@uploadfilm/sdk 0.0.0-stage → 0.2.0

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.
Files changed (267) hide show
  1. package/LICENSE +4 -0
  2. package/README.md +529 -2
  3. package/dist/ai/embed.d.ts +58 -0
  4. package/dist/ai/fake.d.ts +41 -0
  5. package/dist/ai/http.d.ts +17 -0
  6. package/dist/ai/index.d.ts +27 -0
  7. package/dist/ai/languages.d.ts +2 -0
  8. package/dist/ai/ollama.d.ts +23 -0
  9. package/dist/ai/openaiCompatible.d.ts +54 -0
  10. package/dist/ai/passages.d.ts +28 -0
  11. package/dist/ai/transcribe.d.ts +84 -0
  12. package/dist/ai/types.d.ts +141 -0
  13. package/dist/ai/usage.d.ts +66 -0
  14. package/dist/ai/whisperJson.d.ts +15 -0
  15. package/dist/ai/workersAi.d.ts +56 -0
  16. package/dist/ai.js +1 -0
  17. package/dist/ai.mjs +1 -0
  18. package/dist/cards/files.d.ts +91 -0
  19. package/dist/cards/index.d.ts +15 -0
  20. package/dist/cards/layouts.d.ts +121 -0
  21. package/dist/cards/probe.d.ts +69 -0
  22. package/dist/cards/proxies.d.ts +70 -0
  23. package/dist/cards/select.d.ts +68 -0
  24. package/dist/cards/volumes.d.ts +170 -0
  25. package/dist/cards/xml.d.ts +80 -0
  26. package/dist/cards.js +1 -0
  27. package/dist/cards.mjs +1 -0
  28. package/dist/copy/checksums.d.ts +66 -0
  29. package/dist/copy/engine.d.ts +24 -0
  30. package/dist/copy/errors.d.ts +109 -0
  31. package/dist/copy/index.d.ts +15 -0
  32. package/dist/copy/steps.d.ts +28 -0
  33. package/dist/copy/sweep.d.ts +24 -0
  34. package/dist/copy/types.d.ts +363 -0
  35. package/dist/embed/player.js +2 -0
  36. package/dist/host/adapters.d.ts +107 -0
  37. package/dist/host/detect.d.ts +68 -0
  38. package/dist/host/encoders.d.ts +34 -0
  39. package/dist/host/environment.d.ts +53 -0
  40. package/dist/host/index.d.ts +20 -0
  41. package/dist/host/linux.d.ts +94 -0
  42. package/dist/host/mac.d.ts +47 -0
  43. package/dist/host/network.d.ts +54 -0
  44. package/dist/host/node.d.ts +13 -0
  45. package/dist/host/policy.d.ts +101 -0
  46. package/dist/host/tuner.d.ts +48 -0
  47. package/dist/host/types.d.ts +118 -0
  48. package/dist/host-node.js +1 -0
  49. package/dist/host-node.mjs +1 -0
  50. package/dist/host.js +1 -0
  51. package/dist/host.mjs +1 -0
  52. package/dist/images/index.d.ts +11 -0
  53. package/dist/images/settings.d.ts +43 -0
  54. package/dist/images/types.d.ts +83 -0
  55. package/dist/images/url.d.ts +65 -0
  56. package/dist/index.d.ts +16 -0
  57. package/dist/index.js +1 -0
  58. package/dist/index.mjs +1 -0
  59. package/dist/player/adapters/hlsjs.d.ts +28 -0
  60. package/dist/player/bandwidthMemory.d.ts +28 -0
  61. package/dist/player/captions/index.d.ts +55 -0
  62. package/dist/player/embed/code.d.ts +195 -0
  63. package/dist/player/embed/constants.d.ts +4 -0
  64. package/dist/player/embed/controller.d.ts +62 -0
  65. package/dist/player/embed/element.d.ts +32 -0
  66. package/dist/player/embed/frame.d.ts +65 -0
  67. package/dist/player/embed/global.d.ts +14 -0
  68. package/dist/player/embed/index.d.ts +15 -0
  69. package/dist/player/embed/loader.d.ts +1 -0
  70. package/dist/player/embed/oembed.d.ts +74 -0
  71. package/dist/player/embed/protocol.d.ts +64 -0
  72. package/dist/player/engine.d.ts +215 -0
  73. package/dist/player/index.d.ts +18 -0
  74. package/dist/player/start.d.ts +62 -0
  75. package/dist/player/store.d.ts +10 -0
  76. package/dist/player/storyboard/index.d.ts +161 -0
  77. package/dist/player/stream.d.ts +206 -0
  78. package/dist/player/tracks.d.ts +41 -0
  79. package/dist/player.js +1 -0
  80. package/dist/player.mjs +1 -0
  81. package/dist/proxies/args.d.ts +19 -0
  82. package/dist/proxies/budget.d.ts +57 -0
  83. package/dist/proxies/camera.d.ts +104 -0
  84. package/dist/proxies/formats.d.ts +36 -0
  85. package/dist/proxies/index.d.ts +18 -0
  86. package/dist/proxies/jobs.d.ts +98 -0
  87. package/dist/proxies/node.d.ts +32 -0
  88. package/dist/proxies/plan.d.ts +89 -0
  89. package/dist/proxies/probe.d.ts +39 -0
  90. package/dist/proxies/render.d.ts +104 -0
  91. package/dist/proxies/safety.d.ts +45 -0
  92. package/dist/proxies/types.d.ts +83 -0
  93. package/dist/proxies/upload.d.ts +14 -0
  94. package/dist/react/BigPlay.d.ts +26 -0
  95. package/dist/react/Player.d.ts +58 -0
  96. package/dist/react/embed/index.d.ts +45 -0
  97. package/dist/react/examples/CustomControls.d.ts +8 -0
  98. package/dist/react/images/Image.d.ts +16 -0
  99. package/dist/react/images/index.d.ts +4 -0
  100. package/dist/react/images/useImage.d.ts +47 -0
  101. package/dist/react/index.d.ts +17 -0
  102. package/dist/react/storyboard/index.d.ts +58 -0
  103. package/dist/react/theme.d.ts +37 -0
  104. package/dist/react/useAdapterErrors.d.ts +8 -0
  105. package/dist/react/usePlaybackStart.d.ts +18 -0
  106. package/dist/react/usePlayer.d.ts +28 -0
  107. package/dist/react/useReducedMotion.d.ts +2 -0
  108. package/dist/react/useTimedTrack.d.ts +22 -0
  109. package/dist/react.js +1 -0
  110. package/dist/react.mjs +1 -0
  111. package/dist/server/auth/apiKeys.d.ts +30 -0
  112. package/dist/server/auth/cache.d.ts +21 -0
  113. package/dist/server/auth/deviceTokens.d.ts +38 -0
  114. package/dist/server/auth/envelope.d.ts +15 -0
  115. package/dist/server/auth/grants.d.ts +76 -0
  116. package/dist/server/auth/index.d.ts +19 -0
  117. package/dist/server/auth/jwt.d.ts +65 -0
  118. package/dist/server/auth/keyring.d.ts +31 -0
  119. package/dist/server/auth/kv.d.ts +12 -0
  120. package/dist/server/auth/repository.d.ts +45 -0
  121. package/dist/server/auth/revocations.d.ts +21 -0
  122. package/dist/server/auth/scope.d.ts +4 -0
  123. package/dist/server/auth/serviceAuth.d.ts +24 -0
  124. package/dist/server/auth/signedUrls.d.ts +28 -0
  125. package/dist/server/embed/index.d.ts +36 -0
  126. package/dist/server/embed/lifecycle.d.ts +53 -0
  127. package/dist/server/embed/playerScript.d.ts +4 -0
  128. package/dist/server/images/index.d.ts +15 -0
  129. package/dist/server/images/job.d.ts +80 -0
  130. package/dist/server/images/layout.d.ts +24 -0
  131. package/dist/server/images/params.d.ts +90 -0
  132. package/dist/server/images/route.d.ts +101 -0
  133. package/dist/server/images/sign.d.ts +55 -0
  134. package/dist/server/index.d.ts +42 -0
  135. package/dist/server/media/client.d.ts +99 -0
  136. package/dist/server/media/frameCheck.d.ts +62 -0
  137. package/dist/server/media/index.d.ts +10 -0
  138. package/dist/server/media/protocol.d.ts +556 -0
  139. package/dist/server/media/storyboard.d.ts +144 -0
  140. package/dist/server/media/transcript.d.ts +143 -0
  141. package/dist/server/pipeline/index.d.ts +12 -0
  142. package/dist/server/pipeline/match.d.ts +10 -0
  143. package/dist/server/pipeline/pipeline.d.ts +182 -0
  144. package/dist/server/pipeline/queues.d.ts +108 -0
  145. package/dist/server/pipeline/sources.d.ts +54 -0
  146. package/dist/server/r2/client.d.ts +164 -0
  147. package/dist/server/r2/crc32.d.ts +24 -0
  148. package/dist/server/r2/encoding.d.ts +35 -0
  149. package/dist/server/r2/errors.d.ts +9 -0
  150. package/dist/server/r2/integrity.d.ts +25 -0
  151. package/dist/server/r2/parts.d.ts +46 -0
  152. package/dist/server/r2/sign.d.ts +47 -0
  153. package/dist/server/redis/index.d.ts +39 -0
  154. package/dist/server/state/memory.d.ts +18 -0
  155. package/dist/server/state/r2.d.ts +73 -0
  156. package/dist/server/state/types.d.ts +57 -0
  157. package/dist/server/storage/chunked.d.ts +25 -0
  158. package/dist/server/storage/index.d.ts +12 -0
  159. package/dist/server/storage/r2.d.ts +96 -0
  160. package/dist/server/storage/s3.d.ts +144 -0
  161. package/dist/server/testing/fake-r2.d.ts +410 -0
  162. package/dist/server/testing/index.d.ts +10 -0
  163. package/dist/server/testing/kv-contract.d.ts +14 -0
  164. package/dist/server/transfer/handler.d.ts +217 -0
  165. package/dist/server/transfer/keys.d.ts +45 -0
  166. package/dist/server/transfer/protocol.d.ts +190 -0
  167. package/dist/server/zip/client.d.ts +32 -0
  168. package/dist/server/zip/generations.d.ts +83 -0
  169. package/dist/server/zip/index.d.ts +12 -0
  170. package/dist/server/zip/planner.d.ts +94 -0
  171. package/dist/server/zip/sign.d.ts +22 -0
  172. package/dist/server-media.js +1 -0
  173. package/dist/server-media.mjs +1 -0
  174. package/dist/server-redis.js +1 -0
  175. package/dist/server-redis.mjs +1 -0
  176. package/dist/server-testing.js +1 -0
  177. package/dist/server-testing.mjs +1 -0
  178. package/dist/server.js +1 -0
  179. package/dist/server.mjs +1 -0
  180. package/dist/storage/browserFile.d.ts +19 -0
  181. package/dist/storage/index.d.ts +12 -0
  182. package/dist/storage/local.d.ts +9 -0
  183. package/dist/storage/presigned.d.ts +22 -0
  184. package/dist/sync/ffmpeg.d.ts +44 -0
  185. package/dist/sync/index.d.ts +17 -0
  186. package/dist/sync/match.d.ts +28 -0
  187. package/dist/sync/probe.d.ts +30 -0
  188. package/dist/sync/timecode.d.ts +50 -0
  189. package/dist/sync/transcript.d.ts +35 -0
  190. package/dist/sync/types.d.ts +172 -0
  191. package/dist/sync/waveform.d.ts +89 -0
  192. package/dist/sync.js +1 -0
  193. package/dist/sync.mjs +1 -0
  194. package/dist/transfer/claims.d.ts +41 -0
  195. package/dist/transfer/client.d.ts +267 -0
  196. package/dist/transfer/download.d.ts +192 -0
  197. package/dist/transfer/engine.d.ts +402 -0
  198. package/dist/transfer/errors.d.ts +50 -0
  199. package/dist/transfer/growing/bytes.d.ts +17 -0
  200. package/dist/transfer/growing/index.d.ts +12 -0
  201. package/dist/transfer/growing/parts.d.ts +157 -0
  202. package/dist/transfer/growing/retry.d.ts +38 -0
  203. package/dist/transfer/growing/store.d.ts +66 -0
  204. package/dist/transfer/growing/upload.d.ts +126 -0
  205. package/dist/transfer/index.d.ts +40 -0
  206. package/dist/transfer/integrity.d.ts +93 -0
  207. package/dist/transfer/limits.d.ts +34 -0
  208. package/dist/transfer/node.d.ts +63 -0
  209. package/dist/transfer/pool.d.ts +83 -0
  210. package/dist/transfer/protocol.d.ts +172 -0
  211. package/dist/transfer/signer.d.ts +29 -0
  212. package/dist/transfer/storage.d.ts +135 -0
  213. package/dist/transfer/timing.d.ts +132 -0
  214. package/dist/transfer/transports/electronNet.d.ts +62 -0
  215. package/dist/transfer/transports/fetch.d.ts +15 -0
  216. package/dist/transfer/transports/retryAfter.d.ts +12 -0
  217. package/dist/transfer/transports/undici.d.ts +72 -0
  218. package/dist/transfer/tuner.d.ts +56 -0
  219. package/dist/workers/cli.mjs +115 -0
  220. package/dist/workers/media/config.mjs +53 -0
  221. package/dist/workers/media/container/Dockerfile +19 -0
  222. package/dist/workers/media/container/FFMPEG.md +129 -0
  223. package/dist/workers/media/container/runner.mjs +2690 -0
  224. package/dist/workers/media/index.mjs +1011 -0
  225. package/dist/workers/media/init/config.ts +52 -0
  226. package/dist/workers/media/init/worker.ts +8 -0
  227. package/dist/workers/media/init/wrangler.toml +44 -0
  228. package/dist/workers/media/runner.mjs +2656 -0
  229. package/dist/workers/media/types/build/entry/config.d.mts +6 -0
  230. package/dist/workers/media/types/build/entry/index.d.mts +43 -0
  231. package/dist/workers/media/types/build/entry/program.d.mts +13 -0
  232. package/dist/workers/media/types/build/entry/runner.d.mts +12 -0
  233. package/dist/workers/media/types/build/entry/settings.d.mts +40 -0
  234. package/dist/workers/media/types/src/config.d.mts +135 -0
  235. package/dist/workers/media/types/src/jobs.d.mts +73 -0
  236. package/dist/workers/media/types/src/runner/container.d.mts +32 -0
  237. package/dist/workers/media/types/src/runner/ffmpeg.d.mts +227 -0
  238. package/dist/workers/media/types/src/runner/job.d.mts +107 -0
  239. package/dist/workers/media/types/src/runner/node.d.mts +71 -0
  240. package/dist/workers/media/types/src/runner/plugins.d.mts +45 -0
  241. package/dist/workers/media/types/src/serve.d.mts +107 -0
  242. package/dist/workers/media/types/src/steps/audio.d.mts +33 -0
  243. package/dist/workers/media/types/src/steps/edit.d.mts +62 -0
  244. package/dist/workers/media/types/src/steps/frameCheck.d.mts +44 -0
  245. package/dist/workers/media/types/src/steps/gate.d.mts +18 -0
  246. package/dist/workers/media/types/src/steps/paths.d.mts +1 -0
  247. package/dist/workers/media/types/src/steps/prep.d.mts +60 -0
  248. package/dist/workers/media/types/src/steps/render.d.mts +40 -0
  249. package/dist/workers/media/types/src/steps/runner.d.mts +30 -0
  250. package/dist/workers/media/types/src/steps/transcribe.d.mts +89 -0
  251. package/dist/workers/media/types/src/storyboard/index.d.mts +61 -0
  252. package/dist/workers/media/types/src/transcoder.d.mts +30 -0
  253. package/dist/workers/media/types/src/worker.d.mts +30 -0
  254. package/dist/workers/zip/index.mjs +589 -0
  255. package/dist/workers/zip/init/config.ts +22 -0
  256. package/dist/workers/zip/init/worker.ts +5 -0
  257. package/dist/workers/zip/init/wrangler.toml +25 -0
  258. package/dist/workers/zip/types/build/entry/index.d.mts +5 -0
  259. package/dist/workers/zip/types/src/config.d.mts +35 -0
  260. package/dist/workers/zip/types/src/serve.d.mts +86 -0
  261. package/dist/workers/zip/types/src/worker.d.mts +12 -0
  262. package/dist/zip/index.d.ts +13 -0
  263. package/dist/zip/layout.d.ts +129 -0
  264. package/dist/zip/names.d.ts +34 -0
  265. package/dist/zip/plan.d.ts +115 -0
  266. package/dist/zip/sw.d.ts +7 -0
  267. package/package.json +313 -4
@@ -0,0 +1,27 @@
1
+ /**
2
+ * Pluggable AI: Transcriber and Embedder interfaces, their adapters, and the helpers that take any of them.
3
+ *
4
+ * adapters workersAiTranscriber / workersAiEmbedder (binding or workersAiRest), openAiCompatibleTranscriber /
5
+ * openAiCompatibleEmbedder (OpenAI, Groq, OpenRouter, any compatible server), ollamaEmbedder, fakes.
6
+ * Local subprocess models (whisper.cpp, MLX Whisper, sentence-transformers): @uploadfilm/ai-plugins.
7
+ * transcripts transcribeAudio (media-steps' VAD + progress + limits, for any transcriber), settleTranscript
8
+ * embeddings transcriptPassages, embedTranscript / embedPassages / embedFrames → the product's onEmbeddings hook
9
+ * usage every call's latency and cost (AiUsage), usageLog, recordedTranscriber / recordedEmbedder
10
+ *
11
+ * No credentials live here: keys, bindings and fetch are injected. Nothing is stored: vectors go to the hook.
12
+ */
13
+ export declare const SDK_ENTRY: "ai";
14
+ export * from "./types";
15
+ export * from "./usage";
16
+ export { languageCode } from "./languages";
17
+ export { readWhisperJson, confidenceFromLogprob, type ReadWhisper } from "./whisperJson";
18
+ export { credential, MAX_ANSWER_BYTES, type Credential, type FetchLike } from "./http";
19
+ export { WORKERS_AI_WHISPER, WORKERS_AI_TIMEOUT_MS, whisperInput, workersAiEmbedder, workersAiRest, workersAiTranscriber, readVectors, type AiLike, type WorkersAiEmbedderOptions, type WorkersAiTranscriberOptions, } from "./workersAi";
20
+ export { openAiCompatibleEmbedder, openAiCompatibleTranscriber, type OpenAiCompatibleOptions, type OpenAiEmbedderOptions, type OpenAiTranscriberOptions, } from "./openaiCompatible";
21
+ export { ollamaEmbedder, type OllamaEmbedderOptions } from "./ollama";
22
+ export { fakeEmbedder, fakeTranscriber, fakeVector, type FakeAnswer, type FakeEmbedderOptions, type FakeTranscriberOptions, } from "./fake";
23
+ export { MAX_CHUNK_BYTES, MAX_TRANSCRIPT_S, keepAnswer, settleTranscript, transcribeAudio, type TranscribeAudioOptions, type TranscriptOutcome, type TranscriptProgress, type TranscriptRun, } from "./transcribe";
24
+ export { transcriptPassages, type Passage, type PassageOptions } from "./passages";
25
+ export { embedFrames, embedPassages, embedTranscript, type EmbedOutcome, type EmbedRunOptions, type EmbeddingBatch, } from "./embed";
26
+ export type { AudioChunk, AudioIndex, Silence, Transcript, TranscriptSegment, TranscriptWord, VadSettings, WhisperResult, } from "../server/media/transcript";
27
+ export { toWebVtt } from "../server/media/transcript";
@@ -0,0 +1,2 @@
1
+ /** "English" / "en" / "en-US" → "en"; anything else unknown → undefined (never a guess). */
2
+ export declare function languageCode(x: unknown): string | undefined;
@@ -0,0 +1,23 @@
1
+ /**
2
+ * Ollama's own embeddings route (a local model server on the Mac or a self-hosted runner): POST <base>/api/embed
3
+ * {model, input} → {embeddings}. (Ollama's /v1 routes also work with openAiCompatibleEmbedder.)
4
+ */
5
+ import { type FetchLike } from "./http";
6
+ import { type Embedder } from "./types";
7
+ export interface OllamaEmbedderOptions {
8
+ /** e.g. "http://127.0.0.1:11434". */
9
+ baseUrl: string;
10
+ /** e.g. "nomic-embed-text", "bge-m3": the product's choice, always named. */
11
+ model: string;
12
+ fetch?: FetchLike;
13
+ timeoutMs?: number;
14
+ maxBatch?: number;
15
+ /** Ollama's keep_alive (how long the model stays loaded), passed through. */
16
+ keepAlive?: string;
17
+ prefixes?: {
18
+ query?: string;
19
+ document?: string;
20
+ };
21
+ now?: () => number;
22
+ }
23
+ export declare function ollamaEmbedder(o: OllamaEmbedderOptions): Embedder;
@@ -0,0 +1,54 @@
1
+ /**
2
+ * Any OpenAI-compatible endpoint: OpenAI, Groq, OpenRouter, Together, Fireworks, a self-hosted faster-whisper-server,
3
+ * vLLM, LocalAI, Ollama's /v1 … The product passes the base URL and its key (or a function giving it); nothing here
4
+ * names a provider's URL or reads the environment.
5
+ *
6
+ * transcriber: POST <base>/audio/transcriptions (multipart), response_format=verbose_json, word + segment timestamps
7
+ * embedder: POST <base>/embeddings {model, input, encoding_format: "float"}
8
+ */
9
+ import { type Credential, type FetchLike } from "./http";
10
+ import { type AiPrice, type Embedder, type Transcriber } from "./types";
11
+ export interface OpenAiCompatibleOptions {
12
+ /** e.g. "https://api.openai.com/v1", "https://api.groq.com/openai/v1", "https://openrouter.ai/api/v1". */
13
+ baseUrl: string;
14
+ model: string;
15
+ /** Bearer key, or a function giving it (asked per call). Absent: no authorization header (a local server). */
16
+ apiKey?: Credential;
17
+ /** More headers (OpenRouter's HTTP-Referer / X-Title, an org id). Never put a key in a URL. */
18
+ headers?: Record<string, string>;
19
+ fetch?: FetchLike;
20
+ price?: AiPrice;
21
+ timeoutMs?: number;
22
+ /** Shown in usage and errors instead of "openai-compatible" (e.g. "groq"). */
23
+ name?: string;
24
+ now?: () => number;
25
+ }
26
+ export interface OpenAiTranscriberOptions extends OpenAiCompatibleOptions {
27
+ /**
28
+ * "verbose_json" (default: segments, words, language) or "json" (models that only give text, like gpt-4o-transcribe:
29
+ * one segment spanning the audio, no word timings).
30
+ */
31
+ responseFormat?: "verbose_json" | "json";
32
+ /** Ask for word timings (default true; verbose_json only). */
33
+ words?: boolean;
34
+ /** Biggest upload (OpenAI and Groq take 25 MB). */
35
+ maxBytes?: number;
36
+ }
37
+ export declare function openAiCompatibleTranscriber(o: OpenAiTranscriberOptions): Transcriber;
38
+ export interface OpenAiEmbedderOptions extends OpenAiCompatibleOptions {
39
+ /** Output dimensions, for models that can shorten their vectors (text-embedding-3-*). */
40
+ dimensions?: number;
41
+ /** Inputs per call (default 64). */
42
+ maxBatch?: number;
43
+ /**
44
+ * Images: endpoints that embed them through this same route (Jina, vLLM multimodal, Cohere-compatible proxies) take
45
+ * each one as {image: <data URL>}. Off by default: OpenAI's own endpoint embeds text only.
46
+ */
47
+ images?: boolean;
48
+ /** Prefixes some models want ("query: " / "passage: " for e5; "search_query: " / "search_document: " for nomic). */
49
+ prefixes?: {
50
+ query?: string;
51
+ document?: string;
52
+ };
53
+ }
54
+ export declare function openAiCompatibleEmbedder(o: OpenAiEmbedderOptions): Embedder;
@@ -0,0 +1,28 @@
1
+ /**
2
+ * Transcript → time-coded passages to embed (search "where does she say X" → a time to seek to). Passages follow
3
+ * segment boundaries; a segment too long for one passage is split at its words (their own times) or, without word
4
+ * times, at sentence/word breaks with times shared out by length. Pure.
5
+ */
6
+ import type { Transcript } from "../server/media/transcript";
7
+ export interface Passage {
8
+ /** 0-based, in time order: with the asset id, a stable key for the product's own vector store. */
9
+ index: number;
10
+ /** Seconds into the media. */
11
+ start: number;
12
+ end: number;
13
+ text: string;
14
+ }
15
+ export interface PassageOptions {
16
+ /**
17
+ * A passage spans at most this many seconds (the product's choice: there is no default). A single word longer than
18
+ * this (a model's long last word) still makes one passage.
19
+ */
20
+ maxSeconds: number;
21
+ /** …and at most this many characters. */
22
+ maxChars: number;
23
+ /** Seconds of the previous passage repeated at the start of the next (context across a cut). Default 0. */
24
+ overlapS?: number;
25
+ /** A silence (gap between segments) this long ends a passage. Default: never. */
26
+ breakOnGapS?: number;
27
+ }
28
+ export declare function transcriptPassages(t: Pick<Transcript, "segments">, o: PassageOptions): Passage[];
@@ -0,0 +1,84 @@
1
+ /**
2
+ * A whole recording through ANY transcriber, with media-steps' guarantees:
3
+ *
4
+ * audio index (chunks + silences measured on the job's one decode) → each chunk that isn't silence end to end →
5
+ * transcriber.transcribe(chunk) → segments offset by the chunk's start, segments and words inside silence dropped
6
+ * (buildTranscript, the same VAD as the media worker) → a Transcript (transcript.json's shape) + per-call usage
7
+ *
8
+ * - Silence never becomes text: silent chunks are never sent; what a model says over silence is dropped.
9
+ * - Long recordings span several runs: each chunk's answer goes to `progress` as soon as it arrives; past `deadline` no
10
+ * new call starts and the run says how many chunks are left. The next run picks up where this one stopped.
11
+ * - Limits checked before any call: total speech length, each chunk's size (the transcriber's own maximum too).
12
+ * - `settleTranscript` turns any failure into {state: "error", reason}: a transcript never fails a video job.
13
+ *
14
+ * Where the chunks come from is the caller's (`load`): a bucket (the media worker), a local folder (the Node runner),
15
+ * a product's own storage (on demand).
16
+ */
17
+ import { type AudioChunk, type AudioIndex, type Transcript, type VadSettings, type WhisperResult } from "../server/media/transcript";
18
+ import { type AiFailure } from "./usage";
19
+ import { type AiUsage, type Transcriber, type Transcription } from "./types";
20
+ /** Where each chunk's answer is kept between runs (index into the chunks that are sent). Media-steps' shape. */
21
+ export interface TranscriptProgress {
22
+ get(i: number): Promise<WhisperResult | undefined>;
23
+ put(i: number, r: WhisperResult): Promise<void>;
24
+ }
25
+ export type TranscriptRun = {
26
+ done: true;
27
+ transcript: Transcript;
28
+ /** Chunks sent to the model, in all runs. */ sent: number;
29
+ /** Every chunk's call, this run's and those kept from earlier runs. */ usage: AiUsage[];
30
+ }
31
+ /** Out of time with chunks left: run again (their answers so far are in `progress`). */
32
+ | {
33
+ done: false;
34
+ sent: number;
35
+ left: number;
36
+ usage: AiUsage[];
37
+ };
38
+ /** Longest speech transcribed (seconds of chunks sent): every chunk's answer is held to assemble the transcript. */
39
+ export declare const MAX_TRANSCRIPT_S: number;
40
+ /** Biggest chunk loaded (bytes), whatever the transcriber takes. */
41
+ export declare const MAX_CHUNK_BYTES: number;
42
+ /** A transcriber answer → what a run keeps per chunk (media-steps' stored shape: small, and the same for every adapter). */
43
+ export declare function keepAnswer(t: Pick<Transcription, "segments" | "language">): WhisperResult;
44
+ export interface TranscribeAudioOptions {
45
+ transcriber: Transcriber;
46
+ /** audio/audio.json: the chunks and the silences measured with them. */
47
+ index: AudioIndex;
48
+ /** A chunk's bytes (null: missing, a permanent failure). */
49
+ load(chunk: AudioChunk): Promise<Uint8Array | null>;
50
+ /** MIME type of the chunks (the media worker's: "audio/ogg"). */
51
+ mimeType?: string;
52
+ /** VAD: StepSettings.vad (edgeS is what buildTranscript uses). */
53
+ vad: VadSettings;
54
+ language?: string;
55
+ prompt?: string;
56
+ /** The model name stamped on the transcript (default: the transcriber's). */
57
+ model?: string;
58
+ progress?: TranscriptProgress;
59
+ /** Epoch ms: no new call starts after it. */
60
+ deadline?: number;
61
+ now?: () => number;
62
+ maxSeconds?: number;
63
+ maxChunkBytes?: number;
64
+ signal?: AbortSignal;
65
+ onUsage?: (u: AiUsage) => void;
66
+ onFailure?: (f: AiFailure) => void;
67
+ }
68
+ export declare function transcribeAudio(o: TranscribeAudioOptions): Promise<TranscriptRun>;
69
+ /** How a transcript went, for a job's report: never a thrown error. */
70
+ export type TranscriptOutcome = {
71
+ state: "done";
72
+ transcript: Transcript;
73
+ usage: AiUsage[];
74
+ } | {
75
+ state: "pending";
76
+ left: number;
77
+ usage: AiUsage[];
78
+ } | {
79
+ state: "error";
80
+ reason: string;
81
+ permanent: boolean;
82
+ };
83
+ /** Runs `run` and turns any failure into {state: "error"}: a failed transcript never fails the video job. */
84
+ export declare function settleTranscript(run: () => Promise<TranscriptRun>): Promise<TranscriptOutcome>;
@@ -0,0 +1,141 @@
1
+ /**
2
+ * The two AI plugin interfaces: a Transcriber (speech → time-coded text) and an Embedder (texts or frames → vectors).
3
+ * Adapters implement them (Workers AI, any OpenAI-compatible endpoint, Ollama, local subprocesses in
4
+ * @uploadfilm/ai-plugins, a fake for tests); everything else in this folder takes ANY implementation.
5
+ *
6
+ * Rules every adapter keeps:
7
+ * - No credentials in the SDK: an adapter never reads the environment; a key, a binding or a fetch is injected.
8
+ * - Every call reports what it cost and how long it took (`usage`), so a product records it per adapter.
9
+ * - Failures are AiErrors saying whether retrying can help (`permanent`).
10
+ *
11
+ * Erasable TypeScript only: the media worker loads these files with plain Node (type stripping).
12
+ */
13
+ /** One piece of audio to transcribe: the bytes of a self-contained file (Opus in Ogg, WAV, MP3, …). */
14
+ export interface AudioInput {
15
+ bytes: Uint8Array;
16
+ /** Its MIME type ("audio/ogg" for the media worker's chunks). */
17
+ mimeType: string;
18
+ /** A file name some endpoints want for multipart uploads (default from the MIME type). */
19
+ name?: string;
20
+ /** Its length in seconds, when known (cost per audio minute, sanity checks). */
21
+ seconds?: number;
22
+ }
23
+ export interface TranscribeOptions {
24
+ /** Language to force (ISO 639-1, "en"); absent: detected. */
25
+ language?: string;
26
+ /** Vocabulary hint (names, jargon), for models that take one. */
27
+ prompt?: string;
28
+ signal?: AbortSignal;
29
+ }
30
+ export interface TranscriptionWord {
31
+ word: string;
32
+ /** Seconds from the start of the audio given. */
33
+ start: number;
34
+ end: number;
35
+ /** 0..1, when the model said. */
36
+ confidence?: number;
37
+ }
38
+ export interface TranscriptionSegment {
39
+ start: number;
40
+ end: number;
41
+ text: string;
42
+ words?: TranscriptionWord[];
43
+ /** 0..1: exp(average token log-probability), or the mean token probability. */
44
+ confidence?: number;
45
+ /** 0..1: the model's own "this is not speech" probability, when it said. */
46
+ noSpeechProb?: number;
47
+ }
48
+ /** One adapter call's answer. Times are seconds from the start of the audio given. */
49
+ export interface Transcription {
50
+ segments: TranscriptionSegment[];
51
+ /** ISO 639-1 code when the model named one ("en"). */
52
+ language?: string;
53
+ /** 0..1, when the model said how sure it was of the language. */
54
+ languageConfidence?: number;
55
+ text: string;
56
+ usage: AiUsage;
57
+ }
58
+ export interface Transcriber {
59
+ /** The adapter: "workers-ai", "openai-compatible", "whisper.cpp", "mlx-whisper", "fake", … */
60
+ readonly adapter: string;
61
+ readonly model: string;
62
+ /** Biggest audio input (bytes) it takes; bigger chunks fail permanently before any call. */
63
+ readonly maxBytes?: number;
64
+ transcribe(audio: AudioInput, options?: TranscribeOptions): Promise<Transcription>;
65
+ }
66
+ /** A frame (or any image) to embed. */
67
+ export interface ImageInput {
68
+ bytes: Uint8Array;
69
+ mimeType: string;
70
+ /** Seconds into the video, carried through to the vectors' owner. */
71
+ t?: number;
72
+ }
73
+ export type EmbedInput = {
74
+ kind: "text";
75
+ texts: string[];
76
+ } | {
77
+ kind: "image";
78
+ images: ImageInput[];
79
+ };
80
+ export interface EmbedOptions {
81
+ signal?: AbortSignal;
82
+ /** Some models embed queries and documents differently (e5, nomic): passed to adapters that support it. */
83
+ purpose?: "query" | "document";
84
+ }
85
+ export interface Embeddings {
86
+ /** One vector per input, in input order. */
87
+ vectors: number[][];
88
+ dimensions: number;
89
+ usage: AiUsage;
90
+ }
91
+ export interface Embedder {
92
+ readonly adapter: string;
93
+ readonly model: string;
94
+ /** What it embeds. */
95
+ readonly accepts: {
96
+ text: boolean;
97
+ image: boolean;
98
+ };
99
+ /** Most inputs per call (the helpers batch by it). */
100
+ readonly maxBatch?: number;
101
+ embed(input: EmbedInput, options?: EmbedOptions): Promise<Embeddings>;
102
+ }
103
+ /** What one call cost and took. */
104
+ export interface AiUsage {
105
+ adapter: string;
106
+ model: string;
107
+ kind: "transcribe" | "embed";
108
+ /** Wall-clock milliseconds of the call (request to parsed answer). */
109
+ latencyMs: number;
110
+ /** Seconds of audio sent (transcribe). */
111
+ audioSeconds?: number;
112
+ /** Inputs sent (embed). */
113
+ inputs?: number;
114
+ /** Tokens the provider reported. */
115
+ inputTokens?: number;
116
+ /** US dollars, when known: reported by the provider, or estimated from the product's injected price. */
117
+ costUsd?: number;
118
+ /** Where costUsd comes from. "local": a model on our own hardware (no per-call price). */
119
+ costSource: "reported" | "estimated" | "local" | "unknown";
120
+ }
121
+ /**
122
+ * Prices are the product's (they change; the SDK hardcodes none). Any combination: the estimate adds them up.
123
+ */
124
+ export interface AiPrice {
125
+ perAudioMinuteUsd?: number;
126
+ perMillionInputTokensUsd?: number;
127
+ perCallUsd?: number;
128
+ }
129
+ /** A failure; `permanent` when retrying the same input can't help (bad request, unsupported, too big, no auth). */
130
+ export declare class AiError extends Error {
131
+ readonly permanent: boolean;
132
+ readonly adapter: string;
133
+ readonly status?: number;
134
+ constructor(adapter: string, message: string, o: {
135
+ permanent: boolean;
136
+ status?: number;
137
+ cause?: unknown;
138
+ });
139
+ }
140
+ /** Is this HTTP status worth retrying? (Timeouts, rate limits, server errors.) */
141
+ export declare const transientStatus: (status: number) => boolean;
@@ -0,0 +1,66 @@
1
+ /**
2
+ * Cost and latency, per call and per adapter. Adapters fill an AiUsage for every call; a product passes an `onUsage`
3
+ * hook to the helpers (or wraps an adapter with `recorded`) and keeps the numbers wherever it keeps metrics. The SDK
4
+ * stores nothing.
5
+ */
6
+ import { AiError, type AiPrice, type AiUsage, type Embedder, type Transcriber } from "./types";
7
+ /** The product's price applied to what a call used; undefined when no price applies. */
8
+ export declare function estimateCost(price: AiPrice | undefined, used: {
9
+ audioSeconds?: number;
10
+ inputTokens?: number;
11
+ }): number | undefined;
12
+ /** costUsd/costSource from a provider-reported cost, else the estimate. */
13
+ export declare function costOf(price: AiPrice | undefined, used: {
14
+ audioSeconds?: number;
15
+ inputTokens?: number;
16
+ reportedUsd?: number;
17
+ }): Pick<AiUsage, "costUsd" | "costSource">;
18
+ /** Per-adapter totals. */
19
+ export interface UsageTotals {
20
+ adapter: string;
21
+ model: string;
22
+ kind: AiUsage["kind"];
23
+ calls: number;
24
+ failures: number;
25
+ latencyMs: {
26
+ total: number;
27
+ max: number;
28
+ };
29
+ audioSeconds: number;
30
+ inputs: number;
31
+ inputTokens: number;
32
+ /** Sum of the calls whose cost is known; `unpriced` calls are not in it. */
33
+ costUsd: number;
34
+ unpriced: number;
35
+ }
36
+ /** An in-memory tally keyed by adapter + model + kind (a product exports it to its own metrics). */
37
+ export declare function usageLog(): {
38
+ record(u: AiUsage): void;
39
+ failure(u: Pick<AiUsage, "adapter" | "model" | "kind" | "latencyMs">): void;
40
+ totals: () => UsageTotals[];
41
+ };
42
+ export type UsageLog = ReturnType<typeof usageLog>;
43
+ /** A failed call, for `onFailure`: who failed, after how long, and why. */
44
+ export interface AiFailure {
45
+ adapter: string;
46
+ model: string;
47
+ kind: AiUsage["kind"];
48
+ latencyMs: number;
49
+ error: AiError | Error;
50
+ }
51
+ /**
52
+ * The same transcriber, reporting every call (success: its usage; failure: adapter, latency, error). Hooks that throw
53
+ * are ignored: metrics never fail a transcript.
54
+ */
55
+ export declare function recordedTranscriber(t: Transcriber, hooks: {
56
+ onUsage?: (u: AiUsage) => void;
57
+ onFailure?: (f: AiFailure) => void;
58
+ now?: () => number;
59
+ }): Transcriber;
60
+ /** The same embedder, reporting every call. */
61
+ export declare function recordedEmbedder(x: Embedder, hooks: {
62
+ onUsage?: (u: AiUsage) => void;
63
+ onFailure?: (f: AiFailure) => void;
64
+ now?: () => number;
65
+ }): Embedder;
66
+ export declare function quietly(f: () => unknown): void;
@@ -0,0 +1,15 @@
1
+ import { type TranscriptionSegment } from "./types";
2
+ /** exp(average log-probability) as 0..1. */
3
+ export declare const confidenceFromLogprob: (lp: unknown) => number | undefined;
4
+ export interface ReadWhisper {
5
+ segments: TranscriptionSegment[];
6
+ language?: string;
7
+ languageConfidence?: number;
8
+ text: string;
9
+ }
10
+ /**
11
+ * Any Whisper-shaped answer → segments. `seconds`: the audio's length, for an answer with text but no segments (one
12
+ * segment spanning its words, else the whole audio). Times are kept as the model gave them (the transcript rounds them
13
+ * once, after adding the chunk's offset). Throws a non-permanent AiError for an answer that isn't one.
14
+ */
15
+ export declare function readWhisperJson(adapter: string, x: unknown, seconds?: number): ReadWhisper;
@@ -0,0 +1,56 @@
1
+ /**
2
+ * Workers AI: the media worker's transcript step (from media-steps) as an adapter. Takes the AI binding (env.AI) inside
3
+ * a Worker, or `workersAiRest` from anywhere else (a Node runner, a site) with the product's account id and token.
4
+ *
5
+ * Whisper: @cf/openai/whisper-large-v3-turbo (base64 audio) and the older @cf/openai/whisper / whisper-tiny-en
6
+ * (bytes as an array of numbers, so smaller chunks). Embeddings: any text model answering {data | response: number[][]}
7
+ * (@cf/baai/bge-*).
8
+ */
9
+ import { type Credential, type FetchLike } from "./http";
10
+ import { type AiPrice, type Embedder, type Transcriber } from "./types";
11
+ /** The part of Workers AI's binding (env.AI) the adapters use. */
12
+ export interface AiLike {
13
+ run(model: string, input: Record<string, unknown>): Promise<unknown>;
14
+ }
15
+ /** The media worker's model since media-steps. */
16
+ export declare const WORKERS_AI_WHISPER = "@cf/openai/whisper-large-v3-turbo";
17
+ /** Default call timeout (ms). */
18
+ export declare const WORKERS_AI_TIMEOUT_MS: number;
19
+ /** The request body for this model: base64 for whisper-large-v3-turbo (and newer), a byte array for the older ones. */
20
+ export declare function whisperInput(model: string, bytes: Uint8Array): Record<string, unknown>;
21
+ export declare function withTimeout<T>(p: Promise<T>, ms: number, adapter?: string): Promise<T>;
22
+ export interface WorkersAiTranscriberOptions {
23
+ ai: AiLike;
24
+ /** Default WORKERS_AI_WHISPER. */
25
+ model?: string;
26
+ /** The product's price (Workers AI bills neurons; set the per-minute equivalent of the model used). */
27
+ price?: AiPrice;
28
+ timeoutMs?: number;
29
+ now?: () => number;
30
+ }
31
+ export declare function workersAiTranscriber(o: WorkersAiTranscriberOptions): Transcriber;
32
+ export interface WorkersAiEmbedderOptions {
33
+ ai: AiLike;
34
+ /** The embedding model (e.g. "@cf/baai/bge-m3"): the product's choice, always named. */
35
+ model: string;
36
+ price?: AiPrice;
37
+ timeoutMs?: number;
38
+ /** Texts per call (Workers AI's text-embedding models take up to 100). */
39
+ maxBatch?: number;
40
+ now?: () => number;
41
+ }
42
+ export declare function workersAiEmbedder(o: WorkersAiEmbedderOptions): Embedder;
43
+ /** {data: number[][]} or {response: number[][]}: one finite vector per input, all the same length. */
44
+ export declare function readVectors(adapter: string, x: unknown, n: number): number[][];
45
+ export declare function checkVectors(adapter: string, rows: unknown, n: number): number[][];
46
+ /**
47
+ * Workers AI over its REST API (anywhere with fetch): an AiLike for the adapters above. The token is the product's,
48
+ * passed in (a string or a function returning one); the SDK never reads it from the environment.
49
+ */
50
+ export declare function workersAiRest(o: {
51
+ accountId: string;
52
+ token: Credential;
53
+ fetch?: FetchLike;
54
+ timeoutMs?: number;
55
+ baseUrl?: string;
56
+ }): AiLike;
package/dist/ai.js ADDED
@@ -0,0 +1 @@
1
+ "use strict";class e extends Error{permanent;adapter;status;constructor(e,t,n){super(`${e}: ${t}`.slice(0,1e3),void 0===n.cause?void 0:{cause:n.cause}),this.name="AiError",this.adapter=e,this.permanent=n.permanent,void 0!==n.status&&(this.status=n.status)}}const t=e=>408===e||409===e||425===e||429===e||e>=500;function n(e,t){if(!e)return;let n=0,a=!1;return void 0!==e.perCallUsd&&(n+=e.perCallUsd,a=!0),void 0!==e.perAudioMinuteUsd&&void 0!==t.audioSeconds&&(n+=e.perAudioMinuteUsd*t.audioSeconds/60,a=!0),void 0!==e.perMillionInputTokensUsd&&void 0!==t.inputTokens&&(n+=e.perMillionInputTokensUsd*t.inputTokens/1e6,a=!0),a?Math.round(1e8*n)/1e8:void 0}function a(e,t){if("number"==typeof t.reportedUsd&&Number.isFinite(t.reportedUsd)&&t.reportedUsd>=0)return{costUsd:t.reportedUsd,costSource:"reported"};const a=n(e,t);return void 0===a?{costSource:"unknown"}:{costUsd:a,costSource:"estimated"}}function r(e){try{const t=e();t?.catch?.(()=>{})}catch{}}const s={en:"english",zh:"chinese",de:"german",es:"spanish",ru:"russian",ko:"korean",fr:"french",ja:"japanese",pt:"portuguese",tr:"turkish",pl:"polish",ca:"catalan",nl:"dutch",ar:"arabic",sv:"swedish",it:"italian",id:"indonesian",hi:"hindi",fi:"finnish",vi:"vietnamese",he:"hebrew",uk:"ukrainian",el:"greek",ms:"malay",cs:"czech",ro:"romanian",da:"danish",hu:"hungarian",ta:"tamil",no:"norwegian",th:"thai",ur:"urdu",hr:"croatian",bg:"bulgarian",lt:"lithuanian",la:"latin",mi:"maori",ml:"malayalam",cy:"welsh",sk:"slovak",te:"telugu",fa:"persian",lv:"latvian",bn:"bengali",sr:"serbian",az:"azerbaijani",sl:"slovenian",kn:"kannada",et:"estonian",mk:"macedonian",br:"breton",eu:"basque",is:"icelandic",hy:"armenian",ne:"nepali",mn:"mongolian",bs:"bosnian",kk:"kazakh",sq:"albanian",sw:"swahili",gl:"galician",mr:"marathi",pa:"punjabi",si:"sinhala",km:"khmer",sn:"shona",yo:"yoruba",so:"somali",af:"afrikaans",oc:"occitan",ka:"georgian",be:"belarusian",tg:"tajik",sd:"sindhi",gu:"gujarati",am:"amharic",yi:"yiddish",lo:"lao",uz:"uzbek",fo:"faroese",ht:"haitian creole",ps:"pashto",tk:"turkmen",nn:"nynorsk",mt:"maltese",sa:"sanskrit",lb:"luxembourgish",my:"myanmar",bo:"tibetan",tl:"tagalog",mg:"malagasy",as:"assamese",tt:"tatar",haw:"hawaiian",ln:"lingala",ha:"hausa",ba:"bashkir",jw:"javanese",su:"sundanese",yue:"cantonese"},o=Object.fromEntries(Object.entries(s).map(([e,t])=>[t,e]));function i(e){if("string"!=typeof e)return;const t=e.trim().toLowerCase();if(!t)return;if(s[t])return t;const n=t.split(/[-_]/)[0];return/^[a-z]{2,3}$/.test(n)&&s[n]?n:o[t]}Object.assign(o,{burmese:"my",valencian:"ca",flemish:"nl",haitian:"ht",letzeburgesch:"lb",pushto:"ps",panjabi:"pa",moldavian:"ro",moldovan:"ro",sinhalese:"si",castilian:"es",mandarin:"zh"});const c=e=>"number"==typeof e&&Number.isFinite(e),d=e=>Math.min(1,Math.max(0,e)),m=e=>Math.round(1e3*e)/1e3,l=e=>c(e)?m(d(Math.exp(Math.min(0,e)))):void 0;function u(e){if(!e||"object"!=typeof e)return null;const t="string"==typeof e.word?e.word:"string"==typeof e.text?e.text:null;if(null===t||!t.trim()||!c(e.start)||!c(e.end))return null;const n=c(e.probability)?e.probability:c(e.confidence)?e.confidence:void 0;return{word:t.trim(),start:Math.max(0,e.start),end:Math.max(0,e.end),...void 0!==n?{confidence:m(d(n))}:{}}}function p(t,n,a){if(!n||"object"!=typeof n)throw new e(t,"not an answer",{permanent:!1});const r=n,s=i(r.transcription_info?.language??r.language),o=r.transcription_info?.language_probability??r.language_probability,p=c(o)?m(d(o)):void 0,h=Array.isArray(r.words)?r.words.map(u).filter(Boolean):[];let g;if(Array.isArray(r.segments)){g=[];for(const e of r.segments){if(!e||"object"!=typeof e||!c(e.start)||!c(e.end)||"string"!=typeof e.text)continue;const t=Array.isArray(e.words)?e.words.map(e=>u(e)).filter(Boolean):[],n=l(e.avg_logprob);g.push({start:Math.max(0,e.start),end:Math.max(0,e.start,e.end),text:e.text,...t.length?{words:t}:{},...void 0!==n?{confidence:n}:{},...c(e.no_speech_prob)?{noSpeechProb:m(d(e.no_speech_prob))}:{}})}if(h.length&&g.every(e=>!e.words)){let e=0;g.forEach((t,n)=>{const a=g[n+1]?.start??1/0,r=[];for(;e<h.length&&(h[e].start<a||n===g.length-1);)r.push(h[e++]);r.length&&(t.words=r)})}}else{if("string"!=typeof r.text)throw new e(t,"no segments or text in the answer",{permanent:!1});{const e=h.length?{start:h[0].start,end:h[h.length-1].end}:{start:0,end:a??0};g=r.text.trim()?[{...e,text:r.text,...h.length?{words:h}:{}}]:[]}}const f="string"==typeof r.text?r.text.trim():g.map(e=>e.text.trim()).join(" ");return{segments:g,...s?{language:s}:{},...void 0!==p?{languageConfidence:p}:{},text:f}}const h=async e=>"function"==typeof e?await e():e,g=67108864;async function f(n,a,r,s,o){if(o.signal?.aborted)throw new e(n,"aborted",{permanent:!0});const i=function(e,t){const n=new AbortController;let a=!1;const r=setTimeout(()=>{a=!0,n.abort(new Error(`no answer in ${Math.round(t/1e3)} s`))},t);r.unref?.();const s=()=>n.abort(e?.reason);return e?.aborted?n.abort(e.reason):e?.addEventListener("abort",s,{once:!0}),{signal:n.signal,timedOut:()=>a,done:()=>{clearTimeout(r),e?.removeEventListener("abort",s)}}}(o.signal,o.timeoutMs);let c,d;try{c=await a(r,{...s,signal:i.signal})}catch(t){if(i.done(),o.signal?.aborted&&!i.timedOut())throw new e(n,"aborted",{permanent:!0,cause:t});throw new e(n,i.timedOut()?`no answer in ${Math.round(o.timeoutMs/1e3)} s`:`network: ${t?.message??String(t)}`,{permanent:!1,cause:t})}try{d=await async function(e){const t=Number(e.headers.get("content-length"));if(Number.isFinite(t)&&t>g)throw new Error(`answer of ${t} bytes (at most 67108864)`);if(!e.body)return await e.text();const n=e.body.getReader(),a=[];let r=0;for(;;){const{done:e,value:t}=await n.read();if(e)break;if(r+=t.byteLength,r>g)throw await n.cancel().catch(()=>{}),new Error("answer over 67108864 bytes");a.push(t)}const s=new Uint8Array(r);let o=0;for(const e of a)s.set(e,o),o+=e.byteLength;return(new TextDecoder).decode(s)}(c)}catch(t){const a=i.timedOut();if(i.done(),await(c.body?.cancel().catch(()=>{})),o.signal?.aborted&&!a)throw new e(n,"aborted",{permanent:!0,cause:t});throw new e(n,a?`answer not finished in ${Math.round(o.timeoutMs/1e3)} s`:t.message,{permanent:!1,cause:t})}if(i.done(),!c.ok){const a=function(e){try{const t=JSON.parse(e),n=t.error,a="string"==typeof n?n:"string"==typeof n?.message?n.message:"string"==typeof t.message?t.message:void 0;return a?.slice(0,300)}catch{return e.trim().slice(0,200)||void 0}}(d)??c.statusText??"";throw new e(n,`HTTP ${c.status}${a?`: ${a}`:""}`,{permanent:!t(c.status),status:c.status})}try{return{json:JSON.parse(d),headers:c.headers}}catch{throw new e(n,"answer is not JSON",{permanent:!1,status:c.status})}}function w(e){let t="";for(let n=0;n<e.length;n+=32768)t+=String.fromCharCode(...e.subarray(n,n+32768));return btoa(t)}const b=e=>e.replace(/\/+$/,""),x="workers-ai",y="@cf/openai/whisper-large-v3-turbo",M=3e5,k=/^@cf\/openai\/whisper(?:-tiny-en)?$/,v=4194304,$=16777216;function S(t,n){if(k.test(t)){if(n.length>v)throw new e(x,`a ${n.length}-byte chunk is too big for ${t} (at most 4194304)`,{permanent:!0});return{audio:Array.from(n)}}return{audio:w(n)}}function A(t,n,a=x){let r;return Promise.race([t,new Promise((t,s)=>{r=setTimeout(()=>s(new e(a,`no answer in ${Math.round(n/1e3)} s`,{permanent:!1})),n),r.unref?.()})]).finally(()=>clearTimeout(r))}function T(t){if(t instanceof e)return t;const n=t?.message??String(t),a=/\b(?:invalid input|bad input|AiError: 400|unsupported)\b/i.test(n);return new e(x,n,{permanent:a,cause:t})}function U(e,t,n){const a=t;return _(e,Array.isArray(a?.data)?a.data:Array.isArray(a?.response)?a.response:null,n)}function _(t,n,a){if(!Array.isArray(n))throw new e(t,"no vectors in the answer",{permanent:!1});if(n.length!==a)throw new e(t,`${n.length} vectors for ${a} inputs`,{permanent:!1});let r=-1;for(const a of n){if(!Array.isArray(a)||!a.length||!a.every(e=>"number"==typeof e&&Number.isFinite(e)))throw new e(t,"a vector that isn't finite numbers",{permanent:!1});if(-1===r)r=a.length;else if(a.length!==r)throw new e(t,"vectors of different lengths",{permanent:!1})}return n}const j="openai-compatible";function B(t,n){let a;try{a=new URL(t.baseUrl)}catch{throw new e(n,"baseUrl is not a URL",{permanent:!0})}if("https:"!==a.protocol&&("http:"!==a.protocol||!/^(localhost|127\.0\.0\.1|\[::1\])$/.test(a.hostname)))throw new e(n,"baseUrl must be https (http only for localhost)",{permanent:!0});if(a.username||a.password||a.search)throw new e(n,"baseUrl must not carry credentials or a query",{permanent:!0});if(!t.model)throw new e(n,"a model must be named",{permanent:!0});return b(a.href)}async function E(e,t={}){const n=await h(e.apiKey);return{...e.headers,...t,...n?{authorization:`Bearer ${n}`}:{}}}const O={"audio/ogg":"ogg","audio/opus":"ogg","audio/wav":"wav","audio/x-wav":"wav","audio/mpeg":"mp3","audio/mp4":"m4a","audio/flac":"flac","audio/webm":"webm"};const C=e=>"number"==typeof e&&Number.isFinite(e)?e:void 0;const N="ollama";function I(e,t){const n=Array.from({length:t},(t,n)=>function(e,t){let n=2166136261^t;for(let t=0;t<e.length;t++)n^=e.charCodeAt(t),n=Math.imul(n,16777619)>>>0;return n>>>0}(e,n+1)/4294967295*2-1),a=Math.hypot(...n)||1;return n.map(e=>Math.round(e/a*1e6)/1e6)}const F=e=>Math.round(1e3*e)/1e3;function R(e,t,n){let a=0;for(const r of n)a+=Math.max(0,Math.min(t,r.end)-Math.max(e,r.start));return a}function P(e,t,n,a){const r=t-e;if(r<=0)return n.some(t=>e>=t.start&&e<=t.end);const s=R(e,t,n);return s>0&&s>=.8*r&&r-s<=a}function z(e){const t=Math.max(0,Math.round(1e3*e)),n=Math.floor(t/36e5),a=Math.floor(t/6e4)%60,r=Math.floor(t/1e3)%60,s=(e,t=2)=>String(e).padStart(t,"0");return`${s(n)}:${s(a)}:${s(r)}.${s(t%1e3,3)}`}const L=16777216;function D(e){return{...e.language?{transcription_info:{language:e.language}}:{},segments:e.segments.map(e=>({start:e.start,end:e.end,text:e.text,...e.words?.length?{words:e.words.map(e=>({word:e.word,start:e.start,end:e.end}))}:{}}))}}const W=(t,n)=>{if(n instanceof e)return n;const a=n?.permanent,r=String(n?.message??n),s=n?.adapter;return new e("string"==typeof s?s:t,"string"==typeof s&&r.startsWith(`${s}: `)?r.slice(s.length+2):r,{permanent:!0===a,cause:n})};const J=e=>Math.round(1e3*e)/1e3;function K(e,t,n){const a=e.text.trim();if(!a)return[];if(a.length<=t&&e.end-e.start<=n)return[{start:e.start,end:e.end,text:a}];const r=e.words?.filter(e=>e.word.trim()),s=[],o=e=>e.replace(/[^\p{L}\p{N}]+/gu,"");if(r?.length&&o(r.map(e=>e.word).join(""))===o(a)){let e=null;for(const a of r)for(const r of q(a.word.trim(),t))e&&(e.text.length+1+r.length>t||Math.max(e.end,a.end)-e.start>n)&&(s.push(e),e=null),e=e?{start:e.start,end:Math.max(e.end,a.end),text:`${e.text} ${r}`}:{start:a.start,end:a.end,text:r};return e&&s.push(e),s}const i=e.end-e.start,c=i>n?Math.max(1,Math.floor(a.length*n/i)):t;t=Math.min(t,c);let d=0;for(const n of function(e,t){const n=[];let a=e;for(;a.length>t;){const e=a.slice(0,t+1);let r=Math.max(e.lastIndexOf(". "),e.lastIndexOf("? "),e.lastIndexOf("! "),e.lastIndexOf("。"));r>0&&(r+=1),r>t&&(r=-1),r<=t/3&&(r=e.lastIndexOf(" ")),r<=0&&(r=t),r=V(a,r),n.push(a.slice(0,r).trim()),a=a.slice(r).trim()}a&&n.push(a);return n.filter(Boolean)}(a,t)){const t=a.indexOf(n,d),r=t<0?d:t;s.push({start:J(e.start+i*r/a.length),end:J(e.start+i*(r+n.length)/a.length),text:n}),d=r+n.length}return s}function q(e,t){if(e.length<=t)return[e];const n=[];for(let a=0;a<e.length;){const r=V(e,Math.min(e.length,a+t),a);n.push(e.slice(a,r)),a=r}return n}function V(e,t,n=0){const a=e.charCodeAt(t-1);return t<e.length&&a>=55296&&a<=56319&&t-1>n?t-1:t}function H(e,t){if(!(t.maxSeconds>0&&t.maxChars>0))throw new RangeError("maxSeconds and maxChars must be positive");const n=Math.max(0,t.overlapS??0),a=e.segments.flatMap(e=>K(e,t.maxChars,t.maxSeconds)).sort((e,t)=>e.start-t.start),r=[];let s=[];const o=()=>{if(!s.length)return;r.push({index:r.length,start:J(s[0].start),end:J(Math.max(...s.map(e=>e.end))),text:s.map(e=>e.text).join(" ")});const e=Math.max(...s.map(e=>e.end)),t=[];if(n>0)for(let a=s.length-1;a>0&&s[a].start>=e-n;a--)t.unshift(s[a]);s=t};let i=0;for(const e of a){const n=s[s.length-1],a=s.reduce((e,t)=>e+t.text.length+1,0)+e.text.length,r=e.end-(s[0]?.start??e.start),c=n&&void 0!==t.breakOnGapS&&e.start-n.end>=t.breakOnGapS;if(s.length&&(a>t.maxChars||r>t.maxSeconds||c)){i>0&&o(),i=0;const n=s.reduce((e,t)=>e+t.text.length+1,0)+e.text.length;(c||n>t.maxChars||e.end-(s[0]?.start??e.start)>t.maxSeconds)&&(s=[])}s.push(e),i++}return i>0&&o(),r}async function X(e,t,n,a){const s=Math.max(1,Math.min(n.batchSize??n.embedder.maxBatch??32,n.embedder.maxBatch??1/0)),o=Math.ceil(e.length/s),i=[];let c=0,d=0;try{for(let m=0;m<o;m++){if(n.signal?.aborted)throw Object.assign(new Error("aborted"),{permanent:!0});const l=e.slice(m*s,(m+1)*s),u=await n.embedder.embed(t(l),{...n.signal?{signal:n.signal}:{},...a?{purpose:a}:{}});if(u.vectors.length!==l.length)throw new Error(`${u.vectors.length} vectors for ${l.length} inputs`);if(d&&u.dimensions!==d)throw new Error(`vectors changed length (${d} → ${u.dimensions})`);d=u.dimensions,i.push(u.usage),r(()=>n.onUsage?.(u.usage)),await n.onEmbeddings({items:l,vectors:u.vectors,adapter:n.embedder.adapter,model:n.embedder.model,dimensions:u.dimensions,usage:u.usage,batch:m,batches:o}),c++}}catch(e){return{state:"error",reason:String(e?.message??e).slice(0,500),permanent:!!e?.permanent,delivered:c,usage:i}}return{state:"done",items:e.length,batches:o,dimensions:d,usage:i}}function Y(e){return e.embedder.accepts.text?X(e.passages,e=>({kind:"text",texts:e.map(e=>e.text)}),e,"document"):Promise.resolve({state:"error",reason:`${e.embedder.adapter}: does not embed text`,permanent:!0,delivered:0,usage:[]})}exports.AiError=e,exports.MAX_ANSWER_BYTES=g,exports.MAX_CHUNK_BYTES=L,exports.MAX_TRANSCRIPT_S=43200,exports.SDK_ENTRY="ai",exports.WORKERS_AI_TIMEOUT_MS=M,exports.WORKERS_AI_WHISPER=y,exports.confidenceFromLogprob=l,exports.costOf=a,exports.credential=h,exports.embedFrames=function(e){return e.embedder.accepts.image?X(e.frames,e=>({kind:"image",images:e}),e):Promise.resolve({state:"error",reason:`${e.embedder.adapter}: does not embed images`,permanent:!0,delivered:0,usage:[]})},exports.embedPassages=Y,exports.embedTranscript=async function(e){let t;try{t=H(e.transcript,e.passages)}catch(e){return{state:"error",reason:String(e?.message??e),permanent:!0,delivered:0,usage:[]}}return Y({...e,passages:t})},exports.estimateCost=n,exports.fakeEmbedder=function(t={}){const n=t.dimensions??8,a=t.model??"fake-embed",r=[],s=t.maxBatch??16;return{adapter:"fake",model:a,accepts:{text:!0,image:!!t.images},maxBatch:s,calls:r,async embed(o){const i=r.length;r.push(o);const c=t.failOn?.(i);if(c)throw c;const d="text"===o.kind?o.texts:o.images.map(e=>`${e.mimeType}:${e.bytes.length}:${Array.from(e.bytes.subarray(0,64)).join(",")}`);if("image"===o.kind&&!t.images)throw new e("fake","images are not supported",{permanent:!0});if(d.length>s)throw new e("fake",`${d.length} inputs (at most ${s})`,{permanent:!0});return{vectors:d.map(e=>I(e,n)),dimensions:n,usage:{adapter:"fake",model:a,kind:"embed",latencyMs:t.latencyMs??0,inputs:d.length,costSource:"unknown"}}}}},exports.fakeTranscriber=function(t){const n=[],a=t.model??"fake-whisper";return{adapter:"fake",model:a,...void 0!==t.maxBytes?{maxBytes:t.maxBytes}:{},calls:n,async transcribe(r,s={}){const o=n.length;if(n.push({audio:r,options:s}),s.signal?.aborted)throw new e("fake","aborted",{permanent:!0});const i="function"==typeof t.answers?await t.answers(r,o,s):t.answers[Math.min(o,t.answers.length-1)];if(i instanceof Error)throw i;if(!i)throw new e("fake","no scripted answer",{permanent:!0});const c=i.segments.map(e=>({...e,...e.words?{words:e.words.map(e=>({...e}))}:{}})),d=r.seconds;return{...i,segments:c,text:i.text??c.map(e=>e.text.trim()).join(" "),usage:{adapter:"fake",model:a,kind:"transcribe",latencyMs:t.latencyMs??0,...void 0!==d?{audioSeconds:d}:{},...void 0!==t.perAudioMinuteUsd&&void 0!==d?{costUsd:t.perAudioMinuteUsd*d/60,costSource:"estimated"}:{costSource:"unknown"}}}}}},exports.fakeVector=I,exports.keepAnswer=D,exports.languageCode=i,exports.ollamaEmbedder=function(t){if(!t.model)throw new e(N,"a model must be named",{permanent:!0});let n;try{n=`${b(new URL(t.baseUrl).href)}/api/embed`}catch{throw new e(N,"baseUrl is not a URL",{permanent:!0})}const a=t.fetch??((e,t)=>fetch(e,t)),r=t.now??Date.now,s=t.maxBatch??64;return{adapter:N,model:t.model,accepts:{text:!0,image:!1},maxBatch:s,async embed(o,i={}){if("text"!==o.kind)throw new e(N,"images are not supported",{permanent:!0});if(!o.texts.length)return{vectors:[],dimensions:0,usage:{adapter:N,model:t.model,kind:"embed",latencyMs:0,inputs:0,costUsd:0,costSource:"local"}};if(o.texts.length>s)throw new e(N,`${o.texts.length} texts in one call (at most ${s})`,{permanent:!0});const c=i.purpose?t.prefixes?.[i.purpose]??"":"",d=r(),{json:m}=await f(N,a,n,{method:"POST",headers:{"content-type":"application/json"},body:JSON.stringify({model:t.model,input:o.texts.map(e=>c+e),...t.keepAlive?{keep_alive:t.keepAlive}:{}})},{timeoutMs:t.timeoutMs??12e4,signal:i.signal}),l=m,u=_(N,l.embeddings,o.texts.length),p="number"==typeof l.prompt_eval_count?l.prompt_eval_count:void 0;return{vectors:u,dimensions:u[0].length,usage:{adapter:N,model:t.model,kind:"embed",latencyMs:r()-d,inputs:o.texts.length,...void 0!==p?{inputTokens:p}:{},costUsd:0,costSource:"local"}}}}},exports.openAiCompatibleEmbedder=function(t){const n=t.name??j,r=`${B(t,n)}/embeddings`,s=t.fetch??((e,t)=>fetch(e,t)),o=t.now??Date.now,i=t.maxBatch??64;return{adapter:n,model:t.model,accepts:{text:!0,image:!!t.images},maxBatch:i,async embed(c,d={}){const m="text"===c.kind?c.texts.length:c.images.length;if(m>i)throw new e(n,`${m} inputs in one call (at most ${i})`,{permanent:!0});if("image"===c.kind&&!t.images)throw new e(n,`${t.model}: images are not supported (set images: true for an endpoint that embeds them)`,{permanent:!0});const l=d.purpose?t.prefixes?.[d.purpose]??"":"",u="text"===c.kind?c.texts.map(e=>l+e):c.images.map(e=>({image:`data:${e.mimeType};base64,${w(e.bytes)}`})),p=o(),{json:h}=await f(n,s,r,{method:"POST",headers:await E(t,{"content-type":"application/json"}),body:JSON.stringify({model:t.model,input:u,encoding_format:"float",...t.dimensions?{dimensions:t.dimensions}:{}})},{timeoutMs:t.timeoutMs??6e4,signal:d.signal}),g=h;if(!Array.isArray(g.data))throw new e(n,"no data in the answer",{permanent:!1});const b=Array.from({length:m},()=>{});if(g.data.forEach((e,t)=>{const n="number"==typeof e?.index?e.index:t;Number.isInteger(n)&&n>=0&&n<m&&void 0===b[n]&&(b[n]=e.embedding)}),g.data.length!==m||b.some(e=>void 0===e))throw new e(n,`${g.data.length} vectors for ${m} inputs`,{permanent:!1});const x=_(n,b,m),y=C(g.usage?.prompt_tokens)??C(g.usage?.total_tokens);return{vectors:x,dimensions:x[0]?.length??0,usage:{adapter:n,model:t.model,kind:"embed",latencyMs:o()-p,inputs:m,...void 0!==y?{inputTokens:y}:{},...a(t.price,{inputTokens:y,reportedUsd:C(g.usage?.cost)})}}}}},exports.openAiCompatibleTranscriber=function(t){const n=t.name??j,r=`${B(t,n)}/audio/transcriptions`,s=t.fetch??((e,t)=>fetch(e,t)),o=t.now??Date.now,i=t.responseFormat??"verbose_json",c=t.maxBytes??26214400;return{adapter:n,model:t.model,maxBytes:c,async transcribe(d,m={}){if(d.bytes.length>c)throw new e(n,`${d.bytes.length} bytes (at most ${c})`,{permanent:!0});const l=new FormData,u=d.name??`audio.${O[d.mimeType.split(";")[0].trim()]??"bin"}`;l.append("file",new Blob([d.bytes],{type:d.mimeType}),u),l.append("model",t.model),l.append("response_format",i),"verbose_json"===i&&(l.append("timestamp_granularities[]","segment"),!1!==t.words&&l.append("timestamp_granularities[]","word")),m.language&&l.append("language",m.language),m.prompt&&l.append("prompt",m.prompt);const h=o(),{json:g}=await f(n,s,r,{method:"POST",headers:await E(t),body:l},{timeoutMs:t.timeoutMs??3e5,signal:m.signal}),w=g,b=d.seconds??C(w.duration)??C(w.usage?.seconds),x=p(n,g,b),y=C(w.usage?.input_tokens);return{...x,...m.language&&!x.language?{language:m.language}:{},usage:{adapter:n,model:t.model,kind:"transcribe",latencyMs:o()-h,...void 0!==b?{audioSeconds:b}:{},...void 0!==y?{inputTokens:y}:{},...a(t.price,{audioSeconds:b,inputTokens:y,reportedUsd:C(w.usage?.cost)})}}}}},exports.quietly=r,exports.readVectors=U,exports.readWhisperJson=p,exports.recordedEmbedder=function(e,t){const n=t.now??Date.now;return{adapter:e.adapter,model:e.model,accepts:e.accepts,...void 0!==e.maxBatch?{maxBatch:e.maxBatch}:{},async embed(a,s){const o=n();try{const n=await e.embed(a,s);return r(()=>t.onUsage?.(n.usage)),n}catch(a){throw r(()=>t.onFailure?.({adapter:e.adapter,model:e.model,kind:"embed",latencyMs:n()-o,error:a})),a}}}},exports.recordedTranscriber=function(e,t){const n=t.now??Date.now;return{adapter:e.adapter,model:e.model,...void 0!==e.maxBytes?{maxBytes:e.maxBytes}:{},async transcribe(a,s){const o=n();try{const n=await e.transcribe(a,s);return r(()=>t.onUsage?.(n.usage)),n}catch(a){throw r(()=>t.onFailure?.({adapter:e.adapter,model:e.model,kind:"transcribe",latencyMs:n()-o,error:a})),a}}}},exports.settleTranscript=async function(t){try{const e=await t();return e.done?{state:"done",transcript:e.transcript,usage:e.usage}:{state:"pending",left:e.left,usage:e.usage}}catch(t){const n=t instanceof e?t.permanent:!!t?.permanent;return{state:"error",reason:String(t?.message??t).slice(0,500),permanent:n}}},exports.toWebVtt=function(e){const t=["WEBVTT",""];return e.segments.forEach((e,n)=>{const a=e.end>e.start?e.end:e.start+.5;t.push(String(n+1),`${z(e.start)} --\x3e ${z(a)}`,(e=>e.replace(/\r\n?/g,"\n").replace(/\n\s*\n/g,"\n").replace(/-->/g,"→").replace(/&/g,"&amp;").replace(/</g,"&lt;").replace(/>/g,"&gt;").trim())(e.text),"")}),t.join("\n")},exports.transcribeAudio=async function(t){const n=t.transcriber,a=t.index,s=[],o=[];let i=0;const c=a.chunks.filter(e=>{const t=((e,t)=>e.end-e.start-R(e.start,e.end,t)<=.05)(e,a.silences);return t&&i++,!t});if(c.length&&!n?.transcribe)throw new e("transcript","no transcriber",{permanent:!0});const d=c.reduce((e,t)=>e+(t.end-t.start),0),m=t.maxSeconds??43200;if(d>m)throw new e("transcript",`${Math.round(d/360)/10} h of audio (at most ${m/3600} h)`,{permanent:!0});const l=Math.min(t.maxChunkBytes??L,n?.maxBytes??1/0),u=t.now??Date.now,p=new Map,h=t.progress??{get:async e=>p.get(e),put:async(e,t)=>{p.set(e,t)}};for(let a=0;a<c.length;a++){const i=c[a],d=await h.get(a);if(d){const e=d.usage;e&&s.push(e),o.push({chunk:i,result:d});continue}if(t.signal?.aborted)throw new e("transcript","aborted",{permanent:!0});if(void 0!==t.deadline&&u()>t.deadline)return{done:!1,sent:a,left:c.length-a,usage:s};const m=await t.load(i);if(!m)throw new e("transcript",`${i.file} missing`,{permanent:!0});if(m.length>l)throw new e("transcript",`${i.file} is ${m.length} bytes (at most ${l} for ${n.adapter})`,{permanent:!0});const p=u();let g;try{g=await n.transcribe({bytes:m,mimeType:t.mimeType??"audio/ogg",name:i.file.split("/").pop(),seconds:Math.round(1e3*(i.end-i.start))/1e3},{...t.language?{language:t.language}:{},...t.prompt?{prompt:t.prompt}:{},...t.signal?{signal:t.signal}:{}})}catch(e){const a=W(n.adapter,e);throw r(()=>t.onFailure?.({adapter:n.adapter,model:n.model,kind:"transcribe",latencyMs:u()-p,error:a})),a}if(!g||!Array.isArray(g.segments))throw new e(n.adapter,"not a transcription",{permanent:!1});s.push(g.usage),r(()=>t.onUsage?.(g.usage));const f=D(g);await h.put(a,g.usage?{...f,usage:g.usage}:f),o.push({chunk:i,result:f})}return{done:!0,transcript:function(e,t){const n=[];let a=0,r=e.language;for(const{chunk:s,result:o}of e.results){r??=o.transcription_info?.language;const i=s.end-s.start;for(const r of o.segments??[]){const o=String(r.text??"").trim();if(!o||!Number.isFinite(r.start)||!Number.isFinite(r.end))continue;const c=F(s.start+Math.max(0,Math.min(r.start,i))),d=F(s.start+Math.max(0,Math.min(Math.max(r.end,r.start),i)));if(P(c,d,e.silences,t.edgeS)){a++;continue}const m=(r.words??[]).filter(e=>e&&"string"==typeof e.word&&Number.isFinite(e.start)&&Number.isFinite(e.end)).map(e=>({word:e.word.trim(),start:F(s.start+Math.max(0,e.start)),end:F(s.start+Math.max(0,e.end))})).filter(t=>t.word&&!P(t.start,t.end,e.silences,0));n.push({start:c,end:Math.max(c,d),text:o,...m.length?{words:m}:{}})}}return n.sort((e,t)=>e.start-t.start),{format:1,model:e.model,...r?{language:r}:{},seconds:F(e.seconds),segments:n,text:n.map(e=>e.text).join(" "),dropped:a,silentChunks:e.silentChunks??0}}({model:t.model??n?.model??"none",seconds:a.seconds,silences:a.silences,results:o,silentChunks:i,...t.language?{language:t.language}:{}},t.vad),sent:c.length,usage:s}},exports.transcriptPassages=H,exports.transientStatus=t,exports.usageLog=function(){const e=new Map,t=t=>{const n=`${t.kind}\0${t.adapter}\0${t.model}`;let a=e.get(n);return a||e.set(n,a={adapter:t.adapter,model:t.model,kind:t.kind,calls:0,failures:0,latencyMs:{total:0,max:0},audioSeconds:0,inputs:0,inputTokens:0,costUsd:0,unpriced:0}),a};return{record(e){const n=t(e);n.calls++,n.latencyMs.total+=e.latencyMs,n.latencyMs.max=Math.max(n.latencyMs.max,e.latencyMs),n.audioSeconds+=e.audioSeconds??0,n.inputs+=e.inputs??0,n.inputTokens+=e.inputTokens??0,void 0!==e.costUsd?n.costUsd=Math.round(1e8*(n.costUsd+e.costUsd))/1e8:"local"!==e.costSource&&n.unpriced++},failure(e){const n=t(e);n.failures++,n.latencyMs.total+=e.latencyMs,n.latencyMs.max=Math.max(n.latencyMs.max,e.latencyMs)},totals:()=>[...e.values()].map(e=>({...e,latencyMs:{...e.latencyMs}}))}},exports.whisperInput=S,exports.workersAiEmbedder=function(t){if(!t.model)throw new e(x,"an embedding model must be named",{permanent:!0});const n=t.now??Date.now,r=t.maxBatch??100;return{adapter:x,model:t.model,accepts:{text:!0,image:!1},maxBatch:r,async embed(s,o={}){if("text"!==s.kind)throw new e(x,`${t.model}: images are not supported`,{permanent:!0});if(s.texts.length>r)throw new e(x,`${s.texts.length} texts in one call (at most ${r})`,{permanent:!0});if(o.signal?.aborted)throw new e(x,"aborted",{permanent:!0});const i=n();let c;try{c=await A(t.ai.run(t.model,{text:s.texts}),t.timeoutMs??6e4)}catch(e){throw T(e)}const d=U(x,c,s.texts.length);return{vectors:d,dimensions:d[0]?.length??0,usage:{adapter:x,model:t.model,kind:"embed",latencyMs:n()-i,inputs:s.texts.length,...a(t.price,{})}}}}},exports.workersAiRest=function(t){const n=t.fetch??((e,t)=>fetch(e,t)),a=b(t.baseUrl??"https://api.cloudflare.com/client/v4");if(!/^[0-9a-f]{32}$/i.test(t.accountId))throw new e(x,"account id must be 32 hex characters",{permanent:!0});return{async run(r,s){if(!/^@cf\/[\w.-]+\/[\w.@-]+$/.test(r))throw new e(x,`not a Workers AI model id: ${r}`,{permanent:!0});const o=await h(t.token);if(!o)throw new e(x,"no API token",{permanent:!0});const{json:i}=await f(x,n,`${a}/accounts/${t.accountId}/ai/run/${r}`,{method:"POST",headers:{authorization:`Bearer ${o}`,"content-type":"application/json"},body:JSON.stringify(s)},{timeoutMs:t.timeoutMs??M}),c=i;if(!1===c.success)throw new e(x,c.errors?.map(e=>e.message).join("; ")||"failed",{permanent:!1});return c.result??i}}},exports.workersAiTranscriber=function(t){const n=t.model??y,r=t.now??Date.now;return{adapter:x,model:n,maxBytes:k.test(n)?v:$,async transcribe(s,o={}){if(!t.ai?.run)throw new e(x,"no AI binding",{permanent:!0});if(o.signal?.aborted)throw new e(x,"aborted",{permanent:!0});const i=k.test(n)?v:$;if(s.bytes.length>i)throw new e(x,`a ${s.bytes.length}-byte chunk is too big for ${n} (at most ${i})`,{permanent:!0});const c=S(n,s.bytes);o.language&&(c.language=o.language),o.prompt&&(c.initial_prompt=o.prompt);const d=r();let m;try{m=await A(t.ai.run(n,c),t.timeoutMs??M)}catch(e){throw T(e)}const l=p(x,m,s.seconds),u=m.transcription_info,h=s.seconds??("number"==typeof u?.duration?u.duration:void 0);return{...l,usage:{adapter:x,model:n,kind:"transcribe",latencyMs:r()-d,...void 0!==h?{audioSeconds:h}:{},...a(t.price,{audioSeconds:h})}}}}};