@effect-uai/google 0.6.0 → 0.8.0

package/src/LyriaGenerator.ts

@@ -1,11 +1,10 @@
 import {
-  Array as Arr,
   Context,
+  Duration,
   Effect,
   Encoding,
   Layer,
   Match,
-  Option,
   Redacted,
   Result,
   Schema,
@@ -14,13 +13,16 @@ import {
 import { HttpClient, HttpClientRequest } from "effect/unstable/http"
 import * as AiError from "@effect-uai/core/AiError"
 import type { AudioChunk, AudioFormat } from "@effect-uai/core/Audio"
+import { warnDroppedWhen } from "@effect-uai/core/Capabilities"
 import type {
   CommonGenerateMusicRequest,
   CommonStreamGenerateMusicRequest,
+  GenerateResult,
   MusicResult,
   MusicSessionInput,
-  WeightedPrompt,
+  MusicStreamEvent,
 } from "@effect-uai/core/Music"
+import { singleVariant } from "@effect-uai/core/Music"
 import { MusicGenerator, type MusicGeneratorService } from "@effect-uai/core/MusicGenerator"
 import type { LyriaModel } from "./models.js"
 
@@ -29,25 +31,34 @@ import type { LyriaModel } from "./models.js"
 // ---------------------------------------------------------------------------
 
 /**
- * Lyria-typed generate request. Narrows `model` to `LyriaModel`. The
- * Lyria sync API (`generateContent`) accepts only `audio/mp3` and
- * `audio/wav` output; `bpm`, `scale`, `instrumental` etc. on the
- * `CommonGenerateMusicRequest` are flattened into the prompt text
- * because the public API does not expose them as structured fields.
+ * Lyria-typed generate request. Narrows `model` to `LyriaModel`.
+ *
+ * Lyria 3 sync `generateContent` has no structured wire fields for
+ * `lyrics` / `duration` / `seed` / `instrumental` / `bpm` / `scale` /
+ * etc. Those fields on the Common request are logged as
+ * dropped-bucket-2 hints; the adapter never rewrites or augments the
+ * caller's prompt on their behalf. If you want lyrics or vocal
+ * suppression, include them in your `prompt` text yourself.
+ *
+ * Output format: only `audio/mp3` and `audio/wav` are accepted; other
+ * containers fail `Unsupported`. The Clip model variant is fixed at
+ * `audio/mp3`.
  */
 export type LyriaGenerateRequest = Omit<CommonGenerateMusicRequest, "model"> & {
   readonly model: LyriaModel
 }
 
 export type LyriaGeneratorService = {
-  readonly generate: (request: LyriaGenerateRequest) => Effect.Effect<MusicResult, AiError.AiError>
+  readonly generate: (
+    request: LyriaGenerateRequest,
+  ) => Effect.Effect<GenerateResult, AiError.AiError>
   readonly streamGeneration: (
     request: LyriaGenerateRequest,
   ) => Stream.Stream<AudioChunk, AiError.AiError>
   readonly streamGenerationFrom: <E, R>(
     input: Stream.Stream<MusicSessionInput, E, R>,
     request: CommonStreamGenerateMusicRequest,
-  ) => Stream.Stream<AudioChunk, AiError.AiError | E, R>
+  ) => Stream.Stream<MusicStreamEvent, AiError.AiError | E, R>
 }
 
 export class LyriaGenerator extends Context.Service<LyriaGenerator, LyriaGeneratorService>()(
@@ -60,7 +71,7 @@ export type Config = {
 }
 
 // ---------------------------------------------------------------------------
-// Codec — request → prompt text + response_format
+// Codec: AudioFormat → Lyria mimeType
 // ---------------------------------------------------------------------------
 
 export type LyriaResponseMimeType = "audio/mp3" | "audio/wav"
@@ -85,10 +96,10 @@ export const containerToMimeType: (
 )
 
 /**
- * Is this a Lyria 3 Clip model variant? Clip is fixed at 30 s MP3 — it
+ * Is this a Lyria 3 Clip model variant? Clip is fixed at 30 s MP3, it
  * does not accept a `responseFormat` on `generationConfig` (the live
- * API rejects unknown fields, despite what the docs imply). Pro
- * accepts the documented `responseFormat` knob.
+ * API rejects unknown fields). Pro accepts the documented
+ * `responseFormat` knob.
  */
 const isClipModel = (model: string): boolean => model.includes("clip")
 
@@ -110,38 +121,51 @@ export const realizedFormat: (mime: LyriaResponseMimeType) => AudioFormat =
     Match.exhaustive,
   )
 
+// ---------------------------------------------------------------------------
+// Bucket-2 warn-and-drop for Common fields Lyria 3 can't honor
+// ---------------------------------------------------------------------------
+
 /**
- * Collapse `prompts | WeightedPrompt[]` into a single prompt string and
- * splice in non-structural hints (`lyrics`, `bpm`, `scale`,
- * `instrumental`, `durationSeconds`). The Lyria 3 sync API only takes
- * one text part — every condition lives inline.
+ * Warn (don't fail) when the caller passed Common fields Lyria 3 has
+ * no wire field for. Bucket-2 per capabilities policy: provider has
+ * no structured interpretation; the output is still valid music, just
+ * less aligned with the caller's hint. We never embed these into the
+ * prompt on the caller's behalf, prompt construction is the
+ * developer's job.
  */
-export const buildPrompt = (request: CommonGenerateMusicRequest): string => {
-  const promptText = Match.value(request.prompts).pipe(
-    Match.when(Match.string, (s) => s),
-    Match.orElse((ws: ReadonlyArray<WeightedPrompt>) =>
-      Arr.map(ws, (w) =>
-        w.weight !== undefined && w.weight !== 1 ? `${w.text} (weight ${w.weight})` : w.text,
-      ).join(". "),
-    ),
+const warnDroppedHints = (request: LyriaGenerateRequest): Effect.Effect<void> =>
+  Effect.all(
+    [
+      warnDroppedWhen(request.lyrics, {
+        provider: "lyria",
+        capability: "lyrics",
+        field: "lyrics",
+        reason:
+          "Lyria 3 sync has no `lyrics` wire field. If you want vocals to follow specific lyrics, embed them in your prompt yourself (e.g. with `[Verse]` / `[Chorus]` tags).",
+      }),
+      warnDroppedWhen(
+        request.duration !== undefined ? Duration.toMillis(request.duration) : undefined,
+        {
+          provider: "lyria",
+          capability: "duration",
+          field: "duration",
+          reason:
+            "Lyria 3 clip is fixed at 30 s; pro derives duration from prompt content. There is no wire field for duration. If you want to steer length, mention it in your prompt.",
+        },
+      ),
+      warnDroppedWhen(request.seed, {
+        provider: "lyria",
+        capability: "seed",
+        field: "seed",
+        reason:
+          "Lyria 3 (Gemini surface) does not expose a seed parameter. Output is non-deterministic across identical prompts. Use Lyria 2 (Vertex) for seeded generation.",
+      }),
+    ],
+    { discard: true },
   )
-  const hints = Arr.getSomes([
-    request.instrumental === true ? Option.some("Instrumental only — no vocals.") : Option.none(),
-    request.bpm !== undefined ? Option.some(`BPM: ${request.bpm}.`) : Option.none(),
-    request.scale !== undefined ? Option.some(`Key/scale: ${request.scale}.`) : Option.none(),
-    request.durationSeconds !== undefined
-      ? Option.some(`Target duration: ${request.durationSeconds}s.`)
-      : Option.none(),
-  ])
-  const lyricsBlock =
-    request.lyrics !== undefined && request.instrumental !== true
-      ? `\n\nLyrics:\n${request.lyrics}`
-      : ""
-  return [promptText, ...hints].join(" ") + lyricsBlock
-}
 
 // ---------------------------------------------------------------------------
-// Wire response schema — just enough to find audio + optional lyrics text
+// Wire response schema
 // ---------------------------------------------------------------------------
 
 const InlineData = Schema.Struct({
@@ -245,26 +269,23 @@ const generateImpl =
   (cfg: Config) =>
   (
     request: LyriaGenerateRequest,
-  ): Effect.Effect<MusicResult, AiError.AiError, HttpClient.HttpClient> =>
+  ): Effect.Effect<GenerateResult, AiError.AiError, HttpClient.HttpClient> =>
     Effect.gen(function* () {
       const client = yield* HttpClient.HttpClient
       const requestedContainer = (request.outputFormat?.container ??
         "mp3") satisfies AudioFormat["container"]
-      const mimeType = yield* containerToMimeType(requestedContainer)
-      // Clip is fixed 30 s mp3 — reject wav up front rather than sending
-      // a `responseFormat` that the live API ignores or rejects.
-      if (isClipModel(request.model) && mimeType === "audio/wav") {
-        return yield* new AiError.Unsupported({
-          provider: "lyria",
-          capability: "outputFormat",
-          reason: `lyria-3-clip-preview is fixed at audio/mp3. Use lyria-3-pro-preview for audio/wav output.`,
-        })
-      }
-      const promptText = buildPrompt(request)
+      const requestedMime = yield* containerToMimeType(requestedContainer)
+      // lyria-3-clip-preview has no `responseFormat` wire field and always
+      // emits mp3, so there is no format to send and no 400 to surface. Rather
+      // than keep a per-model table that rejects `wav`, report the format the
+      // model actually produces (mp3) so the result is never mislabeled.
+      // lyria-3-pro honours the requested container.
+      const mimeType = isClipModel(request.model) ? ("audio/mp3" as const) : requestedMime
+      yield* warnDroppedHints(request)
       const body = isClipModel(request.model)
-        ? { contents: [{ parts: [{ text: promptText }] }] }
+        ? { contents: [{ parts: [{ text: request.prompt }] }] }
         : {
-            contents: [{ parts: [{ text: promptText }] }],
+            contents: [{ parts: [{ text: request.prompt }] }],
             generationConfig: {
               responseModalities: ["AUDIO", "TEXT"],
               responseFormat: { audio: { mimeType } },
@@ -290,10 +311,6 @@ const generateImpl =
       )
       const { inline, text } = extractParts(wire)
       if (inline === undefined) {
-        // No audio came back. Most common cause: Lyria filtered the prompt
-        // (named-artist references, copyrighted lyrics, etc.). Surface the
-        // full response body — text parts, finish reasons, prompt-feedback
-        // blocks — so the caller can see why.
         return yield* new AiError.GenerationFailed({
           provider: "lyria",
           raw: {
@@ -305,12 +322,13 @@ const generateImpl =
         })
       }
       const bytes = yield* decodeBase64ToBytes(inline.data)
-      return {
-        format: realizedFormat(mimeType),
-        bytes,
+      const result: MusicResult = {
+        audio: { format: realizedFormat(mimeType), bytes },
+        provider: "lyria",
         ...(text !== undefined && { lyrics: text }),
-        watermark: { kind: "synthid" },
-      } satisfies MusicResult
+        watermark: "synthid",
+      }
+      return singleVariant(result)
     })
 
 const streamGenerationImpl =
@@ -320,22 +338,22 @@ const streamGenerationImpl =
   ): Stream.Stream<AudioChunk, AiError.AiError, HttpClient.HttpClient> =>
     Stream.unwrap(
       Effect.map(generateImpl(cfg)(request), (result) =>
-        Stream.succeed<AudioChunk>({ bytes: result.bytes }),
+        Stream.succeed<AudioChunk>({ bytes: result.primary.audio.bytes }),
       ),
     )
 
 /**
- * Lyria 3 sync has no bidirectional session at the wire level — that's
- * Lyria RealTime (Phase 1b, separate Layer). The provider Layer here
- * also does NOT register `MusicInteractiveSession`, so callers using
+ * Lyria 3 sync has no bidirectional session at the wire level, that's
+ * Lyria RealTime (separate Layer). The provider Layer here also does
+ * NOT register `MusicInteractiveSession`, so callers using
  * `MusicGenerator.streamGenerationFrom` against only this Layer get a
  * compile-time error before this runtime fallback ever fires.
  */
 const streamGenerationFromUnsupported = <E, R>(
   _input: Stream.Stream<MusicSessionInput, E, R>,
   _request: CommonStreamGenerateMusicRequest,
-): Stream.Stream<AudioChunk, AiError.AiError | E, R> => {
-  const fail: Stream.Stream<AudioChunk, AiError.AiError | E, R> = Stream.fail(
+): Stream.Stream<MusicStreamEvent, AiError.AiError | E, R> => {
+  const fail: Stream.Stream<MusicStreamEvent, AiError.AiError | E, R> = Stream.fail(
     new AiError.Unsupported({
       provider: "lyria",
       capability: "streamGenerationFrom",
@@ -365,7 +383,7 @@ export const make = (
  * Layer that registers both the provider-specific `LyriaGenerator` tag
  * and the generic `MusicGenerator` tag.
  *
- * Does NOT register the `MusicInteractiveSession` capability marker —
+ * Does NOT register the `MusicInteractiveSession` capability marker,
  * Lyria 3 sync has no bidirectional session at the wire level. Code
  * calling `MusicGenerator.streamGenerationFrom` will fail to typecheck
  * against this Layer alone, which is the intended UX.
@@ -380,9 +398,10 @@ export const layer = (
       Effect.map(
         make(cfg),
         (s): MusicGeneratorService => ({
-          generate: (req: CommonGenerateMusicRequest) => s.generate(req as LyriaGenerateRequest),
+          generate: (req: CommonGenerateMusicRequest) =>
+            s.generate({ ...req, model: req.model as LyriaModel }),
           streamGeneration: (req: CommonStreamGenerateMusicRequest) =>
-            s.streamGeneration(req as LyriaGenerateRequest),
+            s.streamGeneration({ ...req, model: req.model as LyriaModel }),
           streamGenerationFrom: s.streamGenerationFrom,
         }),
       ),

package/src/codec.ts

@@ -8,6 +8,7 @@ import type {
 } from "@effect-uai/core/Items"
 import type { ToolDescriptor } from "@effect-uai/core/Tool"
 import type { Turn } from "@effect-uai/core/Turn"
+import type { CommonRequest } from "@effect-uai/core/LanguageModel"
 
 // ---------------------------------------------------------------------------
 // Wire schemas - the subset of Gemini's streamGenerateContent payload we
@@ -341,10 +342,39 @@ const toolDescriptorsToTools = (tools: ReadonlyArray<ToolDescriptor>): ReadonlyA
         },
       ]
 
+/**
+ * Map our cross-provider `toolChoice` onto Gemini's `functionCallingConfig`.
+ * Gemini honours the same three modes plus a function allow-list, so the
+ * choice is mapped, not dropped. `undefined` keeps the API default (`AUTO`).
+ */
+const toolChoiceToFunctionCallingConfig = (
+  toolChoice: CommonRequest["toolChoice"],
+): RequestToolConfig["functionCallingConfig"] =>
+  Match.value(toolChoice ?? "auto").pipe(
+    Match.when("auto", () => ({ mode: "AUTO" as const })),
+    Match.when("required", () => ({ mode: "ANY" as const })),
+    Match.when("none", () => ({ mode: "NONE" as const })),
+    Match.orElse((tc) => ({ mode: "ANY" as const, allowedFunctionNames: [tc.name] })),
+  )
+
+/**
+ * URL-form images need Gemini's Files API (upload, then `fileData`); we
+ * don't pre-upload, so such a part would otherwise vanish from the request.
+ * The caller fails the request with `Unsupported` instead of silently
+ * dropping the image.
+ */
+export const hasUrlImageSource = (history: ReadonlyArray<HistoryItem>): boolean =>
+  history.some(
+    (item) =>
+      item.type === "message" &&
+      item.content.some((b) => b.type === "input_image" && b.source._tag === "url"),
+  )
+
 export const buildRequestBody = (
   history: ReadonlyArray<HistoryItem>,
   generationConfig: Option.Option<GenerationConfig>,
   tools: ReadonlyArray<ToolDescriptor> = [],
+  toolChoice: CommonRequest["toolChoice"] = undefined,
 ): RequestBody => {
   const systemTexts = pipe(allMessages(history), Arr.filterMap(systemMessageText))
   const contents = pipe(history, Arr.filterMap(itemToContent(history)))
@@ -360,7 +390,7 @@ export const buildRequestBody = (
     }),
     ...(requestTools.length > 0 && {
       tools: requestTools,
-      toolConfig: { functionCallingConfig: { mode: "AUTO" as const } },
+      toolConfig: { functionCallingConfig: toolChoiceToFunctionCallingConfig(toolChoice) },
     }),
   }
 }

package/src/index.ts

@@ -1,7 +1,6 @@
 export * as Gemini from "./Gemini.js"
 export * as GeminiEmbedding from "./GeminiEmbedding.js"
 export * as GeminiSynthesizer from "./GeminiSynthesizer.js"
-export * as GeminiTranscriber from "./GeminiTranscriber.js"
 export * as LyriaGenerator from "./LyriaGenerator.js"
 export * as codec from "./codec.js"
 export * from "./models.js"

package/src/models.ts

@@ -44,23 +44,6 @@ export type GeminiTtsModel =
   // eslint-disable-next-line @typescript-eslint/ban-types
   | (string & {})
 
-/**
- * Known Gemini models that accept audio input via `inlineData` /
- * `fileData` on `generateContent`. Used by `GeminiTranscriber` —
- * transcription is prompt-driven on top of these models, not a
- * dedicated endpoint.
- *
- * Reference: https://ai.google.dev/gemini-api/docs/audio
- */
-export type GeminiSttModel =
-  | "gemini-3-flash-preview"
-  | "gemini-3.1-pro-preview"
-  | "gemini-3.1-flash-lite-preview"
-  | "gemini-2.5-flash"
-  | "gemini-2.5-pro"
-  // eslint-disable-next-line @typescript-eslint/ban-types
-  | (string & {})
-
 /**
  * 30 prebuilt voice names for Gemini TTS. No custom-voice / cloning path
  * is exposed on this surface, so the tail is stock-only (no

package/dist/GeminiTranscriber.d.mts

@@ -1,42 +0,0 @@
-import { GeminiSttModel } from "./models.mjs";
-import { Context, Effect, Layer, Redacted } from "effect";
-import { HttpClient } from "effect/unstable/http";
-import * as AiError from "@effect-uai/core/AiError";
-import { CommonTranscribeRequest, TranscriberService } from "@effect-uai/core/Transcriber";
-import { TranscriptResult } from "@effect-uai/core/Transcript";
-
-//#region src/GeminiTranscriber.d.ts
-declare namespace GeminiTranscriber_d_exports {
-  export { Config, GeminiTranscribeRequest, GeminiTranscriber, GeminiTranscriberService, buildPrompt, layer, make };
-}
-/**
- * Gemini transcription rides on `:generateContent` — the model receives
- * audio plus a textual "transcribe verbatim" prompt and returns plain
- * text. There is no native structured timestamp / diarization channel,
- * so `wordTimestamps` / `diarization` short-circuit with `Unsupported`.
- */
-type GeminiTranscribeRequest = Omit<CommonTranscribeRequest, "model"> & {
-  readonly model: GeminiSttModel;
-};
-type GeminiTranscriberService = {
-  readonly transcribe: (r: GeminiTranscribeRequest) => Effect.Effect<TranscriptResult, AiError.AiError>;
-  readonly streamTranscriptionFrom: TranscriberService["streamTranscriptionFrom"];
-};
-declare const GeminiTranscriber_base: Context.ServiceClass<GeminiTranscriber, "@betalyra/effect-uai/providers/google/GeminiTranscriber", GeminiTranscriberService>;
-declare class GeminiTranscriber extends GeminiTranscriber_base {}
-type Config = {
-  readonly apiKey: Redacted.Redacted;
-  readonly baseUrl?: string;
-};
-declare const buildPrompt: (r: GeminiTranscribeRequest) => string;
-declare const make: (cfg: Config) => Effect.Effect<GeminiTranscriberService, never, HttpClient.HttpClient>;
-/**
- * Layer registers both `GeminiTranscriber` and the generic
- * `Transcriber`. Does NOT register `SttStreaming` — Gemini's transcription
- * is sync-only, so `streamTranscriptionFrom` against this Layer alone
- * is a compile-time error.
- */
-declare const layer: (cfg: Config) => Layer.Layer<unknown, never, HttpClient.HttpClient>;
-//#endregion
-export { Config, GeminiTranscribeRequest, GeminiTranscriber, GeminiTranscriberService, buildPrompt, layer, make, GeminiTranscriber_d_exports as t };
-//# sourceMappingURL=GeminiTranscriber.d.mts.map

package/dist/GeminiTranscriber.d.mts.map

@@ -1 +0,0 @@
-{"version":3,"file":"GeminiTranscriber.d.mts","names":[],"sources":["../src/GeminiTranscriber.ts"],"mappings":";;;;;;;;;;;;;;;;;KAsBY,uBAAA,GAA0B,IAAA,CAAK,uBAAA;EAAA,SAChC,KAAA,EAAO,cAAA;AAAA;AAAA,KAGN,wBAAA;EAAA,SACD,UAAA,GACP,CAAA,EAAG,uBAAA,KACA,MAAA,CAAO,MAAA,CAAO,gBAAA,EAAkB,OAAA,CAAQ,OAAA;EAAA,SACpC,uBAAA,EAAyB,kBAAA;AAAA;AAAA,cACnC,sBAAA;cAEY,iBAAA,SAA0B,sBAAA;AAAA,KAK3B,MAAA;EAAA,SAAoB,MAAA,EAAQ,QAAA,CAAS,QAAA;EAAA,SAAmB,OAAA;AAAA;AAAA,cAqCvD,WAAA,GAAe,CAAA,EAAG,uBAAA;AAAA,cAkGlB,IAAA,GAAQ,GAAA,EAAK,MAAA,KAAM,MAAA,CAAA,MAAA,CAAA,wBAAA,SAAA,UAAA,CAAA,UAAA;;;;;AAnJhC;;cAmKa,KAAA,GAAS,GAAA,EAAK,MAAA,KAAM,KAAA,CAAA,KAAA,iBAAA,UAAA,CAAA,UAAA"}

package/dist/GeminiTranscriber.mjs

@@ -1,80 +0,0 @@
-import { n as __exportAll } from "./chunk-uyGKjUfl.mjs";
-import { audioSourceToInlineData, httpStatusError, transportFailure } from "./geminiSpeechCodec.mjs";
-import { Context, Effect, Layer, Match, Redacted, Schema, Stream } from "effect";
-import { HttpClient, HttpClientRequest } from "effect/unstable/http";
-import * as AiError from "@effect-uai/core/AiError";
-import { Transcriber } from "@effect-uai/core/Transcriber";
-//#region src/GeminiTranscriber.ts
-var GeminiTranscriber_exports = /* @__PURE__ */ __exportAll({
-	GeminiTranscriber: () => GeminiTranscriber,
-	buildPrompt: () => buildPrompt,
-	layer: () => layer,
-	make: () => make
-});
-var GeminiTranscriber = class extends Context.Service()("@betalyra/effect-uai/providers/google/GeminiTranscriber") {};
-const unsupported = (capability, reason) => new AiError.Unsupported({
-	provider: "gemini",
-	capability,
-	reason
-});
-const ensureSupported = Match.type().pipe(Match.when({ wordTimestamps: true }, () => Effect.fail(unsupported("wordTimestamps", "Gemini's prompt-driven transcription returns plain text — no per-word timing. Use Cloud Speech-to-Text (Chirp 2) for word timestamps."))), Match.when({ diarization: true }, () => Effect.fail(unsupported("diarization", "Gemini's prompt-driven transcription does not surface speaker IDs. Use Cloud Speech-to-Text (Chirp 2) for diarization."))), Match.orElse(() => Effect.void));
-const biasingHint = (r) => {
-	if (r.prompt === void 0) return "";
-	if (typeof r.prompt === "string") return `\nContext: ${r.prompt}`;
-	return r.prompt.terms.length === 0 ? "" : `\nThe following terms / names appear in the audio: ${r.prompt.terms.join(", ")}.`;
-};
-const buildPrompt = (r) => (r.language !== void 0 ? `Transcribe the audio verbatim in ${r.language}. Return only the transcript text, no preamble.` : "Transcribe the audio verbatim. Return only the transcript text, no preamble.") + biasingHint(r);
-const Wire = Schema.Struct({ candidates: Schema.optional(Schema.Array(Schema.Struct({ content: Schema.optional(Schema.Struct({ parts: Schema.optional(Schema.Array(Schema.Struct({ text: Schema.optional(Schema.String) }))) })) }))) });
-const decodeWire = Schema.decodeUnknownEffect(Wire);
-const collectText = (wire) => (wire.candidates ?? []).flatMap((c) => c.content?.parts ?? []).flatMap((p) => p.text !== void 0 ? [p.text] : []).join("\n").trim();
-const baseUrl = (cfg) => cfg.baseUrl ?? "https://generativelanguage.googleapis.com/v1beta";
-const transcribeImpl = (cfg) => (request) => Effect.gen(function* () {
-	yield* ensureSupported(request);
-	const client = yield* HttpClient.HttpClient;
-	const inline = yield* audioSourceToInlineData(request.audio);
-	const httpRequest = HttpClientRequest.post(`${baseUrl(cfg)}/models/${request.model}:generateContent`).pipe(HttpClientRequest.setHeader("x-goog-api-key", Redacted.value(cfg.apiKey)), HttpClientRequest.bodyJsonUnsafe({ contents: [{ parts: [{ text: buildPrompt(request) }, { inlineData: inline }] }] }));
-	const response = yield* client.execute(httpRequest).pipe(Effect.mapError(transportFailure));
-	if (response.status >= 400) {
-		const text = yield* response.text.pipe(Effect.orElseSucceed(() => ""));
-		return yield* httpStatusError(response.status, text);
-	}
-	const json = yield* response.json.pipe(Effect.mapError(transportFailure));
-	return {
-		text: yield* decodeWire(json).pipe(Effect.map(collectText), Effect.mapError((cause) => new AiError.GenerationFailed({
-			provider: "gemini",
-			raw: cause
-		})), Effect.flatMap((t) => t.length === 0 ? Effect.fail(new AiError.GenerationFailed({
-			provider: "gemini",
-			raw: {
-				message: "Gemini transcription response had no text part.",
-				hint: "Likely a prompt-safety rejection or the model returned only metadata.",
-				response: json
-			}
-		})) : Effect.succeed(t))),
-		...request.language !== void 0 && { languageCode: request.language },
-		raw: json
-	};
-});
-const unsupportedStreamFrom = () => Stream.fail(new AiError.Unsupported({
-	provider: "gemini",
-	capability: "streamTranscriptionFrom",
-	reason: "Gemini transcription via generateContent is sync-only. Use Cloud Speech-to-Text (StreamingRecognize) for live transcription."
-}));
-const make = (cfg) => Effect.map(HttpClient.HttpClient.asEffect(), (client) => ({
-	transcribe: (r) => transcribeImpl(cfg)(r).pipe(Effect.provideService(HttpClient.HttpClient, client)),
-	streamTranscriptionFrom: unsupportedStreamFrom
-}));
-/**
-* Layer registers both `GeminiTranscriber` and the generic
-* `Transcriber`. Does NOT register `SttStreaming` — Gemini's transcription
-* is sync-only, so `streamTranscriptionFrom` against this Layer alone
-* is a compile-time error.
-*/
-const layer = (cfg) => Layer.merge(Layer.effect(GeminiTranscriber, make(cfg)), Layer.effect(Transcriber, Effect.map(make(cfg), (s) => ({
-	transcribe: (req) => s.transcribe(req),
-	streamTranscriptionFrom: s.streamTranscriptionFrom
-}))));
-//#endregion
-export { GeminiTranscriber, buildPrompt, layer, make, GeminiTranscriber_exports as t };
-
-//# sourceMappingURL=GeminiTranscriber.mjs.map

package/dist/GeminiTranscriber.mjs.map

@@ -1 +0,0 @@
-{"version":3,"file":"GeminiTranscriber.mjs","names":[],"sources":["../src/GeminiTranscriber.ts"],"sourcesContent":["import { Context, Effect, Layer, Match, Redacted, Schema, Stream } from \"effect\"\nimport { HttpClient, HttpClientRequest } from \"effect/unstable/http\"\nimport * as AiError from \"@effect-uai/core/AiError\"\nimport type { TranscriptResult } from \"@effect-uai/core/Transcript\"\nimport {\n  type CommonTranscribeRequest,\n  Transcriber,\n  type TranscriberService,\n} from \"@effect-uai/core/Transcriber\"\nimport { audioSourceToInlineData, httpStatusError, transportFailure } from \"./geminiSpeechCodec.js\"\nimport type { GeminiSttModel } from \"./models.js\"\n\n// ---------------------------------------------------------------------------\n// Public types\n// ---------------------------------------------------------------------------\n\n/**\n * Gemini transcription rides on `:generateContent` — the model receives\n * audio plus a textual \"transcribe verbatim\" prompt and returns plain\n * text. There is no native structured timestamp / diarization channel,\n * so `wordTimestamps` / `diarization` short-circuit with `Unsupported`.\n */\nexport type GeminiTranscribeRequest = Omit<CommonTranscribeRequest, \"model\"> & {\n  readonly model: GeminiSttModel\n}\n\nexport type GeminiTranscriberService = {\n  readonly transcribe: (\n    r: GeminiTranscribeRequest,\n  ) => Effect.Effect<TranscriptResult, AiError.AiError>\n  readonly streamTranscriptionFrom: TranscriberService[\"streamTranscriptionFrom\"]\n}\n\nexport class GeminiTranscriber extends Context.Service<\n  GeminiTranscriber,\n  GeminiTranscriberService\n>()(\"@betalyra/effect-uai/providers/google/GeminiTranscriber\") {}\n\nexport type Config = { readonly apiKey: Redacted.Redacted; readonly baseUrl?: string }\n\n// ---------------------------------------------------------------------------\n// Prompt builder + capability guards\n// ---------------------------------------------------------------------------\n\nconst unsupported = (capability: string, reason: string) =>\n  new AiError.Unsupported({ provider: \"gemini\", capability, reason })\n\nconst ensureSupported = Match.type<GeminiTranscribeRequest>().pipe(\n  Match.when({ wordTimestamps: true }, () =>\n    Effect.fail(\n      unsupported(\n        \"wordTimestamps\",\n        \"Gemini's prompt-driven transcription returns plain text — no per-word timing. Use Cloud Speech-to-Text (Chirp 2) for word timestamps.\",\n      ),\n    ),\n  ),\n  Match.when({ diarization: true }, () =>\n    Effect.fail(\n      unsupported(\n        \"diarization\",\n        \"Gemini's prompt-driven transcription does not surface speaker IDs. Use Cloud Speech-to-Text (Chirp 2) for diarization.\",\n      ),\n    ),\n  ),\n  Match.orElse(() => Effect.void),\n)\n\nconst biasingHint = (r: GeminiTranscribeRequest) => {\n  if (r.prompt === undefined) return \"\"\n  if (typeof r.prompt === \"string\") return `\\nContext: ${r.prompt}`\n  return r.prompt.terms.length === 0\n    ? \"\"\n    : `\\nThe following terms / names appear in the audio: ${r.prompt.terms.join(\", \")}.`\n}\n\nexport const buildPrompt = (r: GeminiTranscribeRequest) =>\n  (r.language !== undefined\n    ? `Transcribe the audio verbatim in ${r.language}. Return only the transcript text, no preamble.`\n    : \"Transcribe the audio verbatim. Return only the transcript text, no preamble.\") +\n  biasingHint(r)\n\n// ---------------------------------------------------------------------------\n// Wire\n// ---------------------------------------------------------------------------\n\nconst Wire = Schema.Struct({\n  candidates: Schema.optional(\n    Schema.Array(\n      Schema.Struct({\n        content: Schema.optional(\n          Schema.Struct({\n            parts: Schema.optional(\n              Schema.Array(Schema.Struct({ text: Schema.optional(Schema.String) })),\n            ),\n          }),\n        ),\n      }),\n    ),\n  ),\n})\nconst decodeWire = Schema.decodeUnknownEffect(Wire)\n\nconst collectText = (wire: typeof Wire.Type) =>\n  (wire.candidates ?? [])\n    .flatMap((c) => c.content?.parts ?? [])\n    .flatMap((p) => (p.text !== undefined ? [p.text] : []))\n    .join(\"\\n\")\n    .trim()\n\n// ---------------------------------------------------------------------------\n// Service implementation\n// ---------------------------------------------------------------------------\n\nconst baseUrl = (cfg: Config) => cfg.baseUrl ?? \"https://generativelanguage.googleapis.com/v1beta\"\n\nconst transcribeImpl = (cfg: Config) => (request: GeminiTranscribeRequest) =>\n  Effect.gen(function* () {\n    yield* ensureSupported(request)\n    const client = yield* HttpClient.HttpClient\n    const inline = yield* audioSourceToInlineData(request.audio)\n    const httpRequest = HttpClientRequest.post(\n      `${baseUrl(cfg)}/models/${request.model}:generateContent`,\n    ).pipe(\n      HttpClientRequest.setHeader(\"x-goog-api-key\", Redacted.value(cfg.apiKey)),\n      HttpClientRequest.bodyJsonUnsafe({\n        contents: [{ parts: [{ text: buildPrompt(request) }, { inlineData: inline }] }],\n      }),\n    )\n    const response = yield* client.execute(httpRequest).pipe(Effect.mapError(transportFailure))\n    if (response.status >= 400) {\n      const text = yield* response.text.pipe(Effect.orElseSucceed(() => \"\"))\n      return yield* httpStatusError(response.status, text)\n    }\n    const json = yield* response.json.pipe(Effect.mapError(transportFailure))\n    const text = yield* decodeWire(json).pipe(\n      Effect.map(collectText),\n      Effect.mapError((cause) => new AiError.GenerationFailed({ provider: \"gemini\", raw: cause })),\n      Effect.flatMap((t) =>\n        t.length === 0\n          ? Effect.fail(\n              new AiError.GenerationFailed({\n                provider: \"gemini\",\n                raw: {\n                  message: \"Gemini transcription response had no text part.\",\n                  hint: \"Likely a prompt-safety rejection or the model returned only metadata.\",\n                  response: json,\n                },\n              }),\n            )\n          : Effect.succeed(t),\n      ),\n    )\n    return {\n      text,\n      ...(request.language !== undefined && { languageCode: request.language }),\n      raw: json,\n    } satisfies TranscriptResult\n  })\n\nconst unsupportedStreamFrom: TranscriberService[\"streamTranscriptionFrom\"] = () =>\n  Stream.fail(\n    new AiError.Unsupported({\n      provider: \"gemini\",\n      capability: \"streamTranscriptionFrom\",\n      reason:\n        \"Gemini transcription via generateContent is sync-only. Use Cloud Speech-to-Text (StreamingRecognize) for live transcription.\",\n    }),\n  )\n\n// ---------------------------------------------------------------------------\n// Constructors\n// ---------------------------------------------------------------------------\n\nexport const make = (cfg: Config) =>\n  Effect.map(\n    HttpClient.HttpClient.asEffect(),\n    (client): GeminiTranscriberService => ({\n      transcribe: (r) =>\n        transcribeImpl(cfg)(r).pipe(Effect.provideService(HttpClient.HttpClient, client)),\n      streamTranscriptionFrom: unsupportedStreamFrom,\n    }),\n  )\n\n/**\n * Layer registers both `GeminiTranscriber` and the generic\n * `Transcriber`. Does NOT register `SttStreaming` — Gemini's transcription\n * is sync-only, so `streamTranscriptionFrom` against this Layer alone\n * is a compile-time error.\n */\nexport const layer = (cfg: Config) =>\n  Layer.merge(\n    Layer.effect(GeminiTranscriber, make(cfg)),\n    Layer.effect(\n      Transcriber,\n      Effect.map(\n        make(cfg),\n        (s): TranscriberService => ({\n          transcribe: (req: CommonTranscribeRequest) =>\n            s.transcribe(req as GeminiTranscribeRequest),\n          streamTranscriptionFrom: s.streamTranscriptionFrom,\n        }),\n      ),\n    ),\n  )\n"],"mappings":";;;;;;;;;;;;;AAiCA,IAAa,oBAAb,cAAuC,QAAQ,SAG5C,CAAC,0DAA0D,CAAC;AAQ/D,MAAM,eAAe,YAAoB,WACvC,IAAI,QAAQ,YAAY;CAAE,UAAU;CAAU;CAAY;CAAQ,CAAC;AAErE,MAAM,kBAAkB,MAAM,MAA+B,CAAC,KAC5D,MAAM,KAAK,EAAE,gBAAgB,MAAM,QACjC,OAAO,KACL,YACE,kBACA,wIACD,CACF,CACF,EACD,MAAM,KAAK,EAAE,aAAa,MAAM,QAC9B,OAAO,KACL,YACE,eACA,yHACD,CACF,CACF,EACD,MAAM,aAAa,OAAO,KAAK,CAChC;AAED,MAAM,eAAe,MAA+B;AAClD,KAAI,EAAE,WAAW,KAAA,EAAW,QAAO;AACnC,KAAI,OAAO,EAAE,WAAW,SAAU,QAAO,cAAc,EAAE;AACzD,QAAO,EAAE,OAAO,MAAM,WAAW,IAC7B,KACA,sDAAsD,EAAE,OAAO,MAAM,KAAK,KAAK,CAAC;;AAGtF,MAAa,eAAe,OACzB,EAAE,aAAa,KAAA,IACZ,oCAAoC,EAAE,SAAS,mDAC/C,kFACJ,YAAY,EAAE;AAMhB,MAAM,OAAO,OAAO,OAAO,EACzB,YAAY,OAAO,SACjB,OAAO,MACL,OAAO,OAAO,EACZ,SAAS,OAAO,SACd,OAAO,OAAO,EACZ,OAAO,OAAO,SACZ,OAAO,MAAM,OAAO,OAAO,EAAE,MAAM,OAAO,SAAS,OAAO,OAAO,EAAE,CAAC,CAAC,CACtE,EACF,CAAC,CACH,EACF,CAAC,CACH,CACF,EACF,CAAC;AACF,MAAM,aAAa,OAAO,oBAAoB,KAAK;AAEnD,MAAM,eAAe,UAClB,KAAK,cAAc,EAAE,EACnB,SAAS,MAAM,EAAE,SAAS,SAAS,EAAE,CAAC,CACtC,SAAS,MAAO,EAAE,SAAS,KAAA,IAAY,CAAC,EAAE,KAAK,GAAG,EAAE,CAAE,CACtD,KAAK,KAAK,CACV,MAAM;AAMX,MAAM,WAAW,QAAgB,IAAI,WAAW;AAEhD,MAAM,kBAAkB,SAAiB,YACvC,OAAO,IAAI,aAAa;AACtB,QAAO,gBAAgB,QAAQ;CAC/B,MAAM,SAAS,OAAO,WAAW;CACjC,MAAM,SAAS,OAAO,wBAAwB,QAAQ,MAAM;CAC5D,MAAM,cAAc,kBAAkB,KACpC,GAAG,QAAQ,IAAI,CAAC,UAAU,QAAQ,MAAM,kBACzC,CAAC,KACA,kBAAkB,UAAU,kBAAkB,SAAS,MAAM,IAAI,OAAO,CAAC,EACzE,kBAAkB,eAAe,EAC/B,UAAU,CAAC,EAAE,OAAO,CAAC,EAAE,MAAM,YAAY,QAAQ,EAAE,EAAE,EAAE,YAAY,QAAQ,CAAC,EAAE,CAAC,EAChF,CAAC,CACH;CACD,MAAM,WAAW,OAAO,OAAO,QAAQ,YAAY,CAAC,KAAK,OAAO,SAAS,iBAAiB,CAAC;AAC3F,KAAI,SAAS,UAAU,KAAK;EAC1B,MAAM,OAAO,OAAO,SAAS,KAAK,KAAK,OAAO,oBAAoB,GAAG,CAAC;AACtE,SAAO,OAAO,gBAAgB,SAAS,QAAQ,KAAK;;CAEtD,MAAM,OAAO,OAAO,SAAS,KAAK,KAAK,OAAO,SAAS,iBAAiB,CAAC;AAmBzE,QAAO;EACL,MAAA,OAnBkB,WAAW,KAAK,CAAC,KACnC,OAAO,IAAI,YAAY,EACvB,OAAO,UAAU,UAAU,IAAI,QAAQ,iBAAiB;GAAE,UAAU;GAAU,KAAK;GAAO,CAAC,CAAC,EAC5F,OAAO,SAAS,MACd,EAAE,WAAW,IACT,OAAO,KACL,IAAI,QAAQ,iBAAiB;GAC3B,UAAU;GACV,KAAK;IACH,SAAS;IACT,MAAM;IACN,UAAU;IACX;GACF,CAAC,CACH,GACD,OAAO,QAAQ,EAAE,CACtB,CACF;EAGC,GAAI,QAAQ,aAAa,KAAA,KAAa,EAAE,cAAc,QAAQ,UAAU;EACxE,KAAK;EACN;EACD;AAEJ,MAAM,8BACJ,OAAO,KACL,IAAI,QAAQ,YAAY;CACtB,UAAU;CACV,YAAY;CACZ,QACE;CACH,CAAC,CACH;AAMH,MAAa,QAAQ,QACnB,OAAO,IACL,WAAW,WAAW,UAAU,GAC/B,YAAsC;CACrC,aAAa,MACX,eAAe,IAAI,CAAC,EAAE,CAAC,KAAK,OAAO,eAAe,WAAW,YAAY,OAAO,CAAC;CACnF,yBAAyB;CAC1B,EACF;;;;;;;AAQH,MAAa,SAAS,QACpB,MAAM,MACJ,MAAM,OAAO,mBAAmB,KAAK,IAAI,CAAC,EAC1C,MAAM,OACJ,aACA,OAAO,IACL,KAAK,IAAI,GACR,OAA2B;CAC1B,aAAa,QACX,EAAE,WAAW,IAA+B;CAC9C,yBAAyB,EAAE;CAC5B,EACF,CACF,CACF"}

package/dist/GeminiTranscriber.test.d.mts

@@ -1 +0,0 @@
-export { };