],{},"drop_params: true"," 会静默丢不兼容字段，调试时容易 confused",[45,646,647,650],{},[48,648,649],{},"PII guardrail 不是 0 延迟","：Presidio 调用增 50–100ms，敏感场景再用",[45,652,653,656,657,660,661,664],{},[48,654,655],{},"monorepo + workspace 模型映射","：业务方调 ",[31,658,659],{},"gpt-4"," → config 映射到 ",[31,662,663],{},"openai\u002Fgpt-5.4-mini","，要规划稳定映射表",[45,666,667,670],{},[48,668,669],{},"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",[22,672,674],{"id":673},"适合-不适合","适合 \u002F 不适合",[42,676,677,680,683,686,689,692,695,698],{},[45,678,679],{},"✅ 中大型团队 \u002F 微服务架构",[45,681,682],{},"✅ 合规 \u002F 数据主权要求",[45,684,685],{},"✅ 多团队预算治理",[45,687,688],{},"✅ 想叠加 Langfuse \u002F Helicone \u002F Prometheus 观测",[45,690,691],{},"✅ BYOK 模式跨多厂商",[45,693,694],{},"❌ 不想运维（走 OpenRouter \u002F Portkey）",[45,696,697],{},"❌ 国内业务支付 + 中文运营（走 one-api \u002F new-api）",[45,699,700],{},"❌ 纯个人项目（SDK 足够，不需要 Proxy）",[22,702,703],{"id":703},"相关阅读",[42,705,706,713,719,725],{},[45,707,708],{},[709,710,712],"a",{"href":711},"\u002Ftools\u002Fcoding\u002Fapi\u002Fopenrouter","OpenRouter 评测",[45,714,715],{},[709,716,718],{"href":717},"\u002Ftools\u002Fcoding\u002Fapi\u002Fone-api","One-API 评测",[45,720,721],{},[709,722,724],{"href":723},"\u002Ftools\u002Fcoding\u002Fapi\u002Fportkey","Portkey 评测",[45,726,727],{},[709,728,730],{"href":729},"\u002Ftools\u002Fcoding\u002Fapi\u002Fhelicone","Helicone 评测",[22,732,733],{"id":733},"来源",[735,736,737,745,752,759],"ol",{},[45,738,739,740],{},"NerdLevelTech — LiteLLM Proxy Production Tutorial 2026（含 v1.85.0 + 供应链事件细节）",[709,741,742],{"href":742,"rel":743},"https:\u002F\u002Fnerdleveltech.com\u002Flitellm-proxy-production-llm-gateway-tutorial",[744],"nofollow",[45,746,747,748],{},"LiteLLM 官方文档 — Fallbacks \u002F Retries \u002F Cooldowns ",[709,749,750],{"href":750,"rel":751},"https:\u002F\u002Fdocs.litellm.ai\u002Fdocs\u002Fproxy\u002Freliability",[744],[45,753,754,755],{},"Youngju.dev — LiteLLM Complete Guide 2026（SDK vs Proxy \u002F cost-based routing）",[709,756,757],{"href":757,"rel":758},"https:\u002F\u002Fwww.youngju.dev\u002Fblog\u002Fculture\u002F2026-03-25-litellm-unified-llm-api-proxy-guide-2025.en",[744],[45,760,761,762],{},"GitHub — BerriAI\u002Flitellm README ",[709,763,764],{"href":764,"rel":765},"https:\u002F\u002Fgithub.com\u002FBerriAI\u002Flitellm",[744],[767,768,769],"style",{},"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":239,"searchDepth":290,"depth":290,"links":771},[772,773,774,775,776,777,778,779,780,781],{"id":24,"depth":268,"text":25},{"id":40,"depth":268,"text":40},{"id":143,"depth":268,"text":143},{"id":169,"depth":268,"text":170},{"id":231,"depth":268,"text":232},{"id":440,"depth":268,"text":440},{"id":606,"depth":268,"text":606},{"id":673,"depth":268,"text":674},{"id":703,"depth":268,"text":703},{"id":733,"depth":268,"text":733},"api","\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。",false,"md",[788,791,794,797],{"q":789,"a":790},"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":792,"a":793},"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":795,"a":796},"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":798,"a":799},"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。",[801],"en",{},"\u002Ftools\u002Fcoding\u002Fapi\u002Flitellm","coding",[806,807,334,808],"sdk","proxy","kubernetes",[810,815],{"plan":811,"price":812,"features":813,"notes":814},"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":156,"price":816,"features":817,"notes":818},"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",[822],"onboarding\u002Frag-pipeline-build",[824],"llm-gateway-comparison",{"power":313,"ux":306,"price":313,"cn_support":290,"stability":306},{"title":12,"description":784},"coding\u002Fapi\u002Flitellm",[829,832,834,836],{"name":830,"url":742,"accessed":831},"NerdLevelTech — LiteLLM Proxy 生产教程 2026","2026-06-24",{"name":833,"url":750,"accessed":831},"LiteLLM 官方文档 — Fallbacks",{"name":835,"url":757,"accessed":831},"Youngju.dev — LiteLLM Complete Guide 2026",{"name":837,"url":764,"accessed":831},"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",[841,807,806,842,843,844,845,846],"llm-gateway","opensource","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","JD7TeoQ-by7hjinNw1E4qfhTneex-5pJiPI0H8uth5Y",{"id":851,"title":852,"alternatives":853,"api_compatible":9,"body":855,"category":1480,"chinese_friendly":246,"cover":1481,"description":1482,"domestic":785,"extension":786,"faq":1483,"free":785,"github":9,"languages":1496,"lastVerified":9,"meta":1497,"models":9,"navigation":309,"notSuitable":9,"opensource":309,"path":1498,"pillar":804,"platforms":1499,"priceTable":1501,"pricing":1513,"published":820,"relatedPlaybooks":1514,"relatedReviews":9,"score":1515,"self_host":309,"seo":1516,"seoTitle":9,"slug":1517,"sources":1518,"stem":1527,"suitable":9,"tagline":1528,"tags":1529,"updated":831,"verdict":1535,"website":1536,"__hash__":1537},"tools\u002Ftools\u002Fcoding\u002Fagent\u002Fraga.md","RagaAI Catalyst",[17,16,827,854],"agent\u002Fplatform\u002Flangflow",{"type":19,"value":856,"toc":1468},[857,859,862,865,867,923,925,944,949,953,957,977,981,1004,1007,1023,1167,1169,1330,1332,1382,1384,1410,1412,1433,1435,1465],[22,858,25],{"id":24},[27,860,861],{},"RagaAI Catalyst 是印度 RagaAI 出品的 AI agent \u002F RAG \u002F LLM 应用『测试 + 观测』专项平台——Python SDK + 云端，300+ 自动化测试库覆盖 hallucination \u002F bias \u002F RAG faithfulness \u002F agentic 协作，多 agent 系统 tracing + debug，主打『部署前发现生产风险减少 90%』。Catalyst SDK 开源（github.com\u002Fraga-ai-hub\u002FRagaAI-Catalyst），Cloud \u002F Enterprise 走销售定价。",[27,863,864],{},"适合：构建生产级 agentic 系统的中大型团队（金融 \u002F 医疗 \u002F 政企）需要 pre-prod 风险量化；RAG pipeline 需要 retrieval \u002F faithfulness 多维评测；多 agent 协作系统需要 trace + debug 工具链。不适合：单一 LLM 调用 + 简单观测（Helicone \u002F Langfuse 更轻）；预算紧的小团队（社区版 SDK 够但 Cloud 走企业销售）；中文 \u002F 国内合规为主（生态弱）。",[22,866,40],{"id":40},[42,868,869,875,881,887,893,899,905,911,917],{},[45,870,871,874],{},[48,872,873],{},"300+ 自动化测试","：LLM（hallucination \u002F toxicity \u002F PII \u002F prompt injection）+ RAG（precision \u002F faithfulness \u002F relevance）+ Agentic（tool 正确性 \u002F 协作一致性 \u002F 任务完成率）",[45,876,877,880],{},[48,878,879],{},"多 agent tracing","：每次 LLM \u002F tool \u002F sub-agent 调用都可追溯 + 时序回放",[45,882,883,886],{},[48,884,885],{},"多 agent debug","：复杂 agent 失败时回放每个决策点 + 上下文",[45,888,889,892],{},[48,890,891],{},"Python SDK","：包装 LangChain \u002F LlamaIndex \u002F 自家框架自动埋点",[45,894,895,898],{},[48,896,897],{},"风险量化","：每个测试出风险分 + 影响面 + 修复建议",[45,900,901,904],{},[48,902,903],{},"数据集管理","：build eval dataset + 跑 regression",[45,906,907,910],{},[48,908,909],{},"报告 \u002F 仪表盘","：团队级风险仪表盘 + 趋势 + CI\u002FCD 集成",[45,912,913,916],{},[48,914,915],{},"自托管 SDK + Cloud 协同","：SDK 本地跑、Cloud 集中可视化",[45,918,919,922],{},[48,920,921],{},"合规友好","：私有部署 + SSO + 审计（Enterprise）",[22,924,143],{"id":143},[42,926,927,933,939],{},[45,928,929,932],{},[48,930,931],{},"Catalyst OSS Python SDK","：$0；功能含 tracing + 部分 eval",[45,934,935,938],{},[48,936,937],{},"Cloud","：Custom（联系销售）；含全 300+ 测试库 + 仪表盘 + 协作",[45,940,941,943],{},[48,942,156],{},"：Custom + SSO + 私有部署 + SLA",[162,945,946],{},[27,947,948],{},"真实场景：先用 OSS SDK 跑 trace 评估价值，进 PoC 后再谈 Cloud 价格。",[22,950,952],{"id":951},"实测中型-rag-agentic-产品-印度欧美客户","实测（中型 RAG \u002F agentic 产品 \u002F 印度欧美客户）",[27,954,955],{},[48,956,175],{},[42,958,959,962,965,968,971,974],{},[45,960,961],{},"300+ 测试库省去 reinvent 各种 eval metric 的工作",[45,963,964],{},"RAG faithfulness \u002F context precision 测试对反 hallucination 很有用",[45,966,967],{},"多 agent trace + 回放在调试复杂协作链时是救命工具",[45,969,970],{},"风险评分给业务方一个 quantitative 沟通口径",[45,972,973],{},"Catalyst SDK 开源，预算紧团队也能先用上",[45,975,976],{},"与 LangChain \u002F LlamaIndex 集成顺滑",[27,978,979],{},[48,980,203],{},[42,982,983,986,989,992,995,998,1001],{},[45,984,985],{},"文档对高级配置 + 自定义测试覆盖不足，社区反馈一致",[45,987,988],{},"Cloud 定价不透明，PoC 才能拿到报价",[45,990,991],{},"报告 \u002F UI 偏英文 + 印度产品风格，中文场景弱",[45,993,994],{},"多 agent trace 在超大调用图（>500 step）下渲染慢",[45,996,997],{},"测试结果质量依赖 dataset 质量，garbage in \u002F garbage out",[45,999,1000],{},"比 Langfuse 更偏『测试』少偏『日常 observability』，两个工具有时要叠用",[45,1002,1003],{},"国内访问 Cloud 延迟 + 合规需要评估",[22,1005,1006],{"id":1006},"上手",[234,1008,1010],{"className":236,"code":1009,"language":238,"meta":239,"style":239},"pip install ragaai-catalyst\n",[31,1011,1012],{"__ignoreMap":239},[243,1013,1014,1017,1020],{"class":245,"line":246},[243,1015,1016],{"class":249},"pip",[243,1018,1019],{"class":253}," install",[243,1021,1022],{"class":253}," ragaai-catalyst\n",[234,1024,1028],{"className":1025,"code":1026,"language":1027,"meta":239,"style":239},"language-python shiki shiki-themes github-light github-dark","from ragaai_catalyst import RagaAICatalyst, Tracer\n\ncatalyst = RagaAICatalyst(\n    access_key=\"...\",\n    secret_key=\"...\",\n    base_url=\"https:\u002F\u002Fcatalyst.raga.ai\"\n)\n\ntracer = Tracer(project_name=\"my-rag-app\", tracer_type=\"langchain\")\ntracer.start()\n\n# 你的 LangChain \u002F LlamaIndex \u002F 自家 agent 代码\n# tracer 自动捕获 LLM \u002F tool \u002F sub-agent 调用\n\ntracer.stop()\n# 登录 Catalyst Cloud 看 trace + 跑 300+ 测试\n","python",[31,1029,1030,1044,1048,1059,1073,1084,1094,1099,1103,1134,1139,1143,1148,1153,1157,1162],{"__ignoreMap":239},[243,1031,1032,1035,1038,1041],{"class":245,"line":246},[243,1033,1034],{"class":283},"from",[243,1036,1037],{"class":257}," ragaai_catalyst ",[243,1039,1040],{"class":283},"import",[243,1042,1043],{"class":257}," RagaAICatalyst, Tracer\n",[243,1045,1046],{"class":245,"line":268},[243,1047,310],{"emptyLinePlaceholder":309},[243,1049,1050,1053,1056],{"class":245,"line":290},[243,1051,1052],{"class":257},"catalyst ",[243,1054,1055],{"class":283},"=",[243,1057,1058],{"class":257}," RagaAICatalyst(\n",[243,1060,1061,1065,1067,1070],{"class":245,"line":306},[243,1062,1064],{"class":1063},"s4XuR","    access_key",[243,1066,1055],{"class":283},[243,1068,1069],{"class":253},"\"...\"",[243,1071,1072],{"class":257},",\n",[243,1074,1075,1078,1080,1082],{"class":245,"line":313},[243,1076,1077],{"class":1063},"    secret_key",[243,1079,1055],{"class":283},[243,1081,1069],{"class":253},[243,1083,1072],{"class":257},[243,1085,1086,1089,1091],{"class":245,"line":320},[243,1087,1088],{"class":1063},"    base_url",[243,1090,1055],{"class":283},[243,1092,1093],{"class":253},"\"https:\u002F\u002Fcatalyst.raga.ai\"\n",[243,1095,1096],{"class":245,"line":326},[243,1097,1098],{"class":257},")\n",[243,1100,1101],{"class":245,"line":331},[243,1102,310],{"emptyLinePlaceholder":309},[243,1104,1105,1108,1110,1113,1116,1118,1121,1124,1127,1129,1132],{"class":245,"line":346},[243,1106,1107],{"class":257},"tracer ",[243,1109,1055],{"class":283},[243,1111,1112],{"class":257}," Tracer(",[243,1114,1115],{"class":1063},"project_name",[243,1117,1055],{"class":283},[243,1119,1120],{"class":253},"\"my-rag-app\"",[243,1122,1123],{"class":257},", ",[243,1125,1126],{"class":1063},"tracer_type",[243,1128,1055],{"class":283},[243,1130,1131],{"class":253},"\"langchain\"",[243,1133,1098],{"class":257},[243,1135,1136],{"class":245,"line":351},[243,1137,1138],{"class":257},"tracer.start()\n",[243,1140,1141],{"class":245,"line":357},[243,1142,310],{"emptyLinePlaceholder":309},[243,1144,1145],{"class":245,"line":375},[243,1146,1147],{"class":316},"# 你的 LangChain \u002F LlamaIndex \u002F 自家 agent 代码\n",[243,1149,1150],{"class":245,"line":392},[243,1151,1152],{"class":316},"# tracer 自动捕获 LLM \u002F tool \u002F sub-agent 调用\n",[243,1154,1155],{"class":245,"line":401},[243,1156,310],{"emptyLinePlaceholder":309},[243,1158,1159],{"class":245,"line":406},[243,1160,1161],{"class":257},"tracer.stop()\n",[243,1163,1164],{"class":245,"line":412},[243,1165,1166],{"class":316},"# 登录 Catalyst Cloud 看 trace + 跑 300+ 测试\n",[22,1168,440],{"id":440},[442,1170,1171,1187],{},[445,1172,1173],{},[448,1174,1175,1177,1179,1182,1185],{},[451,1176,453],{},[451,1178,852],{},[451,1180,1181],{},"Langfuse",[451,1183,1184],{},"Helicone",[451,1186,464],{},[466,1188,1189,1206,1222,1238,1253,1267,1283,1296,1313],{},[448,1190,1191,1194,1197,1200,1203],{},[471,1192,1193],{},"主打",[471,1195,1196],{},"测试 + 风险量化",[471,1198,1199],{},"Tracing + eval",[471,1201,1202],{},"Proxy 观测 + gateway",[471,1204,1205],{},"Gateway + 观测",[448,1207,1208,1211,1214,1217,1220],{},[471,1209,1210],{},"自动化测试库",[471,1212,1213],{},"✅ 300+",[471,1215,1216],{},"部分",[471,1218,1219],{},"浅",[471,1221,1219],{},[448,1223,1224,1227,1230,1233,1235],{},[471,1225,1226],{},"Agentic debug",[471,1228,1229],{},"✅ 强",[471,1231,1232],{},"✅ nested span",[471,1234,1219],{},[471,1236,1237],{},"中",[448,1239,1240,1243,1245,1248,1251],{},[471,1241,1242],{},"RAG 专项测试",[471,1244,1229],{},[471,1246,1247],{},"✅ 中",[471,1249,1250],{},"弱",[471,1252,1250],{},[448,1254,1255,1258,1260,1262,1264],{},[471,1256,1257],{},"Proxy \u002F gateway",[471,1259,509],{},[471,1261,509],{},[471,1263,506],{},[471,1265,1266],{},"✅ 250+",[448,1268,1269,1272,1275,1278,1280],{},[471,1270,1271],{},"自托管 OSS",[471,1273,1274],{},"SDK 部分",[471,1276,1277],{},"✅ MIT 19K+",[471,1279,506],{},[471,1281,1282],{},"OSS Gateway only",[448,1284,1285,1288,1290,1292,1294],{},[471,1286,1287],{},"中文生态",[471,1289,509],{},[471,1291,509],{},[471,1293,509],{},[471,1295,509],{},[448,1297,1298,1301,1304,1307,1310],{},[471,1299,1300],{},"定价透明",[471,1302,1303],{},"❌ Custom",[471,1305,1306],{},"✅ $0\u002F$29\u002FEnterprise",[471,1308,1309],{},"✅ $0\u002F$79\u002F$799",[471,1311,1312],{},"✅ $0\u002F$49\u002F$799",[448,1314,1315,1318,1321,1324,1327],{},[471,1316,1317],{},"适合",[471,1319,1320],{},"pre-prod testing + agentic",[471,1322,1323],{},"日常 tracing + eval",[471,1325,1326],{},"改 URL 快上手",[471,1328,1329],{},"gateway + 治理",[22,1331,606],{"id":606},[42,1333,1334,1340,1346,1352,1358,1364,1370,1376],{},[45,1335,1336,1339],{},[48,1337,1338],{},"OSS SDK ≠ 完整 Cloud","：300+ 测试库主要在 Cloud，OSS 主要 tracing + 部分 eval",[45,1341,1342,1345],{},[48,1343,1344],{},"dataset 质量决定测试质量","：建 eval dataset 时务必含 edge case \u002F adversarial 例",[45,1347,1348,1351],{},[48,1349,1350],{},"大 trace 渲染慢","：>500 step 的 agent 用 filter \u002F sampling",[45,1353,1354,1357],{},[48,1355,1356],{},"Cloud 定价 PoC 谈","：先 SDK 跑两月有数据再谈合同",[45,1359,1360,1363],{},[48,1361,1362],{},"不替代日常 observability","：复杂场景叠 Langfuse \u002F Helicone",[45,1365,1366,1369],{},[48,1367,1368],{},"中文场景弱","：报告 \u002F 文档 \u002F UI 均英文为主，国内项目要评估团队接受度",[45,1371,1372,1375],{},[48,1373,1374],{},"Agentic 测试时间长","：300+ 测试跑一遍可能数小时，CI\u002FCD 集成要 schedule 而非每次 PR",[45,1377,1378,1381],{},[48,1379,1380],{},"风险评分谨慎宣传","：『-90% 生产风险』是营销话术，实际依赖你的实施质量",[22,1383,674],{"id":673},[42,1385,1386,1389,1392,1395,1398,1401,1404,1407],{},[45,1387,1388],{},"✅ 生产级 agentic 系统 \u002F RAG pipeline 的中大型团队",[45,1390,1391],{},"✅ 金融 \u002F 医疗 \u002F 政企需要风险量化沟通",[45,1393,1394],{},"✅ 多 agent 协作系统需要 trace + debug",[45,1396,1397],{},"✅ 想把 LLM eval 从手工脚本升级成系统化平台",[45,1399,1400],{},"❌ 单一 LLM 调用 + 简单观测",[45,1402,1403],{},"❌ 预算紧 + 不要 Cloud（OSS SDK 部分能力够）",[45,1405,1406],{},"❌ 中文 \u002F 国内合规为主",[45,1408,1409],{},"❌ 不愿走企业销售流程",[22,1411,703],{"id":703},[42,1413,1414,1418,1422,1427],{},[45,1415,1416],{},[709,1417,730],{"href":729},[45,1419,1420],{},[709,1421,724],{"href":723},[45,1423,1424],{},[709,1425,1426],{"href":803},"LiteLLM 评测",[45,1428,1429],{},[709,1430,1432],{"href":1431},"\u002Ftools\u002Fagent\u002Flangflow","Langflow 评测",[22,1434,733],{"id":733},[735,1436,1437,1444,1451,1458],{},[45,1438,1439,1440],{},"F6S — RagaAI Catalyst 产品概览 ",[709,1441,1442],{"href":1442,"rel":1443},"https:\u002F\u002Fwww.f6s.com\u002Fsoftware\u002Fragaai-catalyst",[744],[45,1445,1446,1447],{},"AIIndigo — RagaAI Catalyst 评测 + 定价 + 替代品 2026 ",[709,1448,1449],{"href":1449,"rel":1450},"https:\u002F\u002Faiindigo.com\u002Ftool\u002Fragaai-catalyst",[744],[45,1452,1453,1454],{},"SwitchTools — RagaAI Inc 平台（300+ 测试 \u002F 90% 风险下降）",[709,1455,1456],{"href":1456,"rel":1457},"https:\u002F\u002Fwww.switchtools.io\u002Ftool\u002Fragaai-inc",[744],[45,1459,1460,1461],{},"SoftwareSuggest — RagaAI Details \u002F Pricing 2026 ",[709,1462,1463],{"href":1463,"rel":1464},"https:\u002F\u002Fwww.softwaresuggest.com\u002Fragaai",[744],[767,1466,1467],{},"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 .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);}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 .s4XuR, html code.shiki .s4XuR{--shiki-default:#E36209;--shiki-dark:#FFAB70}html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}",{"title":239,"searchDepth":290,"depth":290,"links":1469},[1470,1471,1472,1473,1474,1475,1476,1477,1478,1479],{"id":24,"depth":268,"text":25},{"id":40,"depth":268,"text":40},{"id":143,"depth":268,"text":143},{"id":951,"depth":268,"text":952},{"id":1006,"depth":268,"text":1006},{"id":440,"depth":268,"text":440},{"id":606,"depth":268,"text":606},{"id":673,"depth":268,"text":674},{"id":703,"depth":268,"text":703},{"id":733,"depth":268,"text":733},"agent","\u002Fimg\u002Ftools\u002Fraga.webp","RagaAI Catalyst 真实评测：印度 RagaAI 出品的 AI 测试 + 观测平台（Python SDK + 云端）。300+ 自动化测试覆盖 LLM \u002F RAG \u002F 多 agent，多 agent 系统 tracing + 调试，主打『部署前发现风险减少 90%』。F6S \u002F SoftwareSuggest \u002F SwitchTools 已收录。定价以企业销售为主，社区版 Python SDK 开源。",[1484,1487,1490,1493],{"q":1485,"a":1486},"RagaAI 和 Langfuse \u002F Helicone 怎么选？","RagaAI Catalyst 把『测试』作为一等公民——300+ 自动化测试 + RAG \u002F Agent 风险量化，强调『部署前发现』。Langfuse 是 tracing + eval 双强，Helicone 是 proxy 观测 + gateway。要 agentic 系统的 pre-prod testing + 风险评分 → RagaAI；要日常 LLM trace \u002F 简单 eval → Langfuse；要改 baseURL 看 cost → Helicone。",{"q":1488,"a":1489},"300+ 测试覆盖什么？","覆盖三类系统：(1) LLM 应用——hallucination \u002F toxicity \u002F bias \u002F PII leakage \u002F prompt injection；(2) RAG pipeline——检索准确率 \u002F context precision \u002F answer relevance \u002F faithfulness；(3) Agentic 系统——tool use 正确性 \u002F 多 agent 协作一致性 \u002F 任务完成率 \u002F 安全围栏。",{"q":1491,"a":1492},"Catalyst SDK 怎么用？","pip install ragaai-catalyst → 包装 LangChain \u002F LlamaIndex \u002F 自家 agent 框架 → 自动 trace 每次 LLM \u002F tool \u002F sub-agent 调用 → 上传 Cloud 看仪表盘 \u002F 跑测试。SDK 本身 OSS，Cloud 部分付费。",{"q":1494,"a":1495},"国内能用吗？","Python SDK 自托管 OK。Cloud 服务器在海外，国内延迟 + 合规要评估。中文场景 + 中文文档极弱，prompt \u002F 报告均英文。国内类似定位推荐看 PromptLayer \u002F Helicone OSS 自托管，或自家 LangSmith \u002F Phoenix 组合。",[801],{},"\u002Ftools\u002Fcoding\u002Fagent\u002Fraga",[1027,806,1500],"cloud",[1502,1507,1510],{"plan":1503,"price":1504,"features":1505,"notes":1506},"Catalyst OSS","$0","Python SDK + agent \u002F LLM \u002F tool tracing + 多 agent debug + 部分 eval","github.com\u002Fraga-ai-hub\u002FRagaAI-Catalyst",{"plan":937,"price":816,"features":1508,"notes":1509},"300+ 自动化测试库 + RAG \u002F LLM \u002F Agentic 全景 + 报告 \u002F 仪表盘 + 协作","联系销售",{"plan":156,"price":816,"features":1511,"notes":1512},"SSO + 私有部署 + 合规 + SLA + 专属支持","大型 \u002F 受监管","Catalyst Python SDK 开源 \u002F Cloud Enterprise 定制",[822],{"power":313,"ux":290,"price":290,"cn_support":246,"stability":306},{"title":852,"description":1482},"coding\u002Fagent\u002Fraga",[1519,1521,1523,1525],{"name":1520,"url":1442,"accessed":831},"F6S — RagaAI Catalyst 概览",{"name":1522,"url":1449,"accessed":831},"AIIndigo — RagaAI Catalyst 评测 2026",{"name":1524,"url":1456,"accessed":831},"SwitchTools — RagaAI Inc 平台 + 90% 风险下降",{"name":1526,"url":1463,"accessed":831},"SoftwareSuggest — RagaAI Details 2026","tools\u002Fcoding\u002Fagent\u002Fraga","AI agent 测试 + 观测平台——300+ 自动化测试覆盖 LLM\u002FRAG\u002FAgentic，量化『生产风险 -90%』",[1530,1531,1532,1533,1534,842],"agent-testing","llm-eval","rag-testing","observability","tracing","agentic AI 系统的『生产前测试 + 上线后观测』专项工具。要 300+ 测试库 + 多 agent debug 走 Raga；要简单 LLM trace 用 Langfuse \u002F Helicone；要网关 + 治理用 Portkey。中文 \u002F 国内场景生态弱。","https:\u002F\u002Fraga.ai","D4CXLJkcXBHicMVKL1pAZ4PurzLASuxOKD5agbaXSeQ",1784565441008]