@tanstack/ai-vercel-gateway 0.2.11 → 0.2.14

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tanstack/ai-vercel-gateway",
-  "version": "0.2.11",
+  "version": "0.2.14",
   "description": "Vercel AI Gateway adapter for TanStack AI chat, embeddings, and image generation.",
   "author": "Tanner Linsley",
   "license": "MIT",
@@ -56,12 +56,12 @@
     "vite": "^8.2.1"
   },
   "peerDependencies": {
-    "@tanstack/ai": "^0.55.0"
+    "@tanstack/ai": "^0.57.0"
   },
   "dependencies": {
     "openai": "^6.41.0",
     "@tanstack/ai-utils": "^0.4.0",
-    "@tanstack/openai-base": "^0.10.12"
+    "@tanstack/openai-base": "^0.10.14"
   },
   "scripts": {
     "build": "vite build",

package/src/adapters/evaluate.ts

@@ -0,0 +1,393 @@
+import { BaseEvaluateAdapter } from '@tanstack/ai/adapters'
+import { toRunErrorPayload } from '@tanstack/ai/adapter-internals'
+import {
+  getVercelGatewayApiKeyFromEnv,
+  withVercelGatewayDefaults,
+} from '../utils/client'
+import { mapGatewayModelOptions } from '../utils/map-gateway-options'
+import type {
+  EvaluateAdapterResult,
+  EvaluateOptions,
+  WireAnswer,
+  WireQuestion,
+} from '@tanstack/ai/adapters'
+import type { VercelGatewayClientConfig } from '../utils/client'
+import type { VercelGatewayRoutingOptions } from '../text/text-provider-options'
+
+export interface VercelGatewayEvaluateConfig extends VercelGatewayClientConfig {}
+
+export type VercelGatewayEvaluateModel = 'typesafe-ai/jev'
+
+export type VercelGatewayEvaluateProviderOptions = Record<string, unknown> & {
+  gateway?: VercelGatewayRoutingOptions
+}
+
+/**
+ * Documented evaluate URL from Vercel AI SDK `GatewayEvaluationModel.getUrl()`:
+ * `${baseURL}/evaluation-model` with default baseURL
+ * `https://ai-gateway.vercel.sh/v4/ai`. Not `/v1/chat/completions`.
+ * https://github.com/vercel/ai/blob/main/packages/gateway/src/gateway-evaluation-model.ts
+ */
+const EVALUATE_PATH = '/v4/ai/evaluation-model'
+const DEFAULT_EVALUATE_URL = `https://ai-gateway.vercel.sh${EVALUATE_PATH}`
+
+/** The gateway rejects the request with 400 when this header is absent. */
+const GATEWAY_PROTOCOL_VERSION = '0.0.1'
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === 'object' && value !== null && !Array.isArray(value)
+}
+
+function isNumberRecord(value: unknown): value is Record<string, number> {
+  if (!isRecord(value)) return false
+  const values = Object.values(value)
+  for (const item of values) {
+    if (typeof item !== 'number') return false
+  }
+  return true
+}
+
+function isStringRecord(value: unknown): value is Record<string, string> {
+  if (!isRecord(value)) return false
+  const values = Object.values(value)
+  for (const item of values) {
+    if (typeof item !== 'string') return false
+  }
+  return true
+}
+
+function extraRequestHeaders(defaultHeaders: unknown) {
+  if (defaultHeaders == null) return {}
+  if (typeof Headers !== 'undefined' && defaultHeaders instanceof Headers) {
+    const headers: Record<string, string> = {}
+    const entries = defaultHeaders.entries()
+    for (const [key, value] of entries) {
+      headers[key] = value
+    }
+    return headers
+  }
+  if (Array.isArray(defaultHeaders)) {
+    const headers: Record<string, string> = {}
+    for (const entry of defaultHeaders) {
+      if (!Array.isArray(entry) || entry.length < 2) continue
+      const key = entry[0]
+      const value = entry[1]
+      if (typeof key === 'string' && typeof value === 'string') {
+        headers[key] = value
+      }
+    }
+    return headers
+  }
+  if (!isRecord(defaultHeaders)) return {}
+  const headers: Record<string, string> = {}
+  const keys = Object.keys(defaultHeaders)
+  for (const key of keys) {
+    const value = defaultHeaders[key]
+    if (typeof value === 'string') headers[key] = value
+  }
+  return headers
+}
+
+function resolveEvaluateUrl(baseURL: string | null | undefined) {
+  if (!baseURL) return DEFAULT_EVALUATE_URL
+  try {
+    return `${new URL(baseURL).origin}${EVALUATE_PATH}`
+  } catch {
+    return DEFAULT_EVALUATE_URL
+  }
+}
+
+function toGatewayQuestion(question: WireQuestion) {
+  switch (question.type) {
+    case 'noul': {
+      if (question.criteria === undefined) {
+        return {
+          type: 'boolean' as const,
+          instructions: question.instructions,
+        }
+      }
+      return {
+        type: 'boolean' as const,
+        instructions: question.instructions,
+        criteria: question.criteria,
+      }
+    }
+    case 'choice':
+    case 'score':
+      return question
+  }
+}
+
+function toGatewayQuestions(questions: Record<string, WireQuestion>) {
+  const mapped: Record<string, unknown> = {}
+  const keys = Object.keys(questions)
+  for (const key of keys) {
+    const question = questions[key]
+    if (question === undefined) continue
+    mapped[key] = toGatewayQuestion(question)
+  }
+  return mapped
+}
+
+function readCount(candidates: Array<unknown>) {
+  for (const candidate of candidates) {
+    if (typeof candidate === 'number') return candidate
+  }
+  return 0
+}
+
+function toTokenUsage(usage: unknown) {
+  if (!isRecord(usage)) {
+    return { promptTokens: 0, completionTokens: 0, totalTokens: 0 }
+  }
+  const promptTokens = readCount([
+    usage.inputTokens,
+    usage.input_tokens,
+    usage.promptTokens,
+    usage.prompt_tokens,
+  ])
+  const completionTokens = readCount([
+    usage.outputTokens,
+    usage.output_tokens,
+    usage.completionTokens,
+    usage.completion_tokens,
+  ])
+  const totalTokens = readCount([
+    usage.totalTokens,
+    usage.total_tokens,
+    promptTokens + completionTokens,
+  ])
+  return { promptTokens, completionTokens, totalTokens }
+}
+
+function confidenceMap(providerMetadata: unknown) {
+  if (!isRecord(providerMetadata)) return {}
+  const typesafe = providerMetadata.typesafe
+  if (!isRecord(typesafe)) return {}
+  if (!isNumberRecord(typesafe.confidence)) return {}
+  return typesafe.confidence
+}
+
+function confidenceFor(
+  key: string,
+  answer: Record<string, unknown>,
+  byKey: Record<string, number>,
+) {
+  if (typeof answer.confidence === 'number') return answer.confidence
+  const fromMeta = byKey[key]
+  if (typeof fromMeta === 'number') return fromMeta
+  return 0
+}
+
+function toWireAnswer(answer: Record<string, unknown>, confidence: number) {
+  switch (answer.type) {
+    case 'boolean': {
+      if (typeof answer.probability !== 'number') {
+        throw new Error(
+          'Vercel Gateway evaluate boolean answer was missing probability',
+        )
+      }
+      return { type: 'noul' as const, noul: answer.probability }
+    }
+    case 'choice': {
+      if (typeof answer.choice !== 'string') {
+        throw new Error(
+          'Vercel Gateway evaluate choice answer was missing choice',
+        )
+      }
+      return {
+        type: 'choice' as const,
+        choice: answer.choice,
+        probabilities: isNumberRecord(answer.probabilities)
+          ? answer.probabilities
+          : {},
+        confidence,
+      }
+    }
+    case 'score': {
+      if (typeof answer.score !== 'number') {
+        throw new Error(
+          'Vercel Gateway evaluate score answer was missing score',
+        )
+      }
+      return {
+        type: 'score' as const,
+        score: answer.score,
+        legend: isStringRecord(answer.legend) ? answer.legend : {},
+        probabilities: isNumberRecord(answer.probabilities)
+          ? answer.probabilities
+          : {},
+        confidence,
+      }
+    }
+    default:
+      throw new Error(
+        `Vercel Gateway evaluate answer had an unexpected type: ${String(answer.type)}`,
+      )
+  }
+}
+
+function toWireAnswers(answers: unknown, byKey: Record<string, number>) {
+  if (!isRecord(answers)) {
+    throw new Error('Vercel Gateway evaluate response was missing answers')
+  }
+  const mapped: Record<string, WireAnswer> = {}
+  const keys = Object.keys(answers)
+  for (const key of keys) {
+    const answer = answers[key]
+    if (!isRecord(answer)) {
+      throw new Error(
+        `Vercel Gateway evaluate answer "${key}" had an unexpected shape`,
+      )
+    }
+    mapped[key] = toWireAnswer(answer, confidenceFor(key, answer, byKey))
+  }
+  return mapped
+}
+
+/**
+ * Vercel AI Gateway evaluate adapter.
+ *
+ * Talks to `POST /v4/ai/evaluation-model` with `fetch`. Jev is not a chat
+ * model; this adapter does not use `/v1/chat/completions`.
+ */
+export class VercelGatewayEvaluateAdapter<
+  TModel extends VercelGatewayEvaluateModel,
+> extends BaseEvaluateAdapter<TModel, VercelGatewayEvaluateProviderOptions> {
+  readonly name = 'vercel-gateway' as const
+
+  private readonly apiKey: string
+  private readonly evaluateUrl: string
+  private readonly extraHeaders: Record<string, string>
+
+  constructor(config: VercelGatewayEvaluateConfig, model: TModel) {
+    super({}, model)
+    const defaults = withVercelGatewayDefaults(config)
+    this.apiKey = config.apiKey
+    this.evaluateUrl = resolveEvaluateUrl(defaults.baseURL)
+    this.extraHeaders = extraRequestHeaders(defaults.defaultHeaders)
+  }
+
+  async evaluate(
+    options: EvaluateOptions<VercelGatewayEvaluateProviderOptions>,
+  ) {
+    const { model, state, questions, modelOptions, abortSignal, logger } =
+      options
+    const mapped = mapGatewayModelOptions(modelOptions)
+
+    logger.request(`activity=evaluate provider=${this.name} model=${model}`, {
+      provider: this.name,
+      model,
+    })
+
+    try {
+      const response = await fetch(this.evaluateUrl, {
+        method: 'POST',
+        headers: {
+          ...this.extraHeaders,
+          Authorization: `Bearer ${this.apiKey}`,
+          'Content-Type': 'application/json',
+          'ai-gateway-protocol-version': GATEWAY_PROTOCOL_VERSION,
+          'ai-evaluation-model-specification-version': '4',
+          'ai-model-id': model,
+        },
+        body: JSON.stringify({
+          ...mapped,
+          model,
+          state,
+          questions: toGatewayQuestions(questions),
+        }),
+        ...(abortSignal ? { signal: abortSignal } : {}),
+      })
+
+      if (!response.ok) {
+        const detail = await response.text().catch(() => '')
+        throw new Error(
+          `Vercel Gateway evaluate request failed: ${response.status} ${response.statusText}${
+            detail ? ` — ${detail}` : ''
+          }`,
+        )
+      }
+
+      const json: unknown = await response.json()
+      if (!isRecord(json)) {
+        throw new Error(
+          'Vercel Gateway evaluate response had an unexpected shape',
+        )
+      }
+
+      const result: EvaluateAdapterResult = {
+        model: typeof json.model === 'string' ? json.model : model,
+        answers: toWireAnswers(
+          json.answers,
+          confidenceMap(json.providerMetadata),
+        ),
+        usage: toTokenUsage(json.usage),
+      }
+      return result
+    } catch (error: unknown) {
+      logger.errors(`${this.name}.evaluate fatal`, {
+        error: toRunErrorPayload(error, `${this.name}.evaluate failed`),
+        source: `${this.name}.evaluate`,
+      })
+      throw error
+    }
+  }
+}
+
+/**
+ * Create a Vercel AI Gateway evaluate adapter with an explicit API key.
+ *
+ * @param model Evaluate model id. Use `typesafe-ai/jev`.
+ * @param apiKey Vercel AI Gateway API key.
+ * @param config Optional client config (`baseURL`, `httpReferer`, `xTitle`).
+ *
+ * @example
+ * ```ts
+ * const adapter = createVercelGatewayDecider('typesafe-ai/jev', 'vck_...')
+ * ```
+ */
+export function createVercelGatewayDecider<
+  TModel extends VercelGatewayEvaluateModel,
+>(
+  model: TModel,
+  apiKey: string,
+  config?: Omit<VercelGatewayEvaluateConfig, 'apiKey'>,
+) {
+  return new VercelGatewayEvaluateAdapter({ apiKey, ...config }, model)
+}
+
+/**
+ * Create a Vercel AI Gateway evaluate adapter.
+ *
+ * Reads `AI_GATEWAY_API_KEY`, then `VERCEL_OIDC_TOKEN`.
+ *
+ * @param model Evaluate model id. Use `typesafe-ai/jev`.
+ * @param config Optional client config (`baseURL`, `httpReferer`, `xTitle`).
+ *
+ * @example
+ * ```ts
+ * import { decide, boolean } from '@tanstack/ai'
+ * import { vercelGatewayDecider } from '@tanstack/ai-vercel-gateway'
+ *
+ * const result = await decide({
+ *   adapter: vercelGatewayDecider('typesafe-ai/jev'),
+ *   state: ticket,
+ *   questions: {
+ *     refund: boolean({
+ *       instructions: 'Is the customer asking for a refund?',
+ *     }),
+ *   },
+ * })
+ * ```
+ */
+export function vercelGatewayDecider<TModel extends VercelGatewayEvaluateModel>(
+  model: TModel,
+  config?: Omit<VercelGatewayEvaluateConfig, 'apiKey'>,
+) {
+  return createVercelGatewayDecider(
+    model,
+    getVercelGatewayApiKeyFromEnv(),
+    config,
+  )
+}

package/src/index.ts

@@ -47,6 +47,13 @@ export {
   type VercelGatewayImageConfig,
 } from './adapters/image'
 
+export {
+  VercelGatewayEvaluateAdapter,
+  createVercelGatewayDecider,
+  vercelGatewayDecider,
+  type VercelGatewayEvaluateConfig,
+} from './adapters/evaluate'
+
 export type { VercelGatewayEmbeddingProviderOptions } from './embedding/embedding-provider-options'
 export type { VercelGatewayImageProviderOptions } from './image/image-provider-options'
 export type {

package/src/model-meta.ts

@@ -124,10 +124,6 @@ export const VERCEL_GATEWAY_CHAT_MODELS = [
   'inference-net/schematron-v2-small',
   'inference-net/schematron-v2-turbo',
   'interfaze/interfaze-beta',
-  'kwaipilot/kat-coder-air-v2.5',
-  'kwaipilot/kat-coder-pro-v1',
-  'kwaipilot/kat-coder-pro-v2',
-  'kwaipilot/kat-coder-pro-v2.5',
   'meta/llama-3.1-70b',
   'meta/llama-3.1-8b',
   'meta/llama-3.3-70b',
@@ -274,6 +270,7 @@ export const VERCEL_GATEWAY_CHAT_MODELS = [
   'tencent/hy4-preview',
   'thinkingmachines/inkling',
   'thinkingmachines/inkling-small',
+  'typesafe-ai/jev',
   'xiaomi/mimo-v2.5',
   'xiaomi/mimo-v2.5-pro',
   'zai/glm-4.5',
@@ -312,7 +309,6 @@ export const VERCEL_GATEWAY_PROVIDERS = [
   'inference-net',
   'interfaze',
   'klingai',
-  'kwaipilot',
   'meta',
   'minimax',
   'mistral',
@@ -330,6 +326,7 @@ export const VERCEL_GATEWAY_PROVIDERS = [
   'stepfun',
   'tencent',
   'thinkingmachines',
+  'typesafe-ai',
   'voyage',
   'xiaomi',
   'zai',
@@ -1333,41 +1330,6 @@ export type VercelGatewayChatModelProviderOptionsByName = {
       | 'reasoning'
       | 'include_reasoning'
     >
-  'kwaipilot/kat-coder-air-v2.5': VercelGatewayCommonOptions &
-    Pick<
-      VercelGatewayBaseOptions,
-      | 'max_tokens'
-      | 'max_output_tokens'
-      | 'temperature'
-      | 'stop'
-      | 'reasoning'
-      | 'include_reasoning'
-    >
-  'kwaipilot/kat-coder-pro-v1': VercelGatewayCommonOptions &
-    Pick<
-      VercelGatewayBaseOptions,
-      'max_tokens' | 'max_output_tokens' | 'temperature' | 'stop'
-    >
-  'kwaipilot/kat-coder-pro-v2': VercelGatewayCommonOptions &
-    Pick<
-      VercelGatewayBaseOptions,
-      | 'max_tokens'
-      | 'max_output_tokens'
-      | 'temperature'
-      | 'stop'
-      | 'reasoning'
-      | 'include_reasoning'
-    >
-  'kwaipilot/kat-coder-pro-v2.5': VercelGatewayCommonOptions &
-    Pick<
-      VercelGatewayBaseOptions,
-      | 'max_tokens'
-      | 'max_output_tokens'
-      | 'temperature'
-      | 'stop'
-      | 'reasoning'
-      | 'include_reasoning'
-    >
   'meta/llama-3.1-70b': VercelGatewayCommonOptions &
     Pick<
       VercelGatewayBaseOptions,
@@ -2533,6 +2495,7 @@ export type VercelGatewayChatModelProviderOptionsByName = {
       | 'reasoning'
       | 'include_reasoning'
     >
+  'typesafe-ai/jev': VercelGatewayCommonOptions
   'xiaomi/mimo-v2.5': VercelGatewayCommonOptions &
     Pick<
       VercelGatewayBaseOptions,
@@ -2836,10 +2799,6 @@ export type VercelGatewayModelInputModalitiesByName = {
   'inference-net/schematron-v2-small': readonly ['text']
   'inference-net/schematron-v2-turbo': readonly ['text']
   'interfaze/interfaze-beta': readonly ['text', 'image', 'document']
-  'kwaipilot/kat-coder-air-v2.5': readonly ['text', 'image']
-  'kwaipilot/kat-coder-pro-v1': readonly ['text']
-  'kwaipilot/kat-coder-pro-v2': readonly ['text']
-  'kwaipilot/kat-coder-pro-v2.5': readonly ['text', 'image']
   'meta/llama-3.1-70b': readonly ['text']
   'meta/llama-3.1-8b': readonly ['text']
   'meta/llama-3.3-70b': readonly ['text']
@@ -2869,7 +2828,7 @@ export type VercelGatewayModelInputModalitiesByName = {
   'mistral/mistral-small': readonly ['text', 'image']
   'moonshotai/kimi-k2': readonly ['text']
   'moonshotai/kimi-k2-thinking': readonly ['text']
-  'moonshotai/kimi-k2.5': readonly ['text', 'image', 'video']
+  'moonshotai/kimi-k2.5': readonly ['text', 'image']
   'moonshotai/kimi-k2.6': readonly ['text', 'image', 'video']
   'moonshotai/kimi-k2.7-code': readonly ['text', 'image', 'document', 'video']
   'moonshotai/kimi-k2.7-code-highspeed': readonly [
@@ -2995,6 +2954,7 @@ export type VercelGatewayModelInputModalitiesByName = {
   'tencent/hy4-preview': readonly ['text']
   'thinkingmachines/inkling': readonly ['text', 'image', 'document']
   'thinkingmachines/inkling-small': readonly ['text', 'image', 'document']
+  'typesafe-ai/jev': readonly ['text']
   'xiaomi/mimo-v2.5': readonly ['text', 'image']
   'xiaomi/mimo-v2.5-pro': readonly ['text']
   'zai/glm-4.5': readonly ['text']