workers-ai-provider 3.2.1 → 3.3.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +28 -19
- package/dist/anthropic.d.mts +1 -1
- package/dist/{gateway-delegate-BfaUTwDZ.d.mts → gateway-delegate-D_zkIp5r.d.mts} +90 -15
- package/dist/{gateway-provider-1USFWm7c.mjs → gateway-provider-CQU-v2IO.mjs} +562 -170
- package/dist/gateway-provider-CQU-v2IO.mjs.map +1 -0
- package/dist/gateway-provider.mjs +1 -1
- package/dist/google.d.mts +1 -1
- package/dist/index.d.mts +1 -1
- package/dist/index.mjs +348 -284
- package/dist/index.mjs.map +1 -1
- package/dist/openai.d.mts +1 -1
- package/package.json +5 -4
- package/src/gateway-delegate.ts +318 -89
- package/src/gateway-provider.ts +10 -34
- package/src/gateway-providers.ts +13 -455
- package/src/index.ts +64 -20
- package/src/resumable-stream.ts +9 -221
- package/src/streaming.ts +1 -41
- package/src/utils.ts +40 -125
- package/src/workersai-chat-language-model.ts +37 -16
- package/src/workersai-embedding-model.ts +22 -11
- package/src/workersai-error.ts +70 -0
- package/src/workersai-image-model.ts +19 -11
- package/src/workersai-reranking-model.ts +16 -5
- package/src/workersai-speech-model.ts +35 -13
- package/src/workersai-transcription-model.ts +15 -4
- package/dist/gateway-provider-1USFWm7c.mjs.map +0 -1
package/src/gateway-providers.ts
CHANGED
|
@@ -1,457 +1,15 @@
|
|
|
1
1
|
/**
|
|
2
|
-
*
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
* - **Slug delegate** (`wai("openai/gpt-5")`): `resolverKey` is the slug prefix
|
|
7
|
-
* the user types. `runCatalog` providers dispatch through the resumable run
|
|
8
|
-
* path (`env.AI.run`, unified billing, `cf-aig-run-id`); the rest go through
|
|
9
|
-
* the gateway path (`env.AI.gateway().run`, BYOK, no resume). `wireFormat`
|
|
10
|
-
* selects the built-in `@ai-sdk/*` parser; absent ⇒ the provider is reachable
|
|
11
|
-
* only via the bring-your-own-provider wrapper (it isn't chat/completions
|
|
12
|
-
* shaped, e.g. audio/image providers).
|
|
13
|
-
* - **Bring-your-own-provider** (`createGatewayProvider`): `hostPattern` +
|
|
14
|
-
* `transformEndpoint` map a wrapped provider's request URL to the gateway
|
|
15
|
-
* `provider` id + endpoint path.
|
|
16
|
-
*
|
|
17
|
-
* Slugs mirror the AI Gateway provider directory
|
|
18
|
-
* (developers.cloudflare.com/ai-gateway/usage/providers/); endpoint transforms
|
|
19
|
-
* mirror `ai-gateway-provider`'s provider table. `runCatalog` / `billing` flags
|
|
20
|
-
* follow the documented unified-billing list (OpenAI, Anthropic, Google AI
|
|
21
|
-
* Studio, Google Vertex, xAI/Grok, Groq) and are otherwise conservative — the
|
|
22
|
-
* e2e suite confirms them live, since resume is undocumented upstream.
|
|
2
|
+
* The AI Gateway provider registry now lives in `@cloudflare/gateway-core`
|
|
3
|
+
* (shared with the other Cloudflare AI Gateway packages). This module re-exports
|
|
4
|
+
* it so the existing `workers-ai-provider/src/gateway-providers` import path
|
|
5
|
+
* keeps working unchanged.
|
|
23
6
|
*/
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
29
|
-
|
|
30
|
-
|
|
31
|
-
|
|
32
|
-
|
|
33
|
-
* Slug prefix the user types in `wai("<resolverKey>/<model>")`. For
|
|
34
|
-
* `runCatalog` providers this is also the run-catalog author (so
|
|
35
|
-
* `env.AI.run("<resolverKey>/<model>")` resolves).
|
|
36
|
-
*/
|
|
37
|
-
resolverKey: string;
|
|
38
|
-
/** Provider id for the gateway universal endpoint (`env.AI.gateway().run([{ provider }])`). */
|
|
39
|
-
gatewayProviderId: string;
|
|
40
|
-
/**
|
|
41
|
-
* Built-in parser wire format. `openai` covers the whole OpenAI-compatible
|
|
42
|
-
* long tail (deepseek, grok, groq, mistral, perplexity, …). Absent ⇒ reachable
|
|
43
|
-
* only via the bring-your-own-provider wrapper (provider-native, non-chat, or a
|
|
44
|
-
* gateway-path URL shape we don't reproduce reliably from the slug delegate).
|
|
45
|
-
*/
|
|
46
|
-
wireFormat?: WireFormat;
|
|
47
|
-
/**
|
|
48
|
-
* Wire format the unified-billing **run path** (`env.AI.run`) emits for this
|
|
49
|
-
* provider — which is NOT always the provider's native format. Cloudflare's
|
|
50
|
-
* unified catalog normalizes most providers to OpenAI chat-completions (so
|
|
51
|
-
* `google` is parsed with the `openai` plugin on the run path), but passes
|
|
52
|
-
* **Anthropic through natively** (`content[].text`, native tool shape), so
|
|
53
|
-
* anthropic must be parsed with the `anthropic` plugin. Defaults to `"openai"`
|
|
54
|
-
* for run-catalog providers when omitted. Only meaningful when `runCatalog`.
|
|
55
|
-
*/
|
|
56
|
-
runWireFormat?: WireFormat;
|
|
57
|
-
/**
|
|
58
|
-
* Base URL the wire-format builder should target so the request URL it
|
|
59
|
-
* generates host-strips (via {@link transformEndpoint}) to the provider's
|
|
60
|
-
* gateway-native endpoint. Omit to use the `@ai-sdk` provider's default (the
|
|
61
|
-
* provider's own host — correct for `openai`/`anthropic`/`google`). Required
|
|
62
|
-
* for OpenAI-wire providers that share the `openai` plugin but live on a
|
|
63
|
-
* different host (deepseek, grok, groq, mistral, perplexity, …).
|
|
64
|
-
*/
|
|
65
|
-
baseURL?: string;
|
|
66
|
-
/** On the unified-billing resumable run catalog (`env.AI.run`, `cf-aig-run-id`). */
|
|
67
|
-
runCatalog: boolean;
|
|
68
|
-
/**
|
|
69
|
-
* Whether the provider has a gateway path (`env.AI.gateway().run`). `false` ⇒
|
|
70
|
-
* **run-path only**: the provider is on the unified run catalog but is not a
|
|
71
|
-
* native gateway provider, so caching, server-side fallback, and
|
|
72
|
-
* `transport: "gateway"` are unavailable and the delegate rejects them with a
|
|
73
|
-
* clear error (rather than failing upstream). Defaults to `true`.
|
|
74
|
-
*/
|
|
75
|
-
gatewayPath?: boolean;
|
|
76
|
-
/** Billing model when reached through the gateway. */
|
|
77
|
-
billing: Billing;
|
|
78
|
-
/** Header(s) carrying the upstream provider key (stripped on the gateway path unless BYOK-forwarded). */
|
|
79
|
-
authHeaders: string[];
|
|
80
|
-
/** Host matcher for bring-your-own-provider URL detection. */
|
|
81
|
-
hostPattern?: RegExp;
|
|
82
|
-
/** Strip the provider host, leaving the gateway endpoint path (+ query). */
|
|
83
|
-
transformEndpoint?: (url: string) => string;
|
|
84
|
-
}
|
|
85
|
-
|
|
86
|
-
/** Strip a leading `https://<host>/` prefix, leaving the endpoint path + query. */
|
|
87
|
-
function hostStrip(pattern: RegExp): (url: string) => string {
|
|
88
|
-
return (url: string) => url.replace(pattern, "");
|
|
89
|
-
}
|
|
90
|
-
|
|
91
|
-
const OPENAI_HOST = /^https:\/\/api\.openai\.com\//;
|
|
92
|
-
const ANTHROPIC_HOST = /^https:\/\/api\.anthropic\.com\//;
|
|
93
|
-
const GOOGLE_HOST = /^https:\/\/generativelanguage\.googleapis\.com\//;
|
|
94
|
-
const VERTEX_HOST = /^https:\/\/(?:[a-z0-9-]+-)?aiplatform\.googleapis\.com\//;
|
|
95
|
-
const XAI_HOST = /^https:\/\/api\.x\.ai\//;
|
|
96
|
-
const GROQ_HOST = /^https:\/\/api\.groq\.com\/openai\/v1\//;
|
|
97
|
-
const DEEPSEEK_HOST = /^https:\/\/api\.deepseek\.com\//;
|
|
98
|
-
const MISTRAL_HOST = /^https:\/\/api\.mistral\.ai\//;
|
|
99
|
-
const PERPLEXITY_HOST = /^https:\/\/api\.perplexity\.ai\//;
|
|
100
|
-
const CEREBRAS_HOST = /^https:\/\/api\.cerebras\.ai\//;
|
|
101
|
-
const OPENROUTER_HOST = /^https:\/\/openrouter\.ai\/api\//;
|
|
102
|
-
const FIREWORKS_HOST = /^https:\/\/api\.fireworks\.ai\/inference\/v1\//;
|
|
103
|
-
const COHERE_HOST = /^https:\/\/api\.cohere\.(?:com|ai)\//;
|
|
104
|
-
const REPLICATE_HOST = /^https:\/\/api\.replicate\.com\//;
|
|
105
|
-
const HUGGINGFACE_HOST = /^https:\/\/api-inference\.huggingface\.co\/models\//;
|
|
106
|
-
const CARTESIA_HOST = /^https:\/\/api\.cartesia\.ai\//;
|
|
107
|
-
const FAL_HOST = /^https:\/\/fal\.run\//;
|
|
108
|
-
const IDEOGRAM_HOST = /^https:\/\/api\.ideogram\.ai\//;
|
|
109
|
-
const DEEPGRAM_HOST = /^https:\/\/api\.deepgram\.com\//;
|
|
110
|
-
const ELEVENLABS_HOST = /^https:\/\/api\.elevenlabs\.io\//;
|
|
111
|
-
const GROK_KEY = "grok";
|
|
112
|
-
|
|
113
|
-
// Bedrock's URL carries the AWS region, which the gateway endpoint preserves as
|
|
114
|
-
// `bedrock-runtime/<region>/<rest>` (mirrors ai-gateway-provider).
|
|
115
|
-
const BEDROCK_HOST = /^https:\/\/bedrock-runtime\.(?<region>[^.]+)\.amazonaws\.com\//;
|
|
116
|
-
function bedrockTransform(url: string): string {
|
|
117
|
-
const m = url.match(
|
|
118
|
-
/^https:\/\/bedrock-runtime\.(?<region>[^.]+)\.amazonaws\.com\/(?<rest>.*)$/,
|
|
119
|
-
);
|
|
120
|
-
if (!m?.groups) return url;
|
|
121
|
-
const { region, rest } = m.groups;
|
|
122
|
-
if (!region || rest === undefined) return url;
|
|
123
|
-
return `bedrock-runtime/${region}/${rest}`;
|
|
124
|
-
}
|
|
125
|
-
|
|
126
|
-
// Azure's URL carries the resource + deployment, so it needs a bespoke transform
|
|
127
|
-
// (mirrors ai-gateway-provider). Only used for bring-your-own-provider detection.
|
|
128
|
-
const AZURE_HOST =
|
|
129
|
-
/^https:\/\/(?<resource>[^.]+)\.openai\.azure\.com\/openai\/deployments\/(?<deployment>[^/]+)\/(?<rest>.*)$/;
|
|
130
|
-
function azureTransform(url: string): string {
|
|
131
|
-
const m = url.match(AZURE_HOST);
|
|
132
|
-
if (!m?.groups) return url;
|
|
133
|
-
const { resource, deployment, rest } = m.groups;
|
|
134
|
-
if (!resource || !deployment || !rest) return url;
|
|
135
|
-
return `${resource}/${deployment}/${rest}`;
|
|
136
|
-
}
|
|
137
|
-
|
|
138
|
-
/**
|
|
139
|
-
* The provider table. Order matters only for `detectProviderByUrl` (first match
|
|
140
|
-
* wins); slugs are looked up by `resolverKey`.
|
|
141
|
-
*/
|
|
142
|
-
export const GATEWAY_PROVIDERS: GatewayProviderInfo[] = [
|
|
143
|
-
// ---- Unified-billing run-catalog providers (resumable run path) ----
|
|
144
|
-
{
|
|
145
|
-
resolverKey: "openai",
|
|
146
|
-
gatewayProviderId: "openai",
|
|
147
|
-
wireFormat: "openai",
|
|
148
|
-
runCatalog: true,
|
|
149
|
-
billing: "unified",
|
|
150
|
-
authHeaders: ["authorization"],
|
|
151
|
-
hostPattern: OPENAI_HOST,
|
|
152
|
-
transformEndpoint: hostStrip(OPENAI_HOST),
|
|
153
|
-
},
|
|
154
|
-
{
|
|
155
|
-
resolverKey: "anthropic",
|
|
156
|
-
gatewayProviderId: "anthropic",
|
|
157
|
-
wireFormat: "anthropic",
|
|
158
|
-
// Unified billing passes Anthropic through natively (unlike google, which it
|
|
159
|
-
// normalizes to openai-wire), so the run path also speaks Anthropic Messages.
|
|
160
|
-
runWireFormat: "anthropic",
|
|
161
|
-
runCatalog: true,
|
|
162
|
-
billing: "unified",
|
|
163
|
-
authHeaders: ["x-api-key", "authorization"],
|
|
164
|
-
hostPattern: ANTHROPIC_HOST,
|
|
165
|
-
transformEndpoint: hostStrip(ANTHROPIC_HOST),
|
|
166
|
-
},
|
|
167
|
-
{
|
|
168
|
-
resolverKey: "google",
|
|
169
|
-
gatewayProviderId: "google-ai-studio",
|
|
170
|
-
// Gateway path hits Gemini's native endpoint (google-wire); the unified run
|
|
171
|
-
// path, however, returns openai-wire — so runWireFormat defaults to "openai".
|
|
172
|
-
wireFormat: "google",
|
|
173
|
-
runCatalog: true,
|
|
174
|
-
billing: "unified",
|
|
175
|
-
authHeaders: ["x-goog-api-key", "authorization"],
|
|
176
|
-
hostPattern: GOOGLE_HOST,
|
|
177
|
-
transformEndpoint: hostStrip(GOOGLE_HOST),
|
|
178
|
-
},
|
|
179
|
-
{
|
|
180
|
-
resolverKey: "xai",
|
|
181
|
-
gatewayProviderId: GROK_KEY,
|
|
182
|
-
wireFormat: "openai",
|
|
183
|
-
// Targeted so a forced gateway-path request host-strips correctly (the run
|
|
184
|
-
// path, the default for xai, ignores this).
|
|
185
|
-
baseURL: "https://api.x.ai/v1",
|
|
186
|
-
runCatalog: true,
|
|
187
|
-
billing: "unified",
|
|
188
|
-
authHeaders: ["authorization"],
|
|
189
|
-
hostPattern: XAI_HOST,
|
|
190
|
-
transformEndpoint: hostStrip(XAI_HOST),
|
|
191
|
-
},
|
|
192
|
-
{
|
|
193
|
-
resolverKey: "groq",
|
|
194
|
-
gatewayProviderId: "groq",
|
|
195
|
-
wireFormat: "openai",
|
|
196
|
-
// Groq's gateway-native endpoint strips `/openai/v1/`, so the builder must
|
|
197
|
-
// target that base or a forced gateway request doubles the prefix.
|
|
198
|
-
baseURL: "https://api.groq.com/openai/v1",
|
|
199
|
-
runCatalog: true,
|
|
200
|
-
billing: "unified",
|
|
201
|
-
authHeaders: ["authorization"],
|
|
202
|
-
hostPattern: GROQ_HOST,
|
|
203
|
-
transformEndpoint: hostStrip(GROQ_HOST),
|
|
204
|
-
},
|
|
205
|
-
// Unified-catalog chat providers that are NOT in the native gateway directory:
|
|
206
|
-
// they exist only on the resumable run path (env.AI.run, unified billing), so
|
|
207
|
-
// there's no BYOK gateway path. Both return OpenAI chat-completions wire (so the
|
|
208
|
-
// `openai` plugin parses them) and emit `cf-aig-run-id` on streams (resumable),
|
|
209
|
-
// verified live against the default gateway. Forcing transport:"gateway" for
|
|
210
|
-
// these errors upstream (no native provider) — that's expected.
|
|
211
|
-
{
|
|
212
|
-
// Alibaba Qwen, served via DashScope's OpenAI-compatible endpoint.
|
|
213
|
-
resolverKey: "alibaba",
|
|
214
|
-
gatewayProviderId: "alibaba",
|
|
215
|
-
wireFormat: "openai",
|
|
216
|
-
runCatalog: true,
|
|
217
|
-
gatewayPath: false,
|
|
218
|
-
billing: "unified",
|
|
219
|
-
authHeaders: ["authorization"],
|
|
220
|
-
},
|
|
221
|
-
{
|
|
222
|
-
// MiniMax (M-series). OpenAI-wire with extra fields (reasoning_content,
|
|
223
|
-
// audio_content) the openai parser ignores; core choices[].delta.content is standard.
|
|
224
|
-
resolverKey: "minimax",
|
|
225
|
-
gatewayProviderId: "minimax",
|
|
226
|
-
wireFormat: "openai",
|
|
227
|
-
runCatalog: true,
|
|
228
|
-
gatewayPath: false,
|
|
229
|
-
billing: "unified",
|
|
230
|
-
authHeaders: ["authorization"],
|
|
231
|
-
},
|
|
232
|
-
{
|
|
233
|
-
resolverKey: "google-vertex",
|
|
234
|
-
gatewayProviderId: "google-vertex-ai",
|
|
235
|
-
// Vertex's URL carries project/location/publisher segments that the
|
|
236
|
-
// `@ai-sdk/google` default (AI Studio) does not produce, so the slug
|
|
237
|
-
// delegate can't shape it reliably — reach Vertex via createGatewayProvider.
|
|
238
|
-
runCatalog: false,
|
|
239
|
-
billing: "unified",
|
|
240
|
-
authHeaders: ["authorization"],
|
|
241
|
-
hostPattern: VERTEX_HOST,
|
|
242
|
-
transformEndpoint: hostStrip(VERTEX_HOST),
|
|
243
|
-
},
|
|
244
|
-
|
|
245
|
-
// ---- OpenAI-compatible long tail (gateway path, BYOK) ----
|
|
246
|
-
{
|
|
247
|
-
resolverKey: "deepseek",
|
|
248
|
-
gatewayProviderId: "deepseek",
|
|
249
|
-
wireFormat: "openai",
|
|
250
|
-
baseURL: "https://api.deepseek.com",
|
|
251
|
-
runCatalog: false,
|
|
252
|
-
billing: "byok",
|
|
253
|
-
authHeaders: ["authorization"],
|
|
254
|
-
hostPattern: DEEPSEEK_HOST,
|
|
255
|
-
transformEndpoint: hostStrip(DEEPSEEK_HOST),
|
|
256
|
-
},
|
|
257
|
-
{
|
|
258
|
-
resolverKey: "mistral",
|
|
259
|
-
gatewayProviderId: "mistral",
|
|
260
|
-
wireFormat: "openai",
|
|
261
|
-
baseURL: "https://api.mistral.ai/v1",
|
|
262
|
-
runCatalog: false,
|
|
263
|
-
billing: "byok",
|
|
264
|
-
authHeaders: ["authorization"],
|
|
265
|
-
hostPattern: MISTRAL_HOST,
|
|
266
|
-
transformEndpoint: hostStrip(MISTRAL_HOST),
|
|
267
|
-
},
|
|
268
|
-
{
|
|
269
|
-
resolverKey: "perplexity",
|
|
270
|
-
gatewayProviderId: "perplexity-ai",
|
|
271
|
-
wireFormat: "openai",
|
|
272
|
-
baseURL: "https://api.perplexity.ai",
|
|
273
|
-
runCatalog: false,
|
|
274
|
-
billing: "byok",
|
|
275
|
-
authHeaders: ["authorization"],
|
|
276
|
-
hostPattern: PERPLEXITY_HOST,
|
|
277
|
-
transformEndpoint: hostStrip(PERPLEXITY_HOST),
|
|
278
|
-
},
|
|
279
|
-
{
|
|
280
|
-
resolverKey: "cerebras",
|
|
281
|
-
gatewayProviderId: "cerebras",
|
|
282
|
-
wireFormat: "openai",
|
|
283
|
-
baseURL: "https://api.cerebras.ai/v1",
|
|
284
|
-
runCatalog: false,
|
|
285
|
-
billing: "byok",
|
|
286
|
-
authHeaders: ["authorization"],
|
|
287
|
-
hostPattern: CEREBRAS_HOST,
|
|
288
|
-
transformEndpoint: hostStrip(CEREBRAS_HOST),
|
|
289
|
-
},
|
|
290
|
-
{
|
|
291
|
-
resolverKey: "openrouter",
|
|
292
|
-
gatewayProviderId: "openrouter",
|
|
293
|
-
wireFormat: "openai",
|
|
294
|
-
baseURL: "https://openrouter.ai/api/v1",
|
|
295
|
-
runCatalog: false,
|
|
296
|
-
billing: "byok",
|
|
297
|
-
authHeaders: ["authorization"],
|
|
298
|
-
hostPattern: OPENROUTER_HOST,
|
|
299
|
-
transformEndpoint: hostStrip(OPENROUTER_HOST),
|
|
300
|
-
},
|
|
301
|
-
{
|
|
302
|
-
// Fireworks is OpenAI-compatible. Present on ai-gateway-provider (#409) but
|
|
303
|
-
// not the current provider directory — treat as BYOK long-tail.
|
|
304
|
-
resolverKey: "fireworks",
|
|
305
|
-
gatewayProviderId: "fireworks",
|
|
306
|
-
wireFormat: "openai",
|
|
307
|
-
baseURL: "https://api.fireworks.ai/inference/v1",
|
|
308
|
-
runCatalog: false,
|
|
309
|
-
billing: "byok",
|
|
310
|
-
authHeaders: ["authorization"],
|
|
311
|
-
hostPattern: FIREWORKS_HOST,
|
|
312
|
-
transformEndpoint: hostStrip(FIREWORKS_HOST),
|
|
313
|
-
},
|
|
314
|
-
// Providers whose gateway-path URL shape isn't reliably reproducible from the
|
|
315
|
-
// shared openai builder (cohere's /compat surface, baseten's per-deployment
|
|
316
|
-
// hosts, parallel, azure's resource/deployment path) are bring-your-own-provider
|
|
317
|
-
// only — set your own @ai-sdk provider baseURL and route via createGatewayProvider.
|
|
318
|
-
{
|
|
319
|
-
resolverKey: "cohere",
|
|
320
|
-
gatewayProviderId: "cohere",
|
|
321
|
-
runCatalog: false,
|
|
322
|
-
billing: "byok",
|
|
323
|
-
authHeaders: ["authorization"],
|
|
324
|
-
hostPattern: COHERE_HOST,
|
|
325
|
-
transformEndpoint: hostStrip(COHERE_HOST),
|
|
326
|
-
},
|
|
327
|
-
{
|
|
328
|
-
// Baseten serves per-deployment hosts, so there's no single detectable URL
|
|
329
|
-
// shape — reach it with an explicit `provider` via createGatewayProvider.
|
|
330
|
-
resolverKey: "baseten",
|
|
331
|
-
gatewayProviderId: "baseten",
|
|
332
|
-
runCatalog: false,
|
|
333
|
-
billing: "byok",
|
|
334
|
-
authHeaders: ["authorization"],
|
|
335
|
-
},
|
|
336
|
-
{
|
|
337
|
-
resolverKey: "parallel",
|
|
338
|
-
gatewayProviderId: "parallel",
|
|
339
|
-
runCatalog: false,
|
|
340
|
-
billing: "byok",
|
|
341
|
-
authHeaders: ["authorization", "x-api-key"],
|
|
342
|
-
},
|
|
343
|
-
{
|
|
344
|
-
resolverKey: "azure-openai",
|
|
345
|
-
gatewayProviderId: "azure-openai",
|
|
346
|
-
runCatalog: false,
|
|
347
|
-
billing: "byok",
|
|
348
|
-
authHeaders: ["api-key", "authorization"],
|
|
349
|
-
hostPattern: AZURE_HOST,
|
|
350
|
-
transformEndpoint: azureTransform,
|
|
351
|
-
},
|
|
352
|
-
|
|
353
|
-
// ---- Provider-native only: reachable via the bring-your-own-provider wrapper ----
|
|
354
|
-
// (no `wireFormat` ⇒ not auto-wired by the slug delegate)
|
|
355
|
-
{
|
|
356
|
-
resolverKey: "aws-bedrock",
|
|
357
|
-
gatewayProviderId: "aws-bedrock",
|
|
358
|
-
runCatalog: false,
|
|
359
|
-
billing: "byok",
|
|
360
|
-
authHeaders: ["authorization"],
|
|
361
|
-
hostPattern: BEDROCK_HOST,
|
|
362
|
-
transformEndpoint: bedrockTransform,
|
|
363
|
-
},
|
|
364
|
-
{
|
|
365
|
-
resolverKey: "huggingface",
|
|
366
|
-
gatewayProviderId: "huggingface",
|
|
367
|
-
runCatalog: false,
|
|
368
|
-
billing: "byok",
|
|
369
|
-
authHeaders: ["authorization"],
|
|
370
|
-
hostPattern: HUGGINGFACE_HOST,
|
|
371
|
-
transformEndpoint: hostStrip(HUGGINGFACE_HOST),
|
|
372
|
-
},
|
|
373
|
-
{
|
|
374
|
-
resolverKey: "replicate",
|
|
375
|
-
gatewayProviderId: "replicate",
|
|
376
|
-
runCatalog: false,
|
|
377
|
-
billing: "byok",
|
|
378
|
-
authHeaders: ["authorization"],
|
|
379
|
-
hostPattern: REPLICATE_HOST,
|
|
380
|
-
transformEndpoint: hostStrip(REPLICATE_HOST),
|
|
381
|
-
},
|
|
382
|
-
{
|
|
383
|
-
resolverKey: "fal",
|
|
384
|
-
gatewayProviderId: "fal",
|
|
385
|
-
runCatalog: false,
|
|
386
|
-
billing: "byok",
|
|
387
|
-
authHeaders: ["authorization"],
|
|
388
|
-
hostPattern: FAL_HOST,
|
|
389
|
-
transformEndpoint: hostStrip(FAL_HOST),
|
|
390
|
-
},
|
|
391
|
-
{
|
|
392
|
-
resolverKey: "ideogram",
|
|
393
|
-
gatewayProviderId: "ideogram",
|
|
394
|
-
runCatalog: false,
|
|
395
|
-
billing: "byok",
|
|
396
|
-
authHeaders: ["authorization"],
|
|
397
|
-
hostPattern: IDEOGRAM_HOST,
|
|
398
|
-
transformEndpoint: hostStrip(IDEOGRAM_HOST),
|
|
399
|
-
},
|
|
400
|
-
{
|
|
401
|
-
resolverKey: "cartesia",
|
|
402
|
-
gatewayProviderId: "cartesia",
|
|
403
|
-
runCatalog: false,
|
|
404
|
-
billing: "byok",
|
|
405
|
-
authHeaders: ["authorization", "x-api-key"],
|
|
406
|
-
hostPattern: CARTESIA_HOST,
|
|
407
|
-
transformEndpoint: hostStrip(CARTESIA_HOST),
|
|
408
|
-
},
|
|
409
|
-
{
|
|
410
|
-
resolverKey: "deepgram",
|
|
411
|
-
gatewayProviderId: "deepgram",
|
|
412
|
-
runCatalog: false,
|
|
413
|
-
billing: "byok",
|
|
414
|
-
authHeaders: ["authorization", "token"],
|
|
415
|
-
hostPattern: DEEPGRAM_HOST,
|
|
416
|
-
transformEndpoint: hostStrip(DEEPGRAM_HOST),
|
|
417
|
-
},
|
|
418
|
-
{
|
|
419
|
-
resolverKey: "elevenlabs",
|
|
420
|
-
gatewayProviderId: "elevenlabs",
|
|
421
|
-
runCatalog: false,
|
|
422
|
-
billing: "byok",
|
|
423
|
-
authHeaders: ["xi-api-key", "authorization"],
|
|
424
|
-
hostPattern: ELEVENLABS_HOST,
|
|
425
|
-
transformEndpoint: hostStrip(ELEVENLABS_HOST),
|
|
426
|
-
},
|
|
427
|
-
];
|
|
428
|
-
|
|
429
|
-
/** Aliases that map a friendly slug prefix to a canonical `resolverKey`. */
|
|
430
|
-
const RESOLVER_ALIASES: Record<string, string> = {
|
|
431
|
-
// xAI's run-catalog author is `xai`, but `grok` is the common name.
|
|
432
|
-
grok: "xai",
|
|
433
|
-
"google-ai-studio": "google",
|
|
434
|
-
"google-vertex-ai": "google-vertex",
|
|
435
|
-
bedrock: "aws-bedrock",
|
|
436
|
-
azure: "azure-openai",
|
|
437
|
-
};
|
|
438
|
-
|
|
439
|
-
const BY_RESOLVER_KEY = new Map<string, GatewayProviderInfo>(
|
|
440
|
-
GATEWAY_PROVIDERS.map((p) => [p.resolverKey, p]),
|
|
441
|
-
);
|
|
442
|
-
|
|
443
|
-
/** Look up a provider by the slug prefix the user typed (honoring aliases). */
|
|
444
|
-
export function findProviderBySlug(resolverKey: string): GatewayProviderInfo | undefined {
|
|
445
|
-
const canonical = RESOLVER_ALIASES[resolverKey] ?? resolverKey;
|
|
446
|
-
return BY_RESOLVER_KEY.get(canonical);
|
|
447
|
-
}
|
|
448
|
-
|
|
449
|
-
/** Detect the gateway provider from a wrapped provider's request URL (BYOG). */
|
|
450
|
-
export function detectProviderByUrl(url: string): GatewayProviderInfo | undefined {
|
|
451
|
-
return GATEWAY_PROVIDERS.find((p) => p.hostPattern?.test(url));
|
|
452
|
-
}
|
|
453
|
-
|
|
454
|
-
/** All slug keys with a built-in parser (auto-wireable by the slug delegate). */
|
|
455
|
-
export function wireableProviders(): GatewayProviderInfo[] {
|
|
456
|
-
return GATEWAY_PROVIDERS.filter((p) => p.wireFormat !== undefined);
|
|
457
|
-
}
|
|
7
|
+
export {
|
|
8
|
+
type Billing,
|
|
9
|
+
detectProviderByUrl,
|
|
10
|
+
findProviderBySlug,
|
|
11
|
+
GATEWAY_PROVIDERS,
|
|
12
|
+
type GatewayProviderInfo,
|
|
13
|
+
type WireFormat,
|
|
14
|
+
wireableProviders,
|
|
15
|
+
} from "@cloudflare/gateway-core";
|
package/src/index.ts
CHANGED
|
@@ -305,12 +305,25 @@ export function createWorkersAI(options: WorkersAISettings): WorkersAI {
|
|
|
305
305
|
isBinding,
|
|
306
306
|
});
|
|
307
307
|
|
|
308
|
-
const toGatewayOptions = (
|
|
309
|
-
|
|
310
|
-
|
|
311
|
-
|
|
308
|
+
const toGatewayOptions = (
|
|
309
|
+
gateway: GatewayOptions | string | undefined,
|
|
310
|
+
): GatewayOptions | undefined => (typeof gateway === "string" ? { id: gateway } : gateway);
|
|
311
|
+
|
|
312
|
+
// The bare unified-billing run path (`env.AI.run(model, …)`) shared by two
|
|
313
|
+
// callers that both hit it WITHOUT the gateway delegate:
|
|
314
|
+
// - `dynamic/<route>` AI Gateway dynamic routes, and
|
|
315
|
+
// - `"<vendor>/<model>"` catalog slugs when no `providers` are configured
|
|
316
|
+
// (the #596 passthrough).
|
|
317
|
+
// Neither has a transport choice, resume engine, or BYOK forwarding — those
|
|
318
|
+
// only exist once `providers` are set and the delegate takes over — so this
|
|
319
|
+
// path defaults the gateway (third-party run models need one), folds
|
|
320
|
+
// cache/metadata into the gateway options, and rejects delegate-only options
|
|
321
|
+
// loudly instead of letting `WorkersAIChatLanguageModel` spread them as junk
|
|
322
|
+
// into `binding.run`'s options arg.
|
|
323
|
+
const createRunPathModel = (
|
|
312
324
|
modelId: TextGenerationModels,
|
|
313
|
-
settings:
|
|
325
|
+
settings: WorkersAIChatSettings & DelegateCallOptions = {},
|
|
326
|
+
kind: "dynamic" | "catalog",
|
|
314
327
|
) => {
|
|
315
328
|
if (
|
|
316
329
|
settings.fallback ||
|
|
@@ -320,16 +333,22 @@ export function createWorkersAI(options: WorkersAISettings): WorkersAI {
|
|
|
320
333
|
settings.onResumeExpired ||
|
|
321
334
|
settings.byok
|
|
322
335
|
) {
|
|
336
|
+
const subject =
|
|
337
|
+
kind === "dynamic"
|
|
338
|
+
? `"${modelId}" is an AI Gateway dynamic route`
|
|
339
|
+
: `"${modelId}" routes through the bare unified-billing run path because no \`providers\` are configured`;
|
|
323
340
|
throw new Error(
|
|
324
|
-
|
|
325
|
-
"
|
|
326
|
-
"
|
|
327
|
-
"
|
|
341
|
+
`${subject}. It uses AI.run with OpenAI-compatible chat-completions wire format; ` +
|
|
342
|
+
"fallback, gateway transport, resume, BYOK, and resume callbacks are gateway-delegate " +
|
|
343
|
+
"features — configure provider plugins to use them " +
|
|
344
|
+
"(createWorkersAI({ binding: env.AI, providers: [openai] })), and for dynamic routes " +
|
|
345
|
+
"set caching/fallback on the route or gateway instead of per call.",
|
|
328
346
|
);
|
|
329
347
|
}
|
|
330
348
|
|
|
331
349
|
const gateway = {
|
|
332
|
-
...(toGatewayOptions(settings.gateway) ??
|
|
350
|
+
...(toGatewayOptions(settings.gateway) ??
|
|
351
|
+
options.gateway ?? { id: DEFAULT_GATEWAY_ID }),
|
|
333
352
|
};
|
|
334
353
|
if (settings.metadata) {
|
|
335
354
|
gateway.metadata = {
|
|
@@ -365,7 +384,12 @@ export function createWorkersAI(options: WorkersAISettings): WorkersAI {
|
|
|
365
384
|
|
|
366
385
|
const plugin = options.providers?.find((p) => p.wireFormat === DYNAMIC_ROUTE_WIRE_FORMAT);
|
|
367
386
|
if (!plugin) {
|
|
368
|
-
|
|
387
|
+
// Dynamic routes with providers configured but no OpenAI plugin can't be
|
|
388
|
+
// parsed (they always return openai-wire), so guide the user. The
|
|
389
|
+
// catalog passthrough only reaches here with no providers at all, where
|
|
390
|
+
// the built-in `WorkersAIChatLanguageModel` parser is the intended
|
|
391
|
+
// fallback (#596) — so let it through.
|
|
392
|
+
if (kind === "dynamic" && options.providers?.length) {
|
|
369
393
|
throw new Error(
|
|
370
394
|
`"${modelId}" is an AI Gateway dynamic route. Dynamic routes return OpenAI-compatible ` +
|
|
371
395
|
"chat-completions wire format on the AI.run path, so configure the OpenAI " +
|
|
@@ -375,7 +399,10 @@ export function createWorkersAI(options: WorkersAISettings): WorkersAI {
|
|
|
375
399
|
}
|
|
376
400
|
return createChatModel(modelId, chatSettings);
|
|
377
401
|
}
|
|
378
|
-
const fetchImpl = (async (
|
|
402
|
+
const fetchImpl = (async (
|
|
403
|
+
_input: RequestInfo | URL,
|
|
404
|
+
init?: RequestInit,
|
|
405
|
+
): Promise<Response> => {
|
|
379
406
|
const body = JSON.parse(String(init?.body ?? "{}")) as Record<string, unknown>;
|
|
380
407
|
delete body.model;
|
|
381
408
|
|
|
@@ -385,13 +412,15 @@ export function createWorkersAI(options: WorkersAISettings): WorkersAI {
|
|
|
385
412
|
...(settings.extraHeaders ? { extraHeaders: settings.extraHeaders } : {}),
|
|
386
413
|
...(init?.signal ? { signal: init.signal } : {}),
|
|
387
414
|
};
|
|
388
|
-
const response = await (
|
|
389
|
-
|
|
390
|
-
|
|
391
|
-
|
|
392
|
-
|
|
393
|
-
|
|
394
|
-
|
|
415
|
+
const response = await (
|
|
416
|
+
binding as unknown as {
|
|
417
|
+
run(
|
|
418
|
+
model: string,
|
|
419
|
+
inputs: Record<string, unknown>,
|
|
420
|
+
options: Record<string, unknown>,
|
|
421
|
+
): Promise<Response>;
|
|
422
|
+
}
|
|
423
|
+
).run(modelId, body, runOptions);
|
|
395
424
|
settings.onDispatch?.({
|
|
396
425
|
transport: "run",
|
|
397
426
|
resumeEnabled: false,
|
|
@@ -458,12 +487,27 @@ export function createWorkersAI(options: WorkersAISettings): WorkersAI {
|
|
|
458
487
|
settings?: WorkersAIChatSettings | DelegateCallOptions,
|
|
459
488
|
): WorkersAIChatLanguageModel => {
|
|
460
489
|
if (typeof modelId === "string" && modelId.startsWith("dynamic/")) {
|
|
461
|
-
return
|
|
490
|
+
return createRunPathModel(
|
|
462
491
|
modelId,
|
|
463
492
|
settings as (WorkersAIChatSettings & DelegateCallOptions) | undefined,
|
|
493
|
+
"dynamic",
|
|
464
494
|
);
|
|
465
495
|
}
|
|
466
496
|
if (isGatewaySlug(modelId)) {
|
|
497
|
+
// Without provider plugins there's no gateway/BYOK routing to do, and a
|
|
498
|
+
// bare `"<vendor>/<model>"` id is a valid Workers AI unified-billing run
|
|
499
|
+
// model — `env.AI.run("deepseek/deepseek-v4-pro")`. Route it through the
|
|
500
|
+
// same bare run path as dynamic routes (defaults the gateway, folds
|
|
501
|
+
// cache/metadata, rejects delegate-only options) instead of erroring.
|
|
502
|
+
// Catalog routing (resume, fallback, caching, BYOK) only kicks in once
|
|
503
|
+
// `providers` are set. (#596)
|
|
504
|
+
if (!options.providers?.length) {
|
|
505
|
+
return createRunPathModel(
|
|
506
|
+
modelId,
|
|
507
|
+
settings as (WorkersAIChatSettings & DelegateCallOptions) | undefined,
|
|
508
|
+
"catalog",
|
|
509
|
+
);
|
|
510
|
+
}
|
|
467
511
|
// The delegate returns a `LanguageModelV3` built by the configured plugin.
|
|
468
512
|
// It's structurally compatible with the AI SDK consumers this provider is
|
|
469
513
|
// used with; the cast keeps the public return type unchanged.
|