open-model-selector 0.1.0 → 0.2.0

package/CHANGELOG.md

@@ -7,6 +7,26 @@ This project adheres to [Semantic Versioning](https://semver.org/).
 
 ## [Unreleased]
 
+## [0.2.0] — 2026-02-20
+
+### Added
+
+- **Multi-provider type resolution** — `defaultModelNormalizer` now uses a 6-tier classification strategy to correctly identify model types across providers that use non-standard vocabulary:
+  1. Direct match against canonical types (Venice)
+  2. Non-text alias mapping (Together AI: `"audio"`→`"tts"`, `"transcribe"`→`"asr"`)
+  3. Architecture-based inference from `output_modalities` (OpenRouter)
+  4. Heuristic inference from model ID patterns
+  5. Text-aliased type (Together AI: `"chat"`→`"text"`, Mistral: `"base"`→`"text"`, Vercel: `"language"`→`"text"`)
+  6. Default to `"text"`
+- **`TYPE_ALIASES` export** — maps provider-specific type strings (`chat`, `language`, `base`, `moderation`, `rerank`, `audio`, `transcribe`) to canonical `ModelType` values; exported from both `"open-model-selector"` and `"open-model-selector/utils"` so consumers can inspect or extend
+- **Extended `MODEL_ID_TYPE_PATTERNS`** — new heuristic patterns: `embedqa`, `imagen`, `gpt-image`, `image-preview`, `kling`
+- **Provider compatibility test suite** (`provider-compat.test.ts`) — 418 lines covering type resolution for Together AI, Vercel AI Gateway, Mistral AI, OpenRouter, OpenAI, Nvidia NIM, Helicone, and Venice AI response shapes, plus `defaultResponseExtractor` wrapper shapes
+- **Live provider Storybook stories** (`provider-live.stories.tsx`) — interactive stories for testing against 12 real provider APIs: OpenAI, OpenRouter, Together, Groq, Cerebras, Nvidia, Mistral, DeepSeek, SambaNova, Venice, Helicone, and Vercel
+- **`.env.example`** — environment variable template for 13 API provider keys used by Storybook live stories and the snapshot capture script
+- **Provider snapshot capture script** (`scripts/capture-provider-snapshots.cjs`) — fetches and stores real provider API responses for test fixture generation
+- **`.env` parser utility** (`scripts/parse-env.cjs`) — shared `.env` file parser used by both Storybook config and the snapshot capture script
+- **Storybook provider proxy support** — `.storybook/main.ts` now loads `.env` keys and configures Vite proxies for cross-origin provider API access during development
+
 ## [0.1.0] — 2026-02-06
 
 Initial release of `open-model-selector`.
@@ -68,5 +88,6 @@ Initial release of `open-model-selector`.
   - **Normalizer**: `ModelNormalizer`, `ResponseExtractor`
 - **Dual CJS/ESM output** — built with tsup, sourcemaps and `.d.ts` included
 
-[Unreleased]: https://github.com/sethbang/open-model-selector/compare/v0.1.0...HEAD
+[Unreleased]: https://github.com/sethbang/open-model-selector/compare/v0.2.0...HEAD
+[0.2.0]: https://github.com/sethbang/open-model-selector/compare/v0.1.0...v0.2.0
 [0.1.0]: https://github.com/sethbang/open-model-selector/releases/tag/v0.1.0

package/README.md

@@ -1,6 +1,6 @@
 # open-model-selector
 
-[![npm version](https://img.shields.io/npm/v/open-model-selector)](https://www.npmjs.com/package/open-model-selector) [![npm bundle size](https://img.shields.io/bundlephobia/minzip/open-model-selector)](https://bundlephobia.com/package/open-model-selector) [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT) [![TypeScript](https://img.shields.io/badge/TypeScript-5.9-blue.svg)](https://www.typescriptlang.org/)
+[![npm version](https://img.shields.io/npm/v/open-model-selector)](https://www.npmjs.com/package/open-model-selector) [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT) [![TypeScript](https://img.shields.io/badge/TypeScript-5.9-blue.svg)](https://www.typescriptlang.org/)
 
 ### An accessible, themeable React model-selector combobox for any OpenAI-compatible API — with first-class [Venice.ai](https://venice.ai) support.
 
@@ -37,13 +37,16 @@
 - **"Use System Default"** sentinel option for fallback behavior
 - **8 model types**: text, image, video, inpaint, embedding, TTS, ASR, upscale
 - **Specialized selectors**: `<TextModelSelector>`, `<ImageModelSelector>`, `<VideoModelSelector>`
-- **Built-in normalizers** for Venice.ai, OpenAI, and OpenRouter response shapes
+- **Built-in normalizers** for Venice.ai, OpenAI, OpenRouter, Together AI, Vercel AI Gateway, Mistral, Groq, Cerebras, Nvidia NIM, DeepSeek, SambaNova, and Helicone response shapes
 - **Scoped CSS** with `--oms-` custom property prefix — never pollutes host styles
 - **Dark mode** via `prefers-color-scheme` and `.dark` class
 - **Full TypeScript types** exported
 - **Dual CJS/ESM output** with sourcemaps
 - **React 18 and 19** support
 
+### Check out the [demo here!](https://sethbang.github.io/open-model-selector-demo/)
+
+
 ## Installation
 
 ### Install the package
@@ -444,7 +447,7 @@ import type { ModelNormalizer, ResponseExtractor } from "open-model-selector"
 
 ### Custom Normalizer
 
-The built-in `defaultModelNormalizer` handles Venice.ai, OpenAI, and OpenRouter response shapes automatically — including Venice's nested `model_spec` format with capabilities, privacy, traits, and deprecation info. If your API returns a different shape, provide a custom normalizer:
+The built-in `defaultModelNormalizer` handles response shapes from Venice.ai, OpenAI, OpenRouter, Together AI, Vercel AI Gateway, Mistral, Groq, Cerebras, Nvidia NIM, DeepSeek, SambaNova, and Helicone automatically — using a [6-tier type classification strategy](#type-inference) that resolves provider-specific vocabulary, architecture metadata, and model ID heuristics. Venice's nested `model_spec` format with capabilities, privacy, traits, and deprecation info is fully supported. If your API returns a different shape, provide a custom normalizer:
 
 ```tsx
 import { ModelSelector } from "open-model-selector"
@@ -631,12 +634,42 @@ const imageModel = normalizeImageModel(rawApiObject) // → ImageModel
 
 #### Type Inference
 
-`inferTypeFromId` uses heuristic pattern matching on a model's ID string to determine its type. This is how `defaultModelNormalizer` classifies models from providers that don't include an explicit `type` field (e.g., OpenAI, OpenRouter):
+`defaultModelNormalizer` uses a **6-tier type resolution strategy** to correctly classify models across providers with different metadata formats:
+
+| Tier | Strategy | Providers |
+|------|----------|-----------|
+| 1 | Direct match against canonical types (`text`, `image`, etc.) | Venice |
+| 2 | Non-text alias mapping (e.g. `"audio"`→`"tts"`, `"transcribe"`→`"asr"`) | Together AI |
+| 3 | Architecture-based inference from `output_modalities` | OpenRouter |
+| 4 | Heuristic pattern matching on model ID | OpenAI, Nvidia, Helicone |
+| 5 | Text-aliased type (e.g. `"chat"`→`"text"`, `"base"`→`"text"`) | Together AI, Mistral, Vercel |
+| 6 | Default to `"text"` | All |
+
+> Text aliases (Tier 5) are checked **after** the ID heuristic (Tier 4) so that e.g. Mistral's `mistral-embed` (type: `"base"`) still gets classified as `"embedding"` via its model ID rather than being swallowed by the `"base"`→`"text"` alias.
+
+**`TYPE_ALIASES`** maps provider-specific type vocabulary to canonical `ModelType` values:
+
+```ts
+import { TYPE_ALIASES } from "open-model-selector/utils"
+
+// TYPE_ALIASES = {
+//   chat: 'text',        // Together AI
+//   language: 'text',    // Vercel AI Gateway, Together AI
+//   base: 'text',        // Mistral AI
+//   moderation: 'text',  // Together AI
+//   rerank: 'text',      // Together AI
+//   audio: 'tts',        // Together AI (Cartesia Sonic etc.)
+//   transcribe: 'asr',   // Together AI (Whisper etc.)
+// }
+```
+
+**`inferTypeFromId`** uses heuristic pattern matching on a model's ID string (Tier 4). This is how `defaultModelNormalizer` classifies models from providers that don't include an explicit `type` field (e.g., OpenAI, Nvidia):
 
 ```ts
 import { inferTypeFromId, MODEL_ID_TYPE_PATTERNS } from "open-model-selector/utils"
 
 inferTypeFromId("dall-e-3")           // "image"
+inferTypeFromId("gpt-image-1")       // "image"
 inferTypeFromId("whisper-large-v3")   // "asr"
 inferTypeFromId("gpt-4o")            // undefined (no match → caller defaults to "text")
 ```
@@ -763,6 +796,7 @@ The component implements the [ARIA combobox pattern](https://www.w3.org/WAI/ARIA
 | `npm run test:watch` | Run tests in watch mode. |
 | `npm run storybook` | Start Storybook dev server on port 6006. |
 | `npm run build-storybook` | Build static Storybook site. |
+| `node scripts/capture-provider-snapshots.cjs` | Capture real provider API responses into `test-fixtures/providers/` for test fixture generation. Requires `.env` with API keys. |
 
 > `npm run prepublishOnly` runs typecheck → test → build automatically before publishing.
 
@@ -780,6 +814,7 @@ Test files:
 - [`src/hooks/use-models.test.tsx`](src/hooks/use-models.test.tsx) — Hook tests
 - [`src/utils/format.test.ts`](src/utils/format.test.ts) — Format utility tests
 - [`src/utils/normalizers/normalizers.test.ts`](src/utils/normalizers/normalizers.test.ts) — Normalizer tests
+- [`src/utils/normalizers/provider-compat.test.ts`](src/utils/normalizers/provider-compat.test.ts) — Provider compatibility tests (Together AI, Vercel, Mistral, OpenRouter, OpenAI, Nvidia, Helicone, Venice)
 
 ```bash
 # Run all tests
@@ -793,7 +828,14 @@ npm run test:watch
 
 The project includes comprehensive Storybook stories covering multiple scenarios: Default, PreselectedModel, SystemDefault, CustomPlaceholder, SortByNewest, ControlledFavorites, EmptyState, LoadingState, ErrorState, PopoverTop, WideContainer, DarkMode, MinimalModels, and VeniceLive.
 
+**Live provider stories** are also available for testing against real APIs (OpenAI, OpenRouter, Together, Groq, Cerebras, Nvidia, Mistral, DeepSeek, SambaNova, Venice, Helicone, Vercel). To use them:
+
+1. Copy `.env.example` to `.env` and fill in the API keys you want to test
+2. Run `npm run storybook` — the Storybook config auto-loads your `.env` keys and configures CORS proxies
+
 ```bash
+cp .env.example .env
+# Edit .env with your API keys
 npm run storybook
 # Opens at http://localhost:6006
 ```

package/dist/index.cjs

@@ -10,11 +10,12 @@ var _reactpopover = require('@radix-ui/react-popover'); var PopoverPrimitive = _
 
 // src/utils/normalizers/type-inference.ts
 var MODEL_ID_TYPE_PATTERNS = [
-  [/\b(embed|embedding)\b/i, "embedding"],
-  [/\b(dall-e|stable-diffusion|sdxl|midjourney|flux)\b/i, "image"],
+  [/\b(embed|embedding|embedqa)\b/i, "embedding"],
+  [/\b(dall-e|stable-diffusion|sdxl|midjourney|flux|imagen)\b/i, "image"],
+  [/\b(gpt-image|image-preview)\b/i, "image"],
   [/\b(tts)\b/i, "tts"],
   [/\b(whisper|asr)\b/i, "asr"],
-  [/\b(sora|video|wan)\b/i, "video"],
+  [/\b(sora|video|wan|kling)\b/i, "video"],
   [/\b(inpaint)\b/i, "inpaint"],
   [/\b(upscale|esrgan)\b/i, "upscale"]
 ];
@@ -277,10 +278,45 @@ function defaultResponseExtractor(body) {
   return [];
 }
 var VALID_TYPES = /* @__PURE__ */ new Set(["text", "image", "video", "inpaint", "embedding", "tts", "asr", "upscale"]);
+var TYPE_ALIASES = {
+  // Together AI
+  chat: "text",
+  language: "text",
+  // Vercel AI Gateway, Together AI
+  base: "text",
+  // Mistral AI
+  moderation: "text",
+  // Together AI
+  rerank: "text",
+  // Together AI
+  audio: "tts",
+  // Together AI (Cartesia Sonic etc.)
+  transcribe: "asr"
+  // Together AI (Whisper etc.)
+  // image, video, embedding already match VALID_TYPES directly
+};
+function inferTypeFromArchitecture(raw) {
+  const arch = raw.architecture;
+  if (!arch) return void 0;
+  const outputMods = arch.output_modalities;
+  if (!Array.isArray(outputMods) || outputMods.length === 0) return void 0;
+  if (outputMods.includes("image") && !outputMods.includes("text")) return "image";
+  if (outputMods.length === 1 && outputMods[0] === "audio") return "tts";
+  return void 0;
+}
 function defaultModelNormalizer(raw) {
   const id = _nullishCoalesce(_nullishCoalesce(raw.id, () => ( raw.model_id)), () => ( ""));
   const rawType = raw.type;
-  const type = _nullishCoalesce(_nullishCoalesce((rawType && VALID_TYPES.has(rawType) ? rawType : void 0), () => ( inferTypeFromId(id))), () => ( "text"));
+  const aliasedType = rawType && rawType in TYPE_ALIASES ? TYPE_ALIASES[rawType] : void 0;
+  const type = (
+    // 1. Direct match against canonical types (Venice)
+    _nullishCoalesce(_nullishCoalesce(_nullishCoalesce(_nullishCoalesce(_nullishCoalesce((rawType && VALID_TYPES.has(rawType) ? rawType : void 0), () => ( // 2. Non-text alias (e.g. "audio"→"tts", "transcribe"→"asr") — these are specific enough to trust
+    (aliasedType && aliasedType !== "text" ? aliasedType : void 0))), () => ( // 3. Architecture-based inference (OpenRouter output_modalities)
+    inferTypeFromArchitecture(raw))), () => ( // 4. Heuristic from model ID patterns
+    inferTypeFromId(id))), () => ( // 5. Text-aliased type (e.g. "chat"→"text", "base"→"text", "language"→"text")
+    aliasedType)), () => ( // 6. Default to 'text'
+    "text"))
+  );
   switch (type) {
     case "text":
       return normalizeTextModel(raw);
@@ -325,10 +361,10 @@ function useModels(props) {
       setError(null);
       return;
     }
-    if (!/^https?:\/\//i.test(baseUrl)) {
+    if (!/^(https?:\/\/|\/(?!\/))/i.test(baseUrl)) {
       setAllModels([]);
       setLoading(false);
-      setError(new Error(`Invalid baseUrl scheme: URL must start with http:// or https://`));
+      setError(new Error(`Invalid baseUrl scheme: URL must start with http://, https://, or /`));
       return;
     }
     const controller = new AbortController();
@@ -1370,4 +1406,5 @@ VideoModelSelector.displayName = "VideoModelSelector";
 
 
 
-exports.ImageModelSelector = ImageModelSelector; exports.MODEL_ID_TYPE_PATTERNS = MODEL_ID_TYPE_PATTERNS; exports.ModelSelector = ModelSelector; exports.SYSTEM_DEFAULT_VALUE = SYSTEM_DEFAULT_VALUE; exports.TextModelSelector = TextModelSelector; exports.VideoModelSelector = VideoModelSelector; exports.defaultModelNormalizer = defaultModelNormalizer; exports.defaultResponseExtractor = defaultResponseExtractor; exports.extractBaseFields = extractBaseFields; exports.formatAspectRatios = formatAspectRatios; exports.formatAudioPrice = formatAudioPrice; exports.formatContextLength = formatContextLength; exports.formatDuration = formatDuration; exports.formatFlatPrice = formatFlatPrice; exports.formatPrice = formatPrice; exports.formatResolutions = formatResolutions; exports.inferTypeFromId = inferTypeFromId; exports.isDeprecated = isDeprecated; exports.normalizeAsrModel = normalizeAsrModel; exports.normalizeEmbeddingModel = normalizeEmbeddingModel; exports.normalizeImageModel = normalizeImageModel; exports.normalizeInpaintModel = normalizeInpaintModel; exports.normalizeTextModel = normalizeTextModel; exports.normalizeTtsModel = normalizeTtsModel; exports.normalizeUpscaleModel = normalizeUpscaleModel; exports.normalizeVideoModel = normalizeVideoModel; exports.toNum = toNum; exports.useModels = useModels;
+
+exports.ImageModelSelector = ImageModelSelector; exports.MODEL_ID_TYPE_PATTERNS = MODEL_ID_TYPE_PATTERNS; exports.ModelSelector = ModelSelector; exports.SYSTEM_DEFAULT_VALUE = SYSTEM_DEFAULT_VALUE; exports.TYPE_ALIASES = TYPE_ALIASES; exports.TextModelSelector = TextModelSelector; exports.VideoModelSelector = VideoModelSelector; exports.defaultModelNormalizer = defaultModelNormalizer; exports.defaultResponseExtractor = defaultResponseExtractor; exports.extractBaseFields = extractBaseFields; exports.formatAspectRatios = formatAspectRatios; exports.formatAudioPrice = formatAudioPrice; exports.formatContextLength = formatContextLength; exports.formatDuration = formatDuration; exports.formatFlatPrice = formatFlatPrice; exports.formatPrice = formatPrice; exports.formatResolutions = formatResolutions; exports.inferTypeFromId = inferTypeFromId; exports.isDeprecated = isDeprecated; exports.normalizeAsrModel = normalizeAsrModel; exports.normalizeEmbeddingModel = normalizeEmbeddingModel; exports.normalizeImageModel = normalizeImageModel; exports.normalizeInpaintModel = normalizeInpaintModel; exports.normalizeTextModel = normalizeTextModel; exports.normalizeTtsModel = normalizeTtsModel; exports.normalizeUpscaleModel = normalizeUpscaleModel; exports.normalizeVideoModel = normalizeVideoModel; exports.toNum = toNum; exports.useModels = useModels;

package/dist/index.d.cts

@@ -177,7 +177,10 @@ declare function extractBaseFields(raw: Record<string, unknown>): Omit<BaseModel
 
 /** Known model ID patterns for heuristic type inference.
  *  Used when the API response lacks an explicit `type` field (non-Venice providers).
- *  Exported so consumers can inspect or extend. */
+ *  Exported so consumers can inspect or extend.
+ *
+ *  Patterns are tested in order — first match wins. More specific patterns should
+ *  come before broader ones (e.g. "gpt-image" before generic "image"). */
 declare const MODEL_ID_TYPE_PATTERNS: Array<[RegExp, ModelType]>;
 /** Infer model type from its ID using naming conventions.
  *  Returns undefined if no pattern matches (caller should fall back to 'text'). */
@@ -219,11 +222,22 @@ type ModelNormalizer = (raw: Record<string, unknown>) => AnyModel;
  *  - `{ models: [...] }` → return models
  *  - Fallback → empty array */
 declare function defaultResponseExtractor(body: Record<string, unknown> | unknown[]): Record<string, unknown>[];
+/** Maps provider-specific type strings to our canonical ModelType values.
+ *  Covers vocabulary differences across Together AI, Vercel AI Gateway, Mistral, etc.
+ *  Exported so consumers can inspect or extend. */
+declare const TYPE_ALIASES: Record<string, ModelType>;
 /** Default dispatching model normalizer.
- *  Uses three-tier type resolution:
- *  1. Explicit `raw.type` field (Venice, custom providers)
- *  2. Heuristic inference from model ID patterns (OpenAI, OpenRouter, etc.)
- *  3. Fallback to 'text' — safe default for most providers */
+ *  Uses multi-tier type resolution:
+ *  1. Explicit `raw.type` field matching our canonical types (Venice, custom providers)
+ *  2. Non-text alias mapping for provider-specific vocabulary (Together AI: "audio"→"tts")
+ *  3. Architecture-based inference from `output_modalities` (OpenRouter)
+ *  4. Heuristic inference from model ID patterns (OpenAI, Nvidia, etc.)
+ *  5. Text-aliased type (Together AI: "chat"→"text", Mistral: "base"→"text")
+ *  6. Fallback to 'text' — safe default for most providers
+ *
+ *  Note: aliases that resolve to 'text' are checked AFTER the ID heuristic (Tier 4)
+ *  so that e.g. Mistral's `mistral-embed` (type: "base") still gets classified as
+ *  'embedding' via its model ID rather than being swallowed by the "base"→"text" alias. */
 declare function defaultModelNormalizer(raw: Record<string, unknown>): AnyModel;
 
 /** A fetch-compatible function signature. Used for SSR, testing, or proxy scenarios. */
@@ -466,4 +480,4 @@ type ImageModelSelectorProps = Omit<ModelSelectorProps, 'type'>;
 /** Props for VideoModelSelector — same as ModelSelectorProps but without `type`. */
 type VideoModelSelectorProps = Omit<ModelSelectorProps, 'type'>;
 
-export { type AnyModel, type AsrModel, type AsrPricing, type BaseModel, type Deprecation, type EmbeddingModel, type EmbeddingPricing, type FetchFn, type ImageConstraints, type ImageModel, ImageModelSelector, type ImageModelSelectorProps, type ImagePricing, type InpaintConstraints, type InpaintModel, type InpaintPricing, MODEL_ID_TYPE_PATTERNS, type ModelNormalizer, ModelSelector, type ModelSelectorProps, type ModelType, type ResponseExtractor, SYSTEM_DEFAULT_VALUE, type TextCapabilities, type TextConstraints, type TextModel, TextModelSelector, type TextModelSelectorProps, type TextPricing, type TtsModel, type TtsPricing, type UpscaleModel, type UpscalePricing, type UseModelsProps, type UseModelsResult, type VideoConstraints, type VideoModel, VideoModelSelector, type VideoModelSelectorProps, defaultModelNormalizer, defaultResponseExtractor, extractBaseFields, formatAspectRatios, formatAudioPrice, formatContextLength, formatDuration, formatFlatPrice, formatPrice, formatResolutions, inferTypeFromId, isDeprecated, normalizeAsrModel, normalizeEmbeddingModel, normalizeImageModel, normalizeInpaintModel, normalizeTextModel, normalizeTtsModel, normalizeUpscaleModel, normalizeVideoModel, toNum, useModels };
+export { type AnyModel, type AsrModel, type AsrPricing, type BaseModel, type Deprecation, type EmbeddingModel, type EmbeddingPricing, type FetchFn, type ImageConstraints, type ImageModel, ImageModelSelector, type ImageModelSelectorProps, type ImagePricing, type InpaintConstraints, type InpaintModel, type InpaintPricing, MODEL_ID_TYPE_PATTERNS, type ModelNormalizer, ModelSelector, type ModelSelectorProps, type ModelType, type ResponseExtractor, SYSTEM_DEFAULT_VALUE, TYPE_ALIASES, type TextCapabilities, type TextConstraints, type TextModel, TextModelSelector, type TextModelSelectorProps, type TextPricing, type TtsModel, type TtsPricing, type UpscaleModel, type UpscalePricing, type UseModelsProps, type UseModelsResult, type VideoConstraints, type VideoModel, VideoModelSelector, type VideoModelSelectorProps, defaultModelNormalizer, defaultResponseExtractor, extractBaseFields, formatAspectRatios, formatAudioPrice, formatContextLength, formatDuration, formatFlatPrice, formatPrice, formatResolutions, inferTypeFromId, isDeprecated, normalizeAsrModel, normalizeEmbeddingModel, normalizeImageModel, normalizeInpaintModel, normalizeTextModel, normalizeTtsModel, normalizeUpscaleModel, normalizeVideoModel, toNum, useModels };

package/dist/index.d.ts

@@ -177,7 +177,10 @@ declare function extractBaseFields(raw: Record<string, unknown>): Omit<BaseModel
 
 /** Known model ID patterns for heuristic type inference.
  *  Used when the API response lacks an explicit `type` field (non-Venice providers).
- *  Exported so consumers can inspect or extend. */
+ *  Exported so consumers can inspect or extend.
+ *
+ *  Patterns are tested in order — first match wins. More specific patterns should
+ *  come before broader ones (e.g. "gpt-image" before generic "image"). */
 declare const MODEL_ID_TYPE_PATTERNS: Array<[RegExp, ModelType]>;
 /** Infer model type from its ID using naming conventions.
  *  Returns undefined if no pattern matches (caller should fall back to 'text'). */
@@ -219,11 +222,22 @@ type ModelNormalizer = (raw: Record<string, unknown>) => AnyModel;
  *  - `{ models: [...] }` → return models
  *  - Fallback → empty array */
 declare function defaultResponseExtractor(body: Record<string, unknown> | unknown[]): Record<string, unknown>[];
+/** Maps provider-specific type strings to our canonical ModelType values.
+ *  Covers vocabulary differences across Together AI, Vercel AI Gateway, Mistral, etc.
+ *  Exported so consumers can inspect or extend. */
+declare const TYPE_ALIASES: Record<string, ModelType>;
 /** Default dispatching model normalizer.
- *  Uses three-tier type resolution:
- *  1. Explicit `raw.type` field (Venice, custom providers)
- *  2. Heuristic inference from model ID patterns (OpenAI, OpenRouter, etc.)
- *  3. Fallback to 'text' — safe default for most providers */
+ *  Uses multi-tier type resolution:
+ *  1. Explicit `raw.type` field matching our canonical types (Venice, custom providers)
+ *  2. Non-text alias mapping for provider-specific vocabulary (Together AI: "audio"→"tts")
+ *  3. Architecture-based inference from `output_modalities` (OpenRouter)
+ *  4. Heuristic inference from model ID patterns (OpenAI, Nvidia, etc.)
+ *  5. Text-aliased type (Together AI: "chat"→"text", Mistral: "base"→"text")
+ *  6. Fallback to 'text' — safe default for most providers
+ *
+ *  Note: aliases that resolve to 'text' are checked AFTER the ID heuristic (Tier 4)
+ *  so that e.g. Mistral's `mistral-embed` (type: "base") still gets classified as
+ *  'embedding' via its model ID rather than being swallowed by the "base"→"text" alias. */
 declare function defaultModelNormalizer(raw: Record<string, unknown>): AnyModel;
 
 /** A fetch-compatible function signature. Used for SSR, testing, or proxy scenarios. */
@@ -466,4 +480,4 @@ type ImageModelSelectorProps = Omit<ModelSelectorProps, 'type'>;
 /** Props for VideoModelSelector — same as ModelSelectorProps but without `type`. */
 type VideoModelSelectorProps = Omit<ModelSelectorProps, 'type'>;
 
-export { type AnyModel, type AsrModel, type AsrPricing, type BaseModel, type Deprecation, type EmbeddingModel, type EmbeddingPricing, type FetchFn, type ImageConstraints, type ImageModel, ImageModelSelector, type ImageModelSelectorProps, type ImagePricing, type InpaintConstraints, type InpaintModel, type InpaintPricing, MODEL_ID_TYPE_PATTERNS, type ModelNormalizer, ModelSelector, type ModelSelectorProps, type ModelType, type ResponseExtractor, SYSTEM_DEFAULT_VALUE, type TextCapabilities, type TextConstraints, type TextModel, TextModelSelector, type TextModelSelectorProps, type TextPricing, type TtsModel, type TtsPricing, type UpscaleModel, type UpscalePricing, type UseModelsProps, type UseModelsResult, type VideoConstraints, type VideoModel, VideoModelSelector, type VideoModelSelectorProps, defaultModelNormalizer, defaultResponseExtractor, extractBaseFields, formatAspectRatios, formatAudioPrice, formatContextLength, formatDuration, formatFlatPrice, formatPrice, formatResolutions, inferTypeFromId, isDeprecated, normalizeAsrModel, normalizeEmbeddingModel, normalizeImageModel, normalizeInpaintModel, normalizeTextModel, normalizeTtsModel, normalizeUpscaleModel, normalizeVideoModel, toNum, useModels };
+export { type AnyModel, type AsrModel, type AsrPricing, type BaseModel, type Deprecation, type EmbeddingModel, type EmbeddingPricing, type FetchFn, type ImageConstraints, type ImageModel, ImageModelSelector, type ImageModelSelectorProps, type ImagePricing, type InpaintConstraints, type InpaintModel, type InpaintPricing, MODEL_ID_TYPE_PATTERNS, type ModelNormalizer, ModelSelector, type ModelSelectorProps, type ModelType, type ResponseExtractor, SYSTEM_DEFAULT_VALUE, TYPE_ALIASES, type TextCapabilities, type TextConstraints, type TextModel, TextModelSelector, type TextModelSelectorProps, type TextPricing, type TtsModel, type TtsPricing, type UpscaleModel, type UpscalePricing, type UseModelsProps, type UseModelsResult, type VideoConstraints, type VideoModel, VideoModelSelector, type VideoModelSelectorProps, defaultModelNormalizer, defaultResponseExtractor, extractBaseFields, formatAspectRatios, formatAudioPrice, formatContextLength, formatDuration, formatFlatPrice, formatPrice, formatResolutions, inferTypeFromId, isDeprecated, normalizeAsrModel, normalizeEmbeddingModel, normalizeImageModel, normalizeInpaintModel, normalizeTextModel, normalizeTtsModel, normalizeUpscaleModel, normalizeVideoModel, toNum, useModels };

package/dist/index.js

@@ -10,11 +10,12 @@ import { useState, useEffect, useRef, useMemo } from "react";
 
 // src/utils/normalizers/type-inference.ts
 var MODEL_ID_TYPE_PATTERNS = [
-  [/\b(embed|embedding)\b/i, "embedding"],
-  [/\b(dall-e|stable-diffusion|sdxl|midjourney|flux)\b/i, "image"],
+  [/\b(embed|embedding|embedqa)\b/i, "embedding"],
+  [/\b(dall-e|stable-diffusion|sdxl|midjourney|flux|imagen)\b/i, "image"],
+  [/\b(gpt-image|image-preview)\b/i, "image"],
   [/\b(tts)\b/i, "tts"],
   [/\b(whisper|asr)\b/i, "asr"],
-  [/\b(sora|video|wan)\b/i, "video"],
+  [/\b(sora|video|wan|kling)\b/i, "video"],
   [/\b(inpaint)\b/i, "inpaint"],
   [/\b(upscale|esrgan)\b/i, "upscale"]
 ];
@@ -277,10 +278,45 @@ function defaultResponseExtractor(body) {
   return [];
 }
 var VALID_TYPES = /* @__PURE__ */ new Set(["text", "image", "video", "inpaint", "embedding", "tts", "asr", "upscale"]);
+var TYPE_ALIASES = {
+  // Together AI
+  chat: "text",
+  language: "text",
+  // Vercel AI Gateway, Together AI
+  base: "text",
+  // Mistral AI
+  moderation: "text",
+  // Together AI
+  rerank: "text",
+  // Together AI
+  audio: "tts",
+  // Together AI (Cartesia Sonic etc.)
+  transcribe: "asr"
+  // Together AI (Whisper etc.)
+  // image, video, embedding already match VALID_TYPES directly
+};
+function inferTypeFromArchitecture(raw) {
+  const arch = raw.architecture;
+  if (!arch) return void 0;
+  const outputMods = arch.output_modalities;
+  if (!Array.isArray(outputMods) || outputMods.length === 0) return void 0;
+  if (outputMods.includes("image") && !outputMods.includes("text")) return "image";
+  if (outputMods.length === 1 && outputMods[0] === "audio") return "tts";
+  return void 0;
+}
 function defaultModelNormalizer(raw) {
   const id = raw.id ?? raw.model_id ?? "";
   const rawType = raw.type;
-  const type = (rawType && VALID_TYPES.has(rawType) ? rawType : void 0) ?? inferTypeFromId(id) ?? "text";
+  const aliasedType = rawType && rawType in TYPE_ALIASES ? TYPE_ALIASES[rawType] : void 0;
+  const type = (
+    // 1. Direct match against canonical types (Venice)
+    (rawType && VALID_TYPES.has(rawType) ? rawType : void 0) ?? // 2. Non-text alias (e.g. "audio"→"tts", "transcribe"→"asr") — these are specific enough to trust
+    (aliasedType && aliasedType !== "text" ? aliasedType : void 0) ?? // 3. Architecture-based inference (OpenRouter output_modalities)
+    inferTypeFromArchitecture(raw) ?? // 4. Heuristic from model ID patterns
+    inferTypeFromId(id) ?? // 5. Text-aliased type (e.g. "chat"→"text", "base"→"text", "language"→"text")
+    aliasedType ?? // 6. Default to 'text'
+    "text"
+  );
   switch (type) {
     case "text":
       return normalizeTextModel(raw);
@@ -325,10 +361,10 @@ function useModels(props) {
       setError(null);
       return;
     }
-    if (!/^https?:\/\//i.test(baseUrl)) {
+    if (!/^(https?:\/\/|\/(?!\/))/i.test(baseUrl)) {
       setAllModels([]);
       setLoading(false);
-      setError(new Error(`Invalid baseUrl scheme: URL must start with http:// or https://`));
+      setError(new Error(`Invalid baseUrl scheme: URL must start with http://, https://, or /`));
       return;
     }
     const controller = new AbortController();
@@ -1346,6 +1382,7 @@ export {
   MODEL_ID_TYPE_PATTERNS,
   ModelSelector,
   SYSTEM_DEFAULT_VALUE,
+  TYPE_ALIASES,
   TextModelSelector,
   VideoModelSelector,
   defaultModelNormalizer,

package/dist/utils.cjs

@@ -1,10 +1,11 @@
 "use strict";Object.defineProperty(exports, "__esModule", {value: true}); function _nullishCoalesce(lhs, rhsFn) { if (lhs != null) { return lhs; } else { return rhsFn(); } } function _optionalChain(ops) { let lastAccessLHS = undefined; let value = ops[0]; let i = 1; while (i < ops.length) { const op = ops[i]; const fn = ops[i + 1]; i += 2; if ((op === 'optionalAccess' || op === 'optionalCall') && value == null) { return undefined; } if (op === 'access' || op === 'optionalAccess') { lastAccessLHS = value; value = fn(value); } else if (op === 'call' || op === 'optionalCall') { value = fn((...args) => value.call(lastAccessLHS, ...args)); lastAccessLHS = undefined; } } return value; }// src/utils/normalizers/type-inference.ts
 var MODEL_ID_TYPE_PATTERNS = [
-  [/\b(embed|embedding)\b/i, "embedding"],
-  [/\b(dall-e|stable-diffusion|sdxl|midjourney|flux)\b/i, "image"],
+  [/\b(embed|embedding|embedqa)\b/i, "embedding"],
+  [/\b(dall-e|stable-diffusion|sdxl|midjourney|flux|imagen)\b/i, "image"],
+  [/\b(gpt-image|image-preview)\b/i, "image"],
   [/\b(tts)\b/i, "tts"],
   [/\b(whisper|asr)\b/i, "asr"],
-  [/\b(sora|video|wan)\b/i, "video"],
+  [/\b(sora|video|wan|kling)\b/i, "video"],
   [/\b(inpaint)\b/i, "inpaint"],
   [/\b(upscale|esrgan)\b/i, "upscale"]
 ];
@@ -267,10 +268,45 @@ function defaultResponseExtractor(body) {
   return [];
 }
 var VALID_TYPES = /* @__PURE__ */ new Set(["text", "image", "video", "inpaint", "embedding", "tts", "asr", "upscale"]);
+var TYPE_ALIASES = {
+  // Together AI
+  chat: "text",
+  language: "text",
+  // Vercel AI Gateway, Together AI
+  base: "text",
+  // Mistral AI
+  moderation: "text",
+  // Together AI
+  rerank: "text",
+  // Together AI
+  audio: "tts",
+  // Together AI (Cartesia Sonic etc.)
+  transcribe: "asr"
+  // Together AI (Whisper etc.)
+  // image, video, embedding already match VALID_TYPES directly
+};
+function inferTypeFromArchitecture(raw) {
+  const arch = raw.architecture;
+  if (!arch) return void 0;
+  const outputMods = arch.output_modalities;
+  if (!Array.isArray(outputMods) || outputMods.length === 0) return void 0;
+  if (outputMods.includes("image") && !outputMods.includes("text")) return "image";
+  if (outputMods.length === 1 && outputMods[0] === "audio") return "tts";
+  return void 0;
+}
 function defaultModelNormalizer(raw) {
   const id = _nullishCoalesce(_nullishCoalesce(raw.id, () => ( raw.model_id)), () => ( ""));
   const rawType = raw.type;
-  const type = _nullishCoalesce(_nullishCoalesce((rawType && VALID_TYPES.has(rawType) ? rawType : void 0), () => ( inferTypeFromId(id))), () => ( "text"));
+  const aliasedType = rawType && rawType in TYPE_ALIASES ? TYPE_ALIASES[rawType] : void 0;
+  const type = (
+    // 1. Direct match against canonical types (Venice)
+    _nullishCoalesce(_nullishCoalesce(_nullishCoalesce(_nullishCoalesce(_nullishCoalesce((rawType && VALID_TYPES.has(rawType) ? rawType : void 0), () => ( // 2. Non-text alias (e.g. "audio"→"tts", "transcribe"→"asr") — these are specific enough to trust
+    (aliasedType && aliasedType !== "text" ? aliasedType : void 0))), () => ( // 3. Architecture-based inference (OpenRouter output_modalities)
+    inferTypeFromArchitecture(raw))), () => ( // 4. Heuristic from model ID patterns
+    inferTypeFromId(id))), () => ( // 5. Text-aliased type (e.g. "chat"→"text", "base"→"text", "language"→"text")
+    aliasedType)), () => ( // 6. Default to 'text'
+    "text"))
+  );
   switch (type) {
     case "text":
       return normalizeTextModel(raw);
@@ -368,4 +404,5 @@ function formatAspectRatios(ratios) {
 
 
 
-exports.MODEL_ID_TYPE_PATTERNS = MODEL_ID_TYPE_PATTERNS; exports.defaultModelNormalizer = defaultModelNormalizer; exports.defaultResponseExtractor = defaultResponseExtractor; exports.extractBaseFields = extractBaseFields; exports.formatAspectRatios = formatAspectRatios; exports.formatAudioPrice = formatAudioPrice; exports.formatContextLength = formatContextLength; exports.formatDuration = formatDuration; exports.formatFlatPrice = formatFlatPrice; exports.formatPrice = formatPrice; exports.formatResolutions = formatResolutions; exports.inferTypeFromId = inferTypeFromId; exports.isDeprecated = isDeprecated; exports.normalizeAsrModel = normalizeAsrModel; exports.normalizeEmbeddingModel = normalizeEmbeddingModel; exports.normalizeImageModel = normalizeImageModel; exports.normalizeInpaintModel = normalizeInpaintModel; exports.normalizeTextModel = normalizeTextModel; exports.normalizeTtsModel = normalizeTtsModel; exports.normalizeUpscaleModel = normalizeUpscaleModel; exports.normalizeVideoModel = normalizeVideoModel; exports.toNum = toNum;
+
+exports.MODEL_ID_TYPE_PATTERNS = MODEL_ID_TYPE_PATTERNS; exports.TYPE_ALIASES = TYPE_ALIASES; exports.defaultModelNormalizer = defaultModelNormalizer; exports.defaultResponseExtractor = defaultResponseExtractor; exports.extractBaseFields = extractBaseFields; exports.formatAspectRatios = formatAspectRatios; exports.formatAudioPrice = formatAudioPrice; exports.formatContextLength = formatContextLength; exports.formatDuration = formatDuration; exports.formatFlatPrice = formatFlatPrice; exports.formatPrice = formatPrice; exports.formatResolutions = formatResolutions; exports.inferTypeFromId = inferTypeFromId; exports.isDeprecated = isDeprecated; exports.normalizeAsrModel = normalizeAsrModel; exports.normalizeEmbeddingModel = normalizeEmbeddingModel; exports.normalizeImageModel = normalizeImageModel; exports.normalizeInpaintModel = normalizeInpaintModel; exports.normalizeTextModel = normalizeTextModel; exports.normalizeTtsModel = normalizeTtsModel; exports.normalizeUpscaleModel = normalizeUpscaleModel; exports.normalizeVideoModel = normalizeVideoModel; exports.toNum = toNum;

package/dist/utils.d.cts

@@ -175,7 +175,10 @@ declare function extractBaseFields(raw: Record<string, unknown>): Omit<BaseModel
 
 /** Known model ID patterns for heuristic type inference.
  *  Used when the API response lacks an explicit `type` field (non-Venice providers).
- *  Exported so consumers can inspect or extend. */
+ *  Exported so consumers can inspect or extend.
+ *
+ *  Patterns are tested in order — first match wins. More specific patterns should
+ *  come before broader ones (e.g. "gpt-image" before generic "image"). */
 declare const MODEL_ID_TYPE_PATTERNS: Array<[RegExp, ModelType]>;
 /** Infer model type from its ID using naming conventions.
  *  Returns undefined if no pattern matches (caller should fall back to 'text'). */
@@ -217,11 +220,22 @@ type ModelNormalizer = (raw: Record<string, unknown>) => AnyModel;
  *  - `{ models: [...] }` → return models
  *  - Fallback → empty array */
 declare function defaultResponseExtractor(body: Record<string, unknown> | unknown[]): Record<string, unknown>[];
+/** Maps provider-specific type strings to our canonical ModelType values.
+ *  Covers vocabulary differences across Together AI, Vercel AI Gateway, Mistral, etc.
+ *  Exported so consumers can inspect or extend. */
+declare const TYPE_ALIASES: Record<string, ModelType>;
 /** Default dispatching model normalizer.
- *  Uses three-tier type resolution:
- *  1. Explicit `raw.type` field (Venice, custom providers)
- *  2. Heuristic inference from model ID patterns (OpenAI, OpenRouter, etc.)
- *  3. Fallback to 'text' — safe default for most providers */
+ *  Uses multi-tier type resolution:
+ *  1. Explicit `raw.type` field matching our canonical types (Venice, custom providers)
+ *  2. Non-text alias mapping for provider-specific vocabulary (Together AI: "audio"→"tts")
+ *  3. Architecture-based inference from `output_modalities` (OpenRouter)
+ *  4. Heuristic inference from model ID patterns (OpenAI, Nvidia, etc.)
+ *  5. Text-aliased type (Together AI: "chat"→"text", Mistral: "base"→"text")
+ *  6. Fallback to 'text' — safe default for most providers
+ *
+ *  Note: aliases that resolve to 'text' are checked AFTER the ID heuristic (Tier 4)
+ *  so that e.g. Mistral's `mistral-embed` (type: "base") still gets classified as
+ *  'embedding' via its model ID rather than being swallowed by the "base"→"text" alias. */
 declare function defaultModelNormalizer(raw: Record<string, unknown>): AnyModel;
 
 /** Check whether a deprecation date is in the past.
@@ -301,4 +315,4 @@ declare function formatResolutions(resolutions: string[] | undefined | null): st
  */
 declare function formatAspectRatios(ratios: string[] | undefined | null): string;
 
-export { type AnyModel, type AsrModel, type AsrPricing, type BaseModel, type Deprecation, type EmbeddingModel, type EmbeddingPricing, type ImageConstraints, type ImageModel, type ImagePricing, type InpaintConstraints, type InpaintModel, type InpaintPricing, MODEL_ID_TYPE_PATTERNS, type ModelNormalizer, type ModelType, type ResponseExtractor, type TextCapabilities, type TextConstraints, type TextModel, type TextPricing, type TtsModel, type TtsPricing, type UpscaleModel, type UpscalePricing, type VideoConstraints, type VideoModel, defaultModelNormalizer, defaultResponseExtractor, extractBaseFields, formatAspectRatios, formatAudioPrice, formatContextLength, formatDuration, formatFlatPrice, formatPrice, formatResolutions, inferTypeFromId, isDeprecated, normalizeAsrModel, normalizeEmbeddingModel, normalizeImageModel, normalizeInpaintModel, normalizeTextModel, normalizeTtsModel, normalizeUpscaleModel, normalizeVideoModel, toNum };
+export { type AnyModel, type AsrModel, type AsrPricing, type BaseModel, type Deprecation, type EmbeddingModel, type EmbeddingPricing, type ImageConstraints, type ImageModel, type ImagePricing, type InpaintConstraints, type InpaintModel, type InpaintPricing, MODEL_ID_TYPE_PATTERNS, type ModelNormalizer, type ModelType, type ResponseExtractor, TYPE_ALIASES, type TextCapabilities, type TextConstraints, type TextModel, type TextPricing, type TtsModel, type TtsPricing, type UpscaleModel, type UpscalePricing, type VideoConstraints, type VideoModel, defaultModelNormalizer, defaultResponseExtractor, extractBaseFields, formatAspectRatios, formatAudioPrice, formatContextLength, formatDuration, formatFlatPrice, formatPrice, formatResolutions, inferTypeFromId, isDeprecated, normalizeAsrModel, normalizeEmbeddingModel, normalizeImageModel, normalizeInpaintModel, normalizeTextModel, normalizeTtsModel, normalizeUpscaleModel, normalizeVideoModel, toNum };

package/dist/utils.d.ts

@@ -175,7 +175,10 @@ declare function extractBaseFields(raw: Record<string, unknown>): Omit<BaseModel
 
 /** Known model ID patterns for heuristic type inference.
  *  Used when the API response lacks an explicit `type` field (non-Venice providers).
- *  Exported so consumers can inspect or extend. */
+ *  Exported so consumers can inspect or extend.
+ *
+ *  Patterns are tested in order — first match wins. More specific patterns should
+ *  come before broader ones (e.g. "gpt-image" before generic "image"). */
 declare const MODEL_ID_TYPE_PATTERNS: Array<[RegExp, ModelType]>;
 /** Infer model type from its ID using naming conventions.
  *  Returns undefined if no pattern matches (caller should fall back to 'text'). */
@@ -217,11 +220,22 @@ type ModelNormalizer = (raw: Record<string, unknown>) => AnyModel;
  *  - `{ models: [...] }` → return models
  *  - Fallback → empty array */
 declare function defaultResponseExtractor(body: Record<string, unknown> | unknown[]): Record<string, unknown>[];
+/** Maps provider-specific type strings to our canonical ModelType values.
+ *  Covers vocabulary differences across Together AI, Vercel AI Gateway, Mistral, etc.
+ *  Exported so consumers can inspect or extend. */
+declare const TYPE_ALIASES: Record<string, ModelType>;
 /** Default dispatching model normalizer.
- *  Uses three-tier type resolution:
- *  1. Explicit `raw.type` field (Venice, custom providers)
- *  2. Heuristic inference from model ID patterns (OpenAI, OpenRouter, etc.)
- *  3. Fallback to 'text' — safe default for most providers */
+ *  Uses multi-tier type resolution:
+ *  1. Explicit `raw.type` field matching our canonical types (Venice, custom providers)
+ *  2. Non-text alias mapping for provider-specific vocabulary (Together AI: "audio"→"tts")
+ *  3. Architecture-based inference from `output_modalities` (OpenRouter)
+ *  4. Heuristic inference from model ID patterns (OpenAI, Nvidia, etc.)
+ *  5. Text-aliased type (Together AI: "chat"→"text", Mistral: "base"→"text")
+ *  6. Fallback to 'text' — safe default for most providers
+ *
+ *  Note: aliases that resolve to 'text' are checked AFTER the ID heuristic (Tier 4)
+ *  so that e.g. Mistral's `mistral-embed` (type: "base") still gets classified as
+ *  'embedding' via its model ID rather than being swallowed by the "base"→"text" alias. */
 declare function defaultModelNormalizer(raw: Record<string, unknown>): AnyModel;
 
 /** Check whether a deprecation date is in the past.
@@ -301,4 +315,4 @@ declare function formatResolutions(resolutions: string[] | undefined | null): st
  */
 declare function formatAspectRatios(ratios: string[] | undefined | null): string;
 
-export { type AnyModel, type AsrModel, type AsrPricing, type BaseModel, type Deprecation, type EmbeddingModel, type EmbeddingPricing, type ImageConstraints, type ImageModel, type ImagePricing, type InpaintConstraints, type InpaintModel, type InpaintPricing, MODEL_ID_TYPE_PATTERNS, type ModelNormalizer, type ModelType, type ResponseExtractor, type TextCapabilities, type TextConstraints, type TextModel, type TextPricing, type TtsModel, type TtsPricing, type UpscaleModel, type UpscalePricing, type VideoConstraints, type VideoModel, defaultModelNormalizer, defaultResponseExtractor, extractBaseFields, formatAspectRatios, formatAudioPrice, formatContextLength, formatDuration, formatFlatPrice, formatPrice, formatResolutions, inferTypeFromId, isDeprecated, normalizeAsrModel, normalizeEmbeddingModel, normalizeImageModel, normalizeInpaintModel, normalizeTextModel, normalizeTtsModel, normalizeUpscaleModel, normalizeVideoModel, toNum };
+export { type AnyModel, type AsrModel, type AsrPricing, type BaseModel, type Deprecation, type EmbeddingModel, type EmbeddingPricing, type ImageConstraints, type ImageModel, type ImagePricing, type InpaintConstraints, type InpaintModel, type InpaintPricing, MODEL_ID_TYPE_PATTERNS, type ModelNormalizer, type ModelType, type ResponseExtractor, TYPE_ALIASES, type TextCapabilities, type TextConstraints, type TextModel, type TextPricing, type TtsModel, type TtsPricing, type UpscaleModel, type UpscalePricing, type VideoConstraints, type VideoModel, defaultModelNormalizer, defaultResponseExtractor, extractBaseFields, formatAspectRatios, formatAudioPrice, formatContextLength, formatDuration, formatFlatPrice, formatPrice, formatResolutions, inferTypeFromId, isDeprecated, normalizeAsrModel, normalizeEmbeddingModel, normalizeImageModel, normalizeInpaintModel, normalizeTextModel, normalizeTtsModel, normalizeUpscaleModel, normalizeVideoModel, toNum };

package/dist/utils.js

@@ -1,10 +1,11 @@
 // src/utils/normalizers/type-inference.ts
 var MODEL_ID_TYPE_PATTERNS = [
-  [/\b(embed|embedding)\b/i, "embedding"],
-  [/\b(dall-e|stable-diffusion|sdxl|midjourney|flux)\b/i, "image"],
+  [/\b(embed|embedding|embedqa)\b/i, "embedding"],
+  [/\b(dall-e|stable-diffusion|sdxl|midjourney|flux|imagen)\b/i, "image"],
+  [/\b(gpt-image|image-preview)\b/i, "image"],
   [/\b(tts)\b/i, "tts"],
   [/\b(whisper|asr)\b/i, "asr"],
-  [/\b(sora|video|wan)\b/i, "video"],
+  [/\b(sora|video|wan|kling)\b/i, "video"],
   [/\b(inpaint)\b/i, "inpaint"],
   [/\b(upscale|esrgan)\b/i, "upscale"]
 ];
@@ -267,10 +268,45 @@ function defaultResponseExtractor(body) {
   return [];
 }
 var VALID_TYPES = /* @__PURE__ */ new Set(["text", "image", "video", "inpaint", "embedding", "tts", "asr", "upscale"]);
+var TYPE_ALIASES = {
+  // Together AI
+  chat: "text",
+  language: "text",
+  // Vercel AI Gateway, Together AI
+  base: "text",
+  // Mistral AI
+  moderation: "text",
+  // Together AI
+  rerank: "text",
+  // Together AI
+  audio: "tts",
+  // Together AI (Cartesia Sonic etc.)
+  transcribe: "asr"
+  // Together AI (Whisper etc.)
+  // image, video, embedding already match VALID_TYPES directly
+};
+function inferTypeFromArchitecture(raw) {
+  const arch = raw.architecture;
+  if (!arch) return void 0;
+  const outputMods = arch.output_modalities;
+  if (!Array.isArray(outputMods) || outputMods.length === 0) return void 0;
+  if (outputMods.includes("image") && !outputMods.includes("text")) return "image";
+  if (outputMods.length === 1 && outputMods[0] === "audio") return "tts";
+  return void 0;
+}
 function defaultModelNormalizer(raw) {
   const id = raw.id ?? raw.model_id ?? "";
   const rawType = raw.type;
-  const type = (rawType && VALID_TYPES.has(rawType) ? rawType : void 0) ?? inferTypeFromId(id) ?? "text";
+  const aliasedType = rawType && rawType in TYPE_ALIASES ? TYPE_ALIASES[rawType] : void 0;
+  const type = (
+    // 1. Direct match against canonical types (Venice)
+    (rawType && VALID_TYPES.has(rawType) ? rawType : void 0) ?? // 2. Non-text alias (e.g. "audio"→"tts", "transcribe"→"asr") — these are specific enough to trust
+    (aliasedType && aliasedType !== "text" ? aliasedType : void 0) ?? // 3. Architecture-based inference (OpenRouter output_modalities)
+    inferTypeFromArchitecture(raw) ?? // 4. Heuristic from model ID patterns
+    inferTypeFromId(id) ?? // 5. Text-aliased type (e.g. "chat"→"text", "base"→"text", "language"→"text")
+    aliasedType ?? // 6. Default to 'text'
+    "text"
+  );
   switch (type) {
     case "text":
       return normalizeTextModel(raw);
@@ -347,6 +383,7 @@ function formatAspectRatios(ratios) {
 }
 export {
   MODEL_ID_TYPE_PATTERNS,
+  TYPE_ALIASES,
   defaultModelNormalizer,
   defaultResponseExtractor,
   extractBaseFields,

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "open-model-selector",
-    "version": "0.1.0",
+    "version": "0.2.0",
     "description": "A generic, OpenAI-compatible model selector component for React.",
     "main": "./dist/index.cjs",
     "module": "./dist/index.js",