@hebo-ai/gateway 0.1.2 → 0.2.0

package/src/models/cohere/middleware.ts

@@ -0,0 +1,23 @@
+import type { EmbeddingModelMiddleware } from "ai";
+
+import { modelMiddlewareMatcher } from "../../middleware/matcher";
+
+// Convert `dimensions` (OpenAI) to `outputDimension` (Cohere)
+export const cohereEmbeddingModelMiddleware: EmbeddingModelMiddleware = {
+  specificationVersion: "v3",
+  // eslint-disable-next-line require-await
+  transformParams: async ({ params }) => {
+    const unknown = params.providerOptions?.["unknown"];
+    if (!unknown) return params;
+
+    let dimensions = unknown["dimensions"] as number;
+    if (!dimensions) dimensions = 1024;
+
+    (params.providerOptions!["cohere"] ??= {})["outputDimension"] = dimensions;
+    delete unknown["dimensions"];
+
+    return params;
+  },
+};
+
+modelMiddlewareMatcher.useForModel("cohere/embed-*", { embedding: cohereEmbeddingModelMiddleware });

package/src/models/cohere/presets.ts

@@ -0,0 +1,181 @@
+import type { CanonicalProviderId } from "../../providers/types";
+import type { CanonicalModelId, CatalogModel } from "../types";
+
+import { presetFor, type DeepPartial } from "../../utils/preset";
+
+const COMMAND_BASE = {
+  modalities: {
+    input: ["text"] as const,
+    output: ["text"] as const,
+  },
+  capabilities: ["tool_call", "structured_output", "reasoning", "temperature"] as const,
+  providers: ["cohere"] as const satisfies readonly CanonicalProviderId[],
+  knowledge: "2024-06",
+} satisfies DeepPartial<CatalogModel>;
+
+const COMMAND_VISION_BASE = {
+  modalities: {
+    input: ["text", "image"] as const,
+    output: ["text"] as const,
+  },
+  capabilities: ["structured_output", "reasoning", "temperature"] as const,
+  providers: ["cohere"] as const satisfies readonly CanonicalProviderId[],
+  knowledge: "2024-06",
+} satisfies DeepPartial<CatalogModel>;
+
+const EMBED_V3_BASE = {
+  modalities: {
+    input: ["text", "image"] as const,
+    output: ["embeddings"] as const,
+  },
+  providers: ["cohere", "bedrock"] as const satisfies readonly CanonicalProviderId[],
+} satisfies DeepPartial<CatalogModel>;
+
+const EMBED_V4_BASE = {
+  modalities: {
+    input: ["text", "image", "pdf"] as const,
+    output: ["embeddings"] as const,
+  },
+  providers: ["cohere", "bedrock"] as const satisfies readonly CanonicalProviderId[],
+} satisfies DeepPartial<CatalogModel>;
+
+export const commandA = presetFor<CanonicalModelId, CatalogModel>()("cohere/command-a" as const, {
+  ...COMMAND_BASE,
+  name: "Cohere Command A",
+  created: "2025-03-13",
+  context: 256000,
+} satisfies CatalogModel);
+
+export const commandAReasoning = presetFor<CanonicalModelId, CatalogModel>()(
+  "cohere/command-a-reasoning" as const,
+  {
+    ...COMMAND_BASE,
+    name: "Cohere Command A Reasoning",
+    created: "2025-08-21",
+    context: 256000,
+  } satisfies CatalogModel,
+);
+
+export const commandATranslate = presetFor<CanonicalModelId, CatalogModel>()(
+  "cohere/command-a-translate" as const,
+  {
+    ...COMMAND_BASE,
+    name: "Cohere Command A Translate",
+    created: "2025-08-28",
+    context: 8000,
+  } satisfies CatalogModel,
+);
+
+export const commandAVision = presetFor<CanonicalModelId, CatalogModel>()(
+  "cohere/command-a-vision" as const,
+  {
+    ...COMMAND_VISION_BASE,
+    name: "Cohere Command A Vision",
+    created: "2025-07-31",
+    context: 128000,
+  } satisfies CatalogModel,
+);
+
+export const commandR = presetFor<CanonicalModelId, CatalogModel>()("cohere/command-r" as const, {
+  ...COMMAND_BASE,
+  name: "Cohere Command R",
+  created: "2024-08-01",
+  context: 128000,
+  providers: ["cohere", "bedrock"],
+} satisfies CatalogModel);
+
+export const commandRPlus = presetFor<CanonicalModelId, CatalogModel>()(
+  "cohere/command-r-plus" as const,
+  {
+    ...COMMAND_BASE,
+    name: "Cohere Command R+",
+    created: "2024-08-01",
+    context: 128000,
+    providers: ["cohere", "bedrock"],
+  } satisfies CatalogModel,
+);
+
+export const commandR7b = presetFor<CanonicalModelId, CatalogModel>()(
+  "cohere/command-r7b" as const,
+  {
+    ...COMMAND_BASE,
+    name: "Cohere Command R7B",
+    created: "2024-12-13",
+    context: 128000,
+  } satisfies CatalogModel,
+);
+
+export const embed4 = presetFor<CanonicalModelId, CatalogModel>()("cohere/embed-v4.0" as const, {
+  ...EMBED_V4_BASE,
+  name: "Cohere 4 Embeddings",
+  created: "2025-04-15",
+  context: 128000,
+} satisfies CatalogModel);
+
+export const embedEnglishV3 = presetFor<CanonicalModelId, CatalogModel>()(
+  "cohere/embed-english-v3.0" as const,
+  {
+    ...EMBED_V3_BASE,
+    name: "Cohere Embed English v3",
+    created: "2024-02-07",
+    context: 512,
+  } satisfies CatalogModel,
+);
+
+export const embedEnglishLightV3 = presetFor<CanonicalModelId, CatalogModel>()(
+  "cohere/embed-english-light-v3.0" as const,
+  {
+    ...EMBED_V3_BASE,
+    name: "Cohere Embed English Light v3",
+    created: "2024-02-07",
+    context: 512,
+  } satisfies CatalogModel,
+);
+
+export const embedMultilingualV3 = presetFor<CanonicalModelId, CatalogModel>()(
+  "cohere/embed-multilingual-v3.0" as const,
+  {
+    ...EMBED_V3_BASE,
+    name: "Cohere Embed Multilingual v3",
+    created: "2024-02-07",
+    context: 512,
+  } satisfies CatalogModel,
+);
+
+export const embedMultilingualLightV3 = presetFor<CanonicalModelId, CatalogModel>()(
+  "cohere/embed-multilingual-light-v3.0" as const,
+  {
+    ...EMBED_V3_BASE,
+    name: "Cohere Embed Multilingual Light v3",
+    created: "2024-02-07",
+    context: 512,
+  } satisfies CatalogModel,
+);
+
+const commandAtomic = {
+  A: [commandA, commandAReasoning, commandATranslate, commandAVision],
+  R: [commandR, commandRPlus, commandR7b],
+} as const;
+
+const commandGroups = {} as const;
+
+export const command = {
+  ...commandAtomic,
+  ...commandGroups,
+  latest: [commandA],
+  all: Object.values(commandAtomic).flat(),
+} as const;
+
+const embedAtomic = {
+  v4: [embed4],
+  v3: [embedEnglishV3, embedEnglishLightV3, embedMultilingualV3, embedMultilingualLightV3],
+} as const;
+
+const embedGroups = {} as const;
+
+export const embed = {
+  ...embedAtomic,
+  ...embedGroups,
+  latest: [embed4],
+  all: Object.values(embedAtomic).flat(),
+} as const;

package/src/models/google/index.ts

@@ -0,0 +1,2 @@
+export * from "./presets";
+export * from "./middleware";

package/src/models/google/middleware.ts

@@ -0,0 +1,25 @@
+import type { EmbeddingModelMiddleware } from "ai";
+
+import { modelMiddlewareMatcher } from "../../middleware/matcher";
+
+// Convert `dimensions` (OpenAI) to `outputDimensionality` (Google)
+export const geminiEmbeddingModelMiddleware: EmbeddingModelMiddleware = {
+  specificationVersion: "v3",
+  // eslint-disable-next-line require-await
+  transformParams: async ({ params }) => {
+    const unknown = params.providerOptions?.["unknown"];
+    if (!unknown) return params;
+
+    let dimensions = unknown["dimensions"] as number;
+    if (!dimensions) dimensions = 1024;
+
+    (params.providerOptions!["google"] ??= {})["outputDimensionality"] = dimensions;
+    delete unknown["dimensions"];
+
+    return params;
+  },
+};
+
+modelMiddlewareMatcher.useForModel("google/gemini-*embedding-*", {
+  embedding: geminiEmbeddingModelMiddleware,
+});

package/src/models/{presets/gemini.ts → google/presets.ts}

@@ -19,59 +19,78 @@ const GEMINI_BASE = {
   providers: ["vertex"] as const satisfies readonly CanonicalProviderId[],
 } satisfies DeepPartial<CatalogModel>;
 
+const GEMINI_EMBEDDINGS_BASE = {
+  modalities: {
+    input: ["text"] as const,
+    output: ["embeddings"] as const,
+  },
+  providers: ["vertex"] as const satisfies readonly CanonicalProviderId[],
+} satisfies DeepPartial<CatalogModel>;
+
+export const geminiEmbedding001 = presetFor<CanonicalModelId, CatalogModel>()(
+  "google/embedding-001" as const,
+  {
+    ...GEMINI_EMBEDDINGS_BASE,
+    name: "Gemini Embedding 001",
+    created: "2025-05-20",
+    context: 8192,
+  } satisfies CatalogModel,
+);
+
 export const gemini3FlashPreview = presetFor<CanonicalModelId, CatalogModel>()(
   "google/gemini-3-flash-preview" as const,
   {
+    ...GEMINI_BASE,
     name: "Gemini 3 Flash (Preview)",
     created: "2025-12-17",
     knowledge: "2025-01",
-    ...GEMINI_BASE,
   } satisfies DeepPartial<CatalogModel>,
 );
 
 export const gemini3ProPreview = presetFor<CanonicalModelId, CatalogModel>()(
   "google/gemini-3-pro-preview" as const,
   {
+    ...GEMINI_BASE,
     name: "Gemini 3 Pro (Preview)",
     created: "2025-11-18",
     knowledge: "2025-01",
-    ...GEMINI_BASE,
   } satisfies DeepPartial<CatalogModel>,
 );
 
 export const gemini25FlashLite = presetFor<CanonicalModelId, CatalogModel>()(
   "google/gemini-2.5-flash-lite" as const,
   {
+    ...GEMINI_BASE,
     name: "Gemini 2.5 Flash Lite",
     created: "2025-06-17",
     knowledge: "2025-01",
-    ...GEMINI_BASE,
   } satisfies DeepPartial<CatalogModel>,
 );
 
 export const gemini25Flash = presetFor<CanonicalModelId, CatalogModel>()(
   "google/gemini-2.5-flash" as const,
   {
+    ...GEMINI_BASE,
     name: "Gemini 2.5 Flash",
     created: "2025-03-20",
     knowledge: "2025-01",
-    ...GEMINI_BASE,
   } satisfies DeepPartial<CatalogModel>,
 );
 
 export const gemini25Pro = presetFor<CanonicalModelId, CatalogModel>()(
   "google/gemini-2.5-pro" as const,
   {
+    ...GEMINI_BASE,
     name: "Gemini 2.5 Pro",
     created: "2025-03-20",
     knowledge: "2025-01",
-    ...GEMINI_BASE,
   } satisfies DeepPartial<CatalogModel>,
 );
 
 const geminiAtomic = {
   "v2.5": [gemini25FlashLite, gemini25Flash, gemini25Pro],
   "v3-preview": [gemini3FlashPreview, gemini3ProPreview],
+  embeddings: [geminiEmbedding001],
 } as const;
 
 const geminiGroups = {
@@ -84,5 +103,6 @@ export const gemini = {
   ...geminiGroups,
   latest: [...geminiAtomic["v2.5"]],
   preview: [...geminiAtomic["v3-preview"]],
+  embeddings: [...geminiAtomic["embeddings"]],
   all: Object.values(geminiAtomic).flat(),
 } as const;

package/src/models/meta/index.ts

@@ -0,0 +1 @@
+export * from "./presets";

package/src/models/{presets/llama.ts → meta/presets.ts}

@@ -16,20 +16,80 @@ const LLAMA_3_BASE = {
 export const llama31_8b = presetFor<CanonicalModelId, CatalogModel>()(
   "meta/llama-3.1-8b" as const,
   {
+    ...LLAMA_3_BASE,
     name: "Llama 3.1 8B",
     created: "2024-07-23",
     knowledge: "2023-12",
+  } satisfies CatalogModel,
+);
+
+export const llama31_70b = presetFor<CanonicalModelId, CatalogModel>()(
+  "meta/llama-3.1-70b" as const,
+  {
     ...LLAMA_3_BASE,
+    name: "Llama 3.1 70B",
+    created: "2024-07-23",
+    knowledge: "2023-12",
+  } satisfies CatalogModel,
+);
+
+export const llama31_405b = presetFor<CanonicalModelId, CatalogModel>()(
+  "meta/llama-3.1-405b" as const,
+  {
+    ...LLAMA_3_BASE,
+    name: "Llama 3.1 405B",
+    created: "2024-07-23",
+    knowledge: "2023-12",
+  } satisfies CatalogModel,
+);
+
+export const llama32_1b = presetFor<CanonicalModelId, CatalogModel>()(
+  "meta/llama-3.2-1b" as const,
+  {
+    ...LLAMA_3_BASE,
+    name: "Llama 3.2 1B",
+    created: "2024-09-25",
+    knowledge: "2023-12",
+  } satisfies CatalogModel,
+);
+
+export const llama32_3b = presetFor<CanonicalModelId, CatalogModel>()(
+  "meta/llama-3.2-3b" as const,
+  {
+    ...LLAMA_3_BASE,
+    name: "Llama 3.2 3B",
+    created: "2024-09-25",
+    knowledge: "2023-12",
+  } satisfies CatalogModel,
+);
+
+export const llama32_11b = presetFor<CanonicalModelId, CatalogModel>()(
+  "meta/llama-3.2-11b" as const,
+  {
+    ...LLAMA_3_BASE,
+    name: "Llama 3.2 11B",
+    created: "2024-09-25",
+    knowledge: "2023-12",
+  } satisfies CatalogModel,
+);
+
+export const llama32_90b = presetFor<CanonicalModelId, CatalogModel>()(
+  "meta/llama-3.2-90b" as const,
+  {
+    ...LLAMA_3_BASE,
+    name: "Llama 3.2 90B",
+    created: "2024-09-25",
+    knowledge: "2023-12",
   } satisfies CatalogModel,
 );
 
 export const llama33_70b = presetFor<CanonicalModelId, CatalogModel>()(
   "meta/llama-3.3-70b" as const,
   {
+    ...LLAMA_3_BASE,
     name: "Llama 3.3 70B",
     created: "2024-12-06",
     knowledge: "2023-12",
-    ...LLAMA_3_BASE,
   } satisfies CatalogModel,
 );
 
@@ -38,39 +98,40 @@ const LLAMA_4_BASE = {
     input: ["text", "image", "file"] as const,
     output: ["text"] as const,
   },
-  capabilities: ["attachments", "reasoning", "tool_call", "temperature"] as const,
+  capabilities: ["attachments", "tool_call", "temperature"] as const,
   context: 1000000,
-  providers: ["groq"] as const,
+  providers: ["groq", "vertex", "bedrock"] as const,
 } satisfies DeepPartial<CatalogModel>;
 
 export const llama4Scout = presetFor<CanonicalModelId, CatalogModel>()(
   "meta/llama-4-scout" as const,
   {
+    ...LLAMA_4_BASE,
     name: "Llama 4 Scout",
     created: "2025-08-05",
     knowledge: "2024-06",
-    ...LLAMA_4_BASE,
   } satisfies CatalogModel,
 );
 
 export const llama4Maverick = presetFor<CanonicalModelId, CatalogModel>()(
   "meta/llama-4-maverick" as const,
   {
+    ...LLAMA_4_BASE,
     name: "Llama 4 Maverick",
     created: "2025-08-05",
     knowledge: "2024-06",
-    ...LLAMA_4_BASE,
   } satisfies CatalogModel,
 );
 
 const llamaAtomic = {
-  "v3.1": [llama31_8b],
+  "v3.1": [llama31_8b, llama31_70b, llama31_405b],
+  "v3.2": [llama32_1b, llama32_3b, llama32_11b, llama32_90b],
   "v3.3": [llama33_70b],
   v4: [llama4Scout, llama4Maverick],
 } as const;
 
 const llamaGroups = {
-  "v3.x": [...llamaAtomic["v3.1"], ...llamaAtomic["v3.3"]],
+  "v3.x": [...llamaAtomic["v3.1"], ...llamaAtomic["v3.2"], ...llamaAtomic["v3.3"]],
   "v4.x": [...llamaAtomic["v4"]],
 } as const;
 

package/src/models/openai/index.ts

@@ -0,0 +1,2 @@
+export * from "./presets";
+export * from "./middleware";

package/src/models/openai/middleware.ts

@@ -0,0 +1,25 @@
+import type { EmbeddingModelMiddleware } from "ai";
+
+import { modelMiddlewareMatcher } from "../../middleware/matcher";
+
+// Convert `dimensions` (OpenAI) to `dimensions` (OpenAI)
+export const openAIEmbeddingModelMiddleware: EmbeddingModelMiddleware = {
+  specificationVersion: "v3",
+  // eslint-disable-next-line require-await
+  transformParams: async ({ params }) => {
+    const unknown = params.providerOptions?.["unknown"];
+    if (!unknown) return params;
+
+    let dimensions = unknown["dimensions"] as number;
+    if (!dimensions) dimensions = 1024;
+
+    (params.providerOptions!["openai"] ??= {})["dimensions"] = dimensions;
+    delete unknown["dimensions"];
+
+    return params;
+  },
+};
+
+modelMiddlewareMatcher.useForModel("openai/text-embedding-*", {
+  embedding: openAIEmbeddingModelMiddleware,
+});