✨ feat: add Cohere provider support (#7016)

* ✨ feat: add Cohere provider support

* 💄 style: update cohere model list

* 🐛 fix: fix cohere calling, exclude user and stream_options

* 💄 style: update model description

* 🐛 fix: fix cohere model list fetch

* 🐛 fix: fix cohere calling

* 💄 style: update provider desc

* 💄 style: update model name

Dockerfile

@@ -157,6 +157,8 @@ ENV \
     BAICHUAN_API_KEY="" BAICHUAN_MODEL_LIST="" \
     # Cloudflare
     CLOUDFLARE_API_KEY="" CLOUDFLARE_BASE_URL_OR_ACCOUNT_ID="" CLOUDFLARE_MODEL_LIST="" \
+    # Cohere
+    COHERE_API_KEY="" COHERE_MODEL_LIST="" COHERE_PROXY_URL="" \
     # DeepSeek
     DEEPSEEK_API_KEY="" DEEPSEEK_MODEL_LIST="" \
     # Fireworks AI

Dockerfile.database

@@ -200,6 +200,8 @@ ENV \
     BAICHUAN_API_KEY="" BAICHUAN_MODEL_LIST="" \
     # Cloudflare
     CLOUDFLARE_API_KEY="" CLOUDFLARE_BASE_URL_OR_ACCOUNT_ID="" CLOUDFLARE_MODEL_LIST="" \
+    # Cohere
+    COHERE_API_KEY="" COHERE_MODEL_LIST="" COHERE_PROXY_URL="" \
     # DeepSeek
     DEEPSEEK_API_KEY="" DEEPSEEK_MODEL_LIST="" \
     # Fireworks AI

Dockerfile.pglite

@@ -158,6 +158,8 @@ ENV \
     BAICHUAN_API_KEY="" BAICHUAN_MODEL_LIST="" \
     # Cloudflare
     CLOUDFLARE_API_KEY="" CLOUDFLARE_BASE_URL_OR_ACCOUNT_ID="" CLOUDFLARE_MODEL_LIST="" \
+    # Cohere
+    COHERE_API_KEY="" COHERE_MODEL_LIST="" COHERE_PROXY_URL="" \
     # DeepSeek
     DEEPSEEK_API_KEY="" DEEPSEEK_MODEL_LIST="" \
     # Fireworks AI

src/app/[variants]/(main)/settings/llm/ProviderList/providers.tsx

@@ -5,6 +5,7 @@ import {
   Ai360ProviderCard,
   AnthropicProviderCard,
   BaichuanProviderCard,
+  CohereProviderCard,
   DeepSeekProviderCard,
   FireworksAIProviderCard,
   GiteeAIProviderCard,
@@ -82,6 +83,7 @@ export const useProviderList = (): ProviderItem[] => {
       XAIProviderCard,
       JinaProviderCard,
       SambaNovaProviderCard,
+      CohereProviderCard,
       QwenProviderCard,
       WenxinProviderCard,
       HunyuanProviderCard,

src/config/aiModels/cohere.ts

@@ -0,0 +1,243 @@
+import { AIChatModelCard } from '@/types/aiModel';
+
+const cohereChatModels: AIChatModelCard[] = [
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 256_000,
+    description:
+      'Command A 是我们迄今为止性能最强的模型，在工具使用、代理、检索增强生成（RAG）和多语言应用场景方面表现出色。Command A 具有 256K 的上下文长度，仅需两块 GPU 即可运行，并且相比于 Command R+ 08-2024，吞吐量提高了 150%。',
+    displayName: 'Command A 03-2025',
+    enabled: true,
+    id: 'command-a-03-2025',
+    maxOutput: 8000,
+    pricing: {
+      input: 2.5,
+      output: 10
+    },
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 128_000,
+    description:
+      'command-r-plus 是 command-r-plus-04-2024 的别名，因此如果您在 API 中使用 command-r-plus，实际上指向的就是该模型。',
+    displayName: 'Command R+',
+    enabled: true,
+    id: 'command-r-plus',
+    maxOutput: 4000,
+    pricing: {
+      input: 2.5,
+      output: 10
+    },
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 128_000,
+    description:
+      'Command R+ 是一个遵循指令的对话模型，在语言任务方面表现出更高的质量、更可靠，并且相比以往模型具有更长的上下文长度。它最适用于复杂的 RAG 工作流和多步工具使用。',
+    displayName: 'Command R+ 04-2024',
+    id: 'command-r-plus-04-2024',
+    maxOutput: 4000,
+    pricing: {
+      input: 3,
+      output: 15
+    },
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 128_000,
+    description:
+      'command-r 是 command-c-03-2024 的别名，因此如果您在 API 中使用 command-r，实际上指向的就是该模型。',
+    displayName: 'Command R',
+    enabled: true,
+    id: 'command-r',
+    maxOutput: 4000,
+    pricing: {
+      input: 0.15,
+      output: 0.6
+    },
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 128_000,
+    description:
+      'command-r-08-2024 是 Command R 模型的更新版本，于 2024 年 8 月发布。',
+    displayName: 'Command R 08-2024',
+    id: 'command-r-08-2024',
+    maxOutput: 4000,
+    pricing: {
+      input: 0.15,
+      output: 0.6
+    },
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 128_000,
+    description:
+      'Command R 是一个遵循指令的对话模型，在语言任务方面表现出更高的质量、更可靠，并且相比以往模型具有更长的上下文长度。它可用于复杂的工作流程，如代码生成、检索增强生成（RAG）、工具使用和代理。',
+    displayName: 'Command R 03-2024',
+    id: 'command-r-03-2024',
+    maxOutput: 4000,
+    pricing: {
+      input: 0.5,
+      output: 1.5
+    },
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 128_000,
+    description:
+      'command-r7b-12-2024 是一个小型且高效的更新版本，于 2024 年 12 月发布。它在 RAG、工具使用、代理等需要复杂推理和多步处理的任务中表现出色。',
+    displayName: 'Command R7B 12-2024',
+    enabled: true,
+    id: 'command-r7b-12-2024',
+    maxOutput: 4000,
+    pricing: {
+      input: 0.0375,
+      output: 0.15
+    },
+    type: 'chat'
+  },
+  {
+    contextWindowTokens: 4000,
+    description:
+      '一个遵循指令的对话模型，在语言任务中表现出高质量、更可靠，并且相比我们的基础生成模型具有更长的上下文长度。',
+    displayName: 'Command',
+    enabled: true,
+    id: 'command',
+    maxOutput: 4000,
+    pricing: {
+      input: 1,
+      output: 2
+    },
+    type: 'chat'
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 128_000,
+    description:
+      '为了缩短主要版本发布之间的时间间隔，我们推出了 Command 模型的每夜版本。对于 Command 系列，这一版本称为 command-cightly。请注意，command-nightly 是最新、最具实验性且（可能）不稳定的版本。每夜版本会定期更新，且不会提前通知，因此不建议在生产环境中使用。',
+    displayName: 'Command Nightly',
+    id: 'command-nightly',
+    maxOutput: 4000,
+    pricing: {
+      input: 1,
+      output: 2
+    },
+    type: 'chat'
+  },
+  {
+    contextWindowTokens: 4000,
+    description:
+      '一个更小、更快的 Command 版本，几乎同样强大，但速度更快。',
+    displayName: 'Command Light',
+    enabled: true,
+    id: 'command-light',
+    maxOutput: 4000,
+    pricing: {
+      input: 0.3,
+      output: 0.6
+    },
+    type: 'chat'
+  },
+  {
+    contextWindowTokens: 4000,
+    description:
+      '为了缩短主要版本发布之间的时间间隔，我们推出了 Command 模型的每夜版本。对于 command-light 系列，这一版本称为 command-light-nightly。请注意，command-light-nightly 是最新、最具实验性且（可能）不稳定的版本。每夜版本会定期更新，且不会提前通知，因此不建议在生产环境中使用。',
+    displayName: 'Command Light Nightly',
+    id: 'command-light-nightly',
+    maxOutput: 4000,
+    pricing: {
+      input: 0.3,
+      output: 0.6
+    },
+    type: 'chat'
+  },
+  {
+    contextWindowTokens: 128_000,
+    description:
+      'Aya Expanse 是一款高性能的 32B 多语言模型，旨在通过指令调优、数据套利、偏好训练和模型合并的创新，挑战单语言模型的表现。它支持 23 种语言。',
+    displayName: 'Aya Expanse 32B',
+    enabled: true,
+    id: 'c4ai-aya-expanse-32b',
+    maxOutput: 4000,
+    pricing: {
+      input: 0.5,
+      output: 1.5
+    },
+    type: 'chat'
+  },
+  {
+    contextWindowTokens: 8000,
+    description:
+      'Aya Expanse 是一款高性能的 8B 多语言模型，旨在通过指令调优、数据套利、偏好训练和模型合并的创新，挑战单语言模型的表现。它支持 23 种语言。',
+    displayName: 'Aya Expanse 8B',
+    enabled: true,
+    id: 'c4ai-aya-expanse-8b',
+    maxOutput: 4000,
+    pricing: {
+      input: 0.5,
+      output: 1.5
+    },
+    type: 'chat'
+  },
+  {
+    abilities: {
+      vision: true,
+    },
+    contextWindowTokens: 16_000,
+    description:
+      'Aya Vision 是一款最先进的多模态模型，在语言、文本和图像能力的多个关键基准上表现出色。它支持 23 种语言。这个 320 亿参数的版本专注于最先进的多语言表现。',
+    displayName: 'Aya Vision 32B',
+    enabled: true,
+    id: 'c4ai-aya-vision-32b',
+    maxOutput: 4000,
+    pricing: {
+      input: 0.5,
+      output: 1.5
+    },
+    type: 'chat'
+  },
+  {
+    abilities: {
+      vision: true,
+    },
+    contextWindowTokens: 16_000,
+    description:
+      'Aya Vision 是一款最先进的多模态模型，在语言、文本和图像能力的多个关键基准上表现出色。这个 80 亿参数的版本专注于低延迟和最佳性能。',
+    displayName: 'Aya Vision 8B',
+    enabled: true,
+    id: 'c4ai-aya-vision-8b',
+    maxOutput: 4000,
+    pricing: {
+      input: 0.5,
+      output: 1.5
+    },
+    type: 'chat'
+  },
+]
+
+export const allModels = [...cohereChatModels];
+
+export default allModels;

src/config/aiModels/index.ts

@@ -8,6 +8,7 @@ import { default as azureai } from './azureai';
 import { default as baichuan } from './baichuan';
 import { default as bedrock } from './bedrock';
 import { default as cloudflare } from './cloudflare';
+import { default as cohere } from './cohere';
 import { default as deepseek } from './deepseek';
 import { default as doubao } from './doubao';
 import { default as fireworksai } from './fireworksai';
@@ -77,6 +78,7 @@ export const LOBE_DEFAULT_MODEL_LIST = buildDefaultModelList({
   baichuan,
   bedrock,
   cloudflare,
+  cohere,
   deepseek,
   doubao,
   fireworksai,
@@ -127,6 +129,7 @@ export { default as azureai } from './azureai';
 export { default as baichuan } from './baichuan';
 export { default as bedrock } from './bedrock';
 export { default as cloudflare } from './cloudflare';
+export { default as cohere } from './cohere';
 export { default as deepseek } from './deepseek';
 export { default as doubao } from './doubao';
 export { default as fireworksai } from './fireworksai';

src/config/llm.ts

@@ -150,6 +150,9 @@ export const getLLMConfig = () => {
 
       ENABLED_PPIO: z.boolean(),
       PPIO_API_KEY: z.string().optional(),
+
+      ENABLED_COHERE: z.boolean(),
+      COHERE_API_KEY: z.string().optional(),
     },
     runtimeEnv: {
       API_KEY_SELECT_MODE: process.env.API_KEY_SELECT_MODE,
@@ -298,6 +301,9 @@ export const getLLMConfig = () => {
 
       ENABLED_PPIO: !!process.env.PPIO_API_KEY,
       PPIO_API_KEY: process.env.PPIO_API_KEY,
+
+      ENABLED_COHERE: !!process.env.COHERE_API_KEY,
+      COHERE_API_KEY: process.env.COHERE_API_KEY,
     },
   });
 };

src/config/modelProviders/cohere.ts

@@ -0,0 +1,19 @@
+import { ModelProviderCard } from '@/types/llm';
+
+const Cohere: ModelProviderCard = {
+  chatModels: [],
+  checkModel: 'command-r7b-12-2024',
+  description: 'Cohere 为您带来最前沿的多语言模型、先进的检索功能以及为现代企业量身定制的 AI 工作空间 — 一切都集成在一个安全的平台中。',
+  id: 'cohere',
+  modelsUrl: 'https://docs.cohere.com/v2/docs/models',
+  name: 'Cohere',
+  settings: {
+    proxyUrl: {
+      placeholder: 'https://api.cohere.ai/compatibility/v1',
+    },
+    sdkType: 'openai',
+  },
+  url: 'https://cohere.com',
+};
+
+export default Cohere;

src/config/modelProviders/index.ts

@@ -8,6 +8,7 @@ import AzureAIProvider from './azureai';
 import BaichuanProvider from './baichuan';
 import BedrockProvider from './bedrock';
 import CloudflareProvider from './cloudflare';
+import CohereProvider from './cohere';
 import DeepSeekProvider from './deepseek';
 import DoubaoProvider from './doubao';
 import FireworksAIProvider from './fireworksai';
@@ -75,6 +76,7 @@ export const LOBE_DEFAULT_MODEL_LIST: ChatModelCard[] = [
   XAIProvider.chatModels,
   JinaProvider.chatModels,
   SambaNovaProvider.chatModels,
+  CohereProvider.chatModels,
   ZeroOneProvider.chatModels,
   StepfunProvider.chatModels,
   NovitaProvider.chatModels,
@@ -125,6 +127,7 @@ export const DEFAULT_MODEL_PROVIDER_LIST = [
   XAIProvider,
   JinaProvider,
   SambaNovaProvider,
+  CohereProvider,
   QwenProvider,
   WenxinProvider,
   TencentcloudProvider,
@@ -165,6 +168,7 @@ export { default as AzureAIProviderCard } from './azureai';
 export { default as BaichuanProviderCard } from './baichuan';
 export { default as BedrockProviderCard } from './bedrock';
 export { default as CloudflareProviderCard } from './cloudflare';
+export { default as CohereProviderCard } from './cohere';
 export { default as DeepSeekProviderCard } from './deepseek';
 export { default as DoubaoProviderCard } from './doubao';
 export { default as FireworksAIProviderCard } from './fireworksai';

src/libs/agent-runtime/cohere/index.ts

@@ -0,0 +1,69 @@
+import { ModelProvider } from '../types';
+import { LobeOpenAICompatibleFactory } from '../utils/openaiCompatibleFactory';
+
+import type { ChatModelCard } from '@/types/llm';
+
+export interface CohereModelCard {
+  context_length: number;
+  features: string[] | null;
+  name: string;
+  supports_vision: boolean;
+}
+
+export const LobeCohereAI = LobeOpenAICompatibleFactory({
+  baseURL: 'https://api.cohere.ai/compatibility/v1',
+  chatCompletion: {
+    // https://docs.cohere.com/v2/docs/compatibility-api#unsupported-parameters
+    excludeUsage: true,
+    handlePayload: (payload) => {
+      const { frequency_penalty, presence_penalty, top_p, ...rest } = payload;
+
+      return {
+        ...rest,
+        frequency_penalty:
+          frequency_penalty !== undefined && frequency_penalty > 0 && frequency_penalty <= 1
+            ? frequency_penalty
+            : undefined,
+        presence_penalty:
+          presence_penalty !== undefined && presence_penalty > 0 && presence_penalty <= 1
+            ? presence_penalty
+            : undefined,
+        top_p: top_p !== undefined && top_p > 0 && top_p < 1 ? top_p : undefined,
+      } as any;
+    },
+    noUserId: true,
+  },
+  debug: {
+    chatCompletion: () => process.env.DEBUG_COHERE_CHAT_COMPLETION === '1',
+  },
+  models: async ({ client }) => {
+    const { LOBE_DEFAULT_MODEL_LIST } = await import('@/config/aiModels');
+
+    client.baseURL = 'https://api.cohere.com/v1';
+
+    const modelsPage = await client.models.list() as any;
+    const modelList: CohereModelCard[] = modelsPage.body.models;
+
+    return modelList
+      .map((model) => {
+        const knownModel = LOBE_DEFAULT_MODEL_LIST.find((m) => model.name.toLowerCase() === m.id.toLowerCase());
+
+        return {
+          contextWindowTokens: model.context_length,
+          displayName: knownModel?.displayName ?? undefined,
+          enabled: knownModel?.enabled || false,
+          functionCall:
+            (model.features && model.features.includes("tools"))
+            || knownModel?.abilities?.functionCall
+            || false,
+          id: model.name,
+          vision:
+            model.supports_vision
+            || knownModel?.abilities?.vision
+            || false,
+        };
+      })
+      .filter(Boolean) as ChatModelCard[];
+  },
+  provider: ModelProvider.Cohere,
+});

src/libs/agent-runtime/runtimeMap.ts

@@ -6,6 +6,7 @@ import { LobeAzureAI } from './azureai';
 import { LobeBaichuanAI } from './baichuan';
 import LobeBedrockAI from './bedrock';
 import { LobeCloudflareAI } from './cloudflare';
+import { LobeCohereAI } from './cohere';
 import { LobeDeepSeekAI } from './deepseek';
 import { LobeFireworksAI } from './fireworksai';
 import { LobeGiteeAI } from './giteeai';
@@ -54,6 +55,7 @@ export const providerRuntimeMap = {
   baichuan: LobeBaichuanAI,
   bedrock: LobeBedrockAI,
   cloudflare: LobeCloudflareAI,
+  cohere: LobeCohereAI,
   deepseek: LobeDeepSeekAI,
   doubao: LobeVolcengineAI,
   fireworksai: LobeFireworksAI,

src/libs/agent-runtime/types/type.ts

@@ -30,6 +30,7 @@ export enum ModelProvider {
   Baichuan = 'baichuan',
   Bedrock = 'bedrock',
   Cloudflare = 'cloudflare',
+  Cohere = 'cohere',
   DeepSeek = 'deepseek',
   /**
    * @deprecated

src/types/user/settings/keyVaults.ts

@@ -41,6 +41,7 @@ export interface UserKeyVaults extends SearchEngineKeyVaults {
   baichuan?: OpenAICompatibleKeyVault;
   bedrock?: AWSBedrockKeyVault;
   cloudflare?: CloudflareKeyVault;
+  cohere?: OpenAICompatibleKeyVault;
   deepseek?: OpenAICompatibleKeyVault;
   doubao?: OpenAICompatibleKeyVault;
   fireworksai?: OpenAICompatibleKeyVault;