@startsimpli/llm 0.4.24 → 0.4.26

package/src/providers/openai.ts

@@ -19,6 +19,7 @@ import type {
   NativeToolResponse,
   ReasoningEffort,
   LLMImage,
+  LLMDocument,
   TokenUsage,
 } from '../types';
 import { LLM_DEFAULTS } from '../types';
@@ -27,19 +28,28 @@ import type { LLMErrorCode } from '../errors';
 
 type OpenAIContentPart =
   | { type: 'text'; text: string }
-  | { type: 'image_url'; image_url: { url: string } };
+  | { type: 'image_url'; image_url: { url: string } }
+  | { type: 'file'; file: { filename: string; file_data: string } };
 
 interface OpenAIMessage {
   role: 'system' | 'user' | 'assistant';
   content: string | OpenAIContentPart[];
 }
 
-/** The user message: plain text, or text plus images for vision models. */
-function userContent(prompt: string, images?: LLMImage[]): string | OpenAIContentPart[] {
-  if (!images?.length) return prompt;
+/**
+ * The user message: plain text, or text followed by documents (file parts) and
+ * images. OpenAI requires a filename on a file part, so an unnamed document
+ * gets a default one.
+ */
+function userContent(prompt: string, images?: LLMImage[], documents?: LLMDocument[]): string | OpenAIContentPart[] {
+  if (!images?.length && !documents?.length) return prompt;
   return [
     { type: 'text', text: prompt },
-    ...images.map((img) => ({ type: 'image_url' as const, image_url: { url: `data:${img.mediaType};base64,${img.data}` } })),
+    ...(documents ?? []).map((doc, i) => ({
+      type: 'file' as const,
+      file: { filename: doc.filename || `document-${i + 1}.pdf`, file_data: `data:${doc.mediaType};base64,${doc.data}` },
+    })),
+    ...(images ?? []).map((img) => ({ type: 'image_url' as const, image_url: { url: `data:${img.mediaType};base64,${img.data}` } })),
   ];
 }
 
@@ -364,7 +374,7 @@ export class OpenAIProvider implements ILLMProvider {
       model,
       messages: [
         { role: 'system' as const, content: systemPrompt },
-        { role: 'user' as const, content: userContent(userPrompt, options?.images) },
+        { role: 'user' as const, content: userContent(userPrompt, options?.images, options?.documents) },
       ],
       response_format: { type: 'json_object' as const },
       stream: true,
@@ -466,7 +476,7 @@ export class OpenAIProvider implements ILLMProvider {
       model,
       messages: [
         { role: 'system' as const, content: systemPrompt },
-        { role: 'user' as const, content: userContent(userPrompt, options?.images) },
+        { role: 'user' as const, content: userContent(userPrompt, options?.images, options?.documents) },
       ],
       response_format: { type: 'json_object' as const },
     };

package/src/providers/registry.ts

@@ -0,0 +1,59 @@
+/**
+ * One provider registry for every modality (text, speech, and whatever comes next).
+ *
+ * WHY THIS EXISTS. `LLMService` used to hard-code its providers in a private
+ * method, so adding one meant editing the service. Speech needed the same shape
+ * and would have copied that method. Now a modality owns a registry of
+ * name -> factory, built-ins register themselves, and a new provider — in this
+ * package or in an app — is a single `register(...)` call with no service edit.
+ *
+ * A factory returns `null` when its credentials are absent, which is how
+ * `createOpenAIProvider()` already behaved; the registry just formalises it.
+ */
+export interface Selectable {
+  readonly name: string;
+  isAvailable(): boolean;
+}
+
+export type ProviderFactory<P extends Selectable> = () => P | null;
+
+export class ProviderRegistry<P extends Selectable> {
+  private readonly factories = new Map<string, ProviderFactory<P>>();
+
+  /** Register (or replace) a provider by name. Later registrations win. */
+  register(name: string, factory: ProviderFactory<P>): this {
+    this.factories.set(name, factory);
+    return this;
+  }
+
+  unregister(name: string): boolean {
+    return this.factories.delete(name);
+  }
+
+  names(): string[] {
+    return [...this.factories.keys()];
+  }
+
+  /** Build every registered provider that its own environment allows. */
+  instantiate(): Map<string, P> {
+    const out = new Map<string, P>();
+    for (const [name, factory] of this.factories) {
+      const provider = factory();
+      if (provider) out.set(name, provider);
+    }
+    return out;
+  }
+}
+
+/** The preferred provider if it is usable, otherwise the first one that is. */
+export function pickProvider<P extends Selectable>(
+  providers: Map<string, P>,
+  preferred?: string,
+): P | undefined {
+  const first = preferred ? providers.get(preferred) : undefined;
+  if (first?.isAvailable()) return first;
+  for (const provider of providers.values()) {
+    if (provider.isAvailable()) return provider;
+  }
+  return undefined;
+}

package/src/service.ts

@@ -26,6 +26,20 @@ import {
 import { OpenAIProvider, createOpenAIProvider } from './providers/openai';
 import { AnthropicProvider, createAnthropicProvider } from './providers/anthropic';
 import { MockLLMProvider } from './providers/mock';
+import { ProviderRegistry, pickProvider } from './providers/registry';
+
+/**
+ * The text-provider registry. Built-ins register below; anything else —
+ * a self-hosted model, a new vendor — is `registerLLMProvider(name, factory)`
+ * and needs no change to `LLMService`.
+ */
+export const llmProviders = new ProviderRegistry<ILLMProvider>()
+  .register('openai', () => createOpenAIProvider())
+  .register('anthropic', () => createAnthropicProvider());
+
+export function registerLLMProvider(name: string, factory: () => ILLMProvider | null): void {
+  llmProviders.register(name, factory);
+}
 
 /**
  * Metadata about a generation attempt (tokens, model, retries).
@@ -83,22 +97,14 @@ export class LLMService<TOutput> {
       return;
     }
 
-    const openai = createOpenAIProvider();
-    if (openai) this.providers.set('openai', openai);
-
-    const anthropic = createAnthropicProvider();
-    if (anthropic) this.providers.set('anthropic', anthropic);
+    for (const [name, provider] of llmProviders.instantiate()) {
+      this.providers.set(name, provider);
+    }
   }
 
   private getProvider(): ILLMProvider {
-    // Try preferred first
-    const preferred = this.providers.get(this.config.preferredProvider);
-    if (preferred?.isAvailable()) return preferred;
-
-    // Fallback to any available
-    for (const provider of this.providers.values()) {
-      if (provider.isAvailable()) return provider;
-    }
+    const picked = pickProvider(this.providers, this.config.preferredProvider);
+    if (picked) return picked;
 
     throw new LLMProviderError(
       'No LLM providers available. Please configure OPENAI_API_KEY or ANTHROPIC_API_KEY.',

package/src/speech/http.ts

@@ -0,0 +1,73 @@
+import { LLMProviderError } from '../errors';
+import type { LLMErrorCode } from '../errors';
+import type { TranscribeRequest } from './types';
+
+/** Container extension a vendor can recognise, from the recorder's MIME type. */
+const EXTENSIONS: Record<string, string> = {
+  'audio/webm': 'webm', 'video/webm': 'webm', 'audio/ogg': 'ogg', 'audio/mp4': 'mp4',
+  'audio/x-m4a': 'm4a', 'audio/m4a': 'm4a', 'audio/mpeg': 'mp3', 'audio/mp3': 'mp3',
+  'audio/wav': 'wav', 'audio/x-wav': 'wav', 'audio/flac': 'flac',
+};
+
+/**
+ * THE FILE NAME MATTERS. Vendors infer the container from it, and browsers
+ * disagree: Chrome/Firefox record webm/ogg, Safari records mp4. A webm blob
+ * uploaded as `.mp4` is rejected as unsupported, so the name comes from the
+ * MIME type, never a constant.
+ */
+export function audioExtension(mimeType: string): string {
+  const base = (mimeType || '').split(';')[0].trim().toLowerCase();
+  return EXTENSIONS[base] ?? 'webm';
+}
+
+export function audioBlob(request: Pick<TranscribeRequest, 'audio' | 'mimeType'>): { blob: Blob; filename: string } {
+  const base = (request.mimeType || 'audio/webm').split(';')[0].trim() || 'audio/webm';
+  return {
+    blob: new Blob([new Uint8Array(request.audio)], { type: base }),
+    filename: `audio.${audioExtension(request.mimeType)}`,
+  };
+}
+
+export function assertAudio(request: Pick<TranscribeRequest, 'audio'>, provider: string): void {
+  if (!request.audio?.byteLength) {
+    throw new LLMProviderError('There is no audio to transcribe.', 'INVALID_REQUEST', false, undefined, provider);
+  }
+}
+
+/** One HTTP-status -> error-code mapping, so every vendor classifies failures the same way. */
+export function speechHttpError(status: number, body: string, provider: string): LLMProviderError {
+  const code: LLMErrorCode =
+    status === 401 || status === 403 ? 'AUTHENTICATION_ERROR'
+    : status === 429 ? 'RATE_LIMITED'
+    : status === 408 || status === 504 ? 'TIMEOUT'
+    : status >= 500 ? 'SERVICE_UNAVAILABLE'
+    : status === 400 || status === 413 || status === 415 || status === 422 ? 'INVALID_REQUEST'
+    : 'UNKNOWN_ERROR';
+  const retryable = code === 'RATE_LIMITED' || code === 'SERVICE_UNAVAILABLE' || code === 'TIMEOUT';
+  return new LLMProviderError(
+    `${provider} transcription failed (${status}): ${body.slice(0, 300)}`,
+    code, retryable, status, provider,
+  );
+}
+
+/** fetch with a deadline; a timeout or network drop becomes a typed, retryable error. */
+export async function speechFetch(
+  provider: string,
+  url: string,
+  init: RequestInit,
+  timeoutMs = 240_000,
+  fetchImpl: typeof fetch = fetch,
+): Promise<Response> {
+  try {
+    const response = await fetchImpl(url, { ...init, signal: AbortSignal.timeout(timeoutMs) });
+    if (!response.ok) throw speechHttpError(response.status, await response.text().catch(() => ''), provider);
+    return response;
+  } catch (error) {
+    if (error instanceof LLMProviderError) throw error;
+    const timedOut = error instanceof Error && (error.name === 'TimeoutError' || error.name === 'AbortError');
+    throw new LLMProviderError(
+      timedOut ? `${provider} did not answer within ${Math.round(timeoutMs / 1000)}s.` : `${provider} request failed: ${(error as Error)?.message ?? error}`,
+      timedOut ? 'TIMEOUT' : 'SERVICE_UNAVAILABLE', true, undefined, provider,
+    );
+  }
+}

package/src/speech/index.ts

@@ -0,0 +1,17 @@
+/**
+ * `@startsimpli/llm/speech` — pluggable speech-to-text.
+ *
+ *   const speech = new SpeechService();            // SPEECH_PROVIDER or ElevenLabs, fallback to OpenAI
+ *   const { text } = await speech.transcribe({ audio, mimeType: 'audio/webm' });
+ *
+ * A new vendor: implement `ISpeechProvider`, then `registerSpeechProvider(name, factory)`.
+ */
+export { SpeechService, registerSpeechProvider, speechProviders } from './service';
+export type { ISpeechProvider } from './interface';
+export type {
+  SpeechServiceConfig, SpeechUsage, TranscribeRequest, TranscriptResult, TranscriptWord,
+} from './types';
+export { ElevenLabsSpeechProvider, createElevenLabsSpeechProvider } from './providers/elevenlabs';
+export { OpenAISpeechProvider, createOpenAISpeechProvider } from './providers/openai';
+export { MockSpeechProvider, setMockSpeechBehavior, resetMockSpeechBehavior } from './providers/mock';
+export { audioExtension, speechHttpError } from './http';

package/src/speech/interface.ts

@@ -0,0 +1,19 @@
+import type { Selectable } from '../providers/registry';
+import type { TranscribeRequest, TranscriptResult } from './types';
+
+/**
+ * What a speech vendor implements. A provider supports whichever directions its
+ * API does — `capabilities` says which, so a caller (or the service) never has
+ * to call a method and catch "not supported".
+ *
+ * ADDING A VENDOR: implement this, then
+ * `registerSpeechProvider('name', () => new Mine(process.env.MINE_KEY) )` — no
+ * edit to `SpeechService`. Text-to-speech joins here as an optional
+ * `synthesize` alongside `transcribe` (bd startsim-2lnyg); nothing in the
+ * service needs to change for it either.
+ */
+export interface ISpeechProvider extends Selectable {
+  readonly capabilities: { transcribe: boolean };
+  readonly defaultModel: string;
+  transcribe?(request: TranscribeRequest): Promise<TranscriptResult>;
+}

package/src/speech/providers/elevenlabs.ts

@@ -0,0 +1,72 @@
+/**
+ * ElevenLabs Scribe speech-to-text. https://elevenlabs.io/docs/api-reference/speech-to-text
+ * Env: ELEVENLABS_API_KEY.
+ */
+import type { ISpeechProvider } from '../interface';
+import { assertAudio, audioBlob, speechFetch } from '../http';
+import type { TranscribeRequest, TranscriptResult, TranscriptWord } from '../types';
+
+export interface ElevenLabsOptions {
+  apiKey?: string;
+  baseUrl?: string;
+  fetchImpl?: typeof fetch;
+}
+
+interface ScribeWord { text: string; start?: number; end?: number; type?: string; speaker_id?: string }
+
+export class ElevenLabsSpeechProvider implements ISpeechProvider {
+  readonly name = 'elevenlabs';
+  readonly capabilities = { transcribe: true };
+  readonly defaultModel = 'scribe_v1';
+  private readonly apiKey: string;
+  private readonly baseUrl: string;
+  private readonly fetchImpl?: typeof fetch;
+
+  constructor(options: ElevenLabsOptions = {}) {
+    this.apiKey = options.apiKey ?? '';
+    this.baseUrl = (options.baseUrl ?? 'https://api.elevenlabs.io').replace(/\/+$/, '');
+    this.fetchImpl = options.fetchImpl;
+  }
+
+  isAvailable(): boolean {
+    return Boolean(this.apiKey);
+  }
+
+  async transcribe(request: TranscribeRequest): Promise<TranscriptResult> {
+    assertAudio(request, this.name);
+    const model = request.model ?? this.defaultModel;
+    const { blob, filename } = audioBlob(request);
+    const form = new FormData();
+    form.append('file', blob, filename);
+    form.append('model_id', model);
+    form.append('timestamps_granularity', request.timestamps === 'none' ? 'none' : 'word');
+    if (request.language) form.append('language_code', request.language);
+    if (request.diarize) form.append('diarize', 'true');
+    // Scribe has no free-text prompt; `keyterms` is its bias list on models that take it.
+    for (const term of request.hints ?? []) form.append('keyterms', term);
+
+    const response = await speechFetch(
+      this.name, `${this.baseUrl}/v1/speech-to-text`,
+      { method: 'POST', headers: { 'xi-api-key': this.apiKey }, body: form },
+      request.timeoutMs, this.fetchImpl,
+    );
+    const json = (await response.json()) as { text?: string; language_code?: string; words?: ScribeWord[] };
+    const words: TranscriptWord[] | undefined = json.words
+      ?.filter((w) => (w.type ?? 'word') === 'word' && typeof w.start === 'number' && typeof w.end === 'number')
+      .map((w) => ({ text: w.text, start: w.start as number, end: w.end as number, ...(w.speaker_id ? { speaker: w.speaker_id } : {}) }));
+    const audioSeconds = words?.length ? words[words.length - 1].end : undefined;
+    return {
+      text: (json.text ?? '').trim(),
+      ...(json.language_code ? { language: json.language_code } : {}),
+      ...(words?.length ? { words } : {}),
+      model, provider: this.name,
+      usage: audioSeconds !== undefined ? { audioSeconds } : {},
+    };
+  }
+}
+
+export function createElevenLabsSpeechProvider(
+  env: Record<string, string | undefined> = typeof process !== 'undefined' ? process.env : {},
+): ElevenLabsSpeechProvider | null {
+  return env.ELEVENLABS_API_KEY ? new ElevenLabsSpeechProvider({ apiKey: env.ELEVENLABS_API_KEY }) : null;
+}

package/src/speech/providers/mock.ts

@@ -0,0 +1,41 @@
+import { LLMProviderError } from '../../errors';
+import type { ISpeechProvider } from '../interface';
+import type { TranscribeRequest, TranscriptResult } from '../types';
+
+export interface MockSpeechBehavior {
+  text?: string;
+  /** Throw this many retryable failures before succeeding. */
+  failTimes?: number;
+  /** Throw a non-retryable auth error every time. */
+  authError?: boolean;
+}
+
+let behavior: MockSpeechBehavior = {};
+let remainingFailures = 0;
+export function setMockSpeechBehavior(next: MockSpeechBehavior) {
+  behavior = next;
+  remainingFailures = next.failTimes ?? 0;
+}
+export function resetMockSpeechBehavior() {
+  behavior = {};
+  remainingFailures = 0;
+}
+
+/** Deterministic transcription for tests — selected with `SPEECH_PROVIDER=mock`. */
+export class MockSpeechProvider implements ISpeechProvider {
+  readonly name = 'mock';
+  readonly capabilities = { transcribe: true };
+  readonly defaultModel = 'mock-transcribe';
+  isAvailable() { return true; }
+  async transcribe(request: TranscribeRequest): Promise<TranscriptResult> {
+    if (behavior.authError) throw new LLMProviderError('mock auth failure', 'AUTHENTICATION_ERROR', false, 401, this.name);
+    if (remainingFailures > 0) {
+      remainingFailures -= 1;
+      throw new LLMProviderError('mock outage', 'SERVICE_UNAVAILABLE', true, 503, this.name);
+    }
+    return {
+      text: behavior.text ?? `mock transcript of ${request.audio.byteLength} bytes`,
+      model: this.defaultModel, provider: this.name, usage: { audioSeconds: 1 },
+    };
+  }
+}

package/src/speech/providers/openai.ts

@@ -0,0 +1,89 @@
+/**
+ * OpenAI speech-to-text. Env: OPENAI_API_KEY (the same key the text provider uses).
+ *
+ * `gpt-4o-transcribe` for plain text; per-word timings are only offered by
+ * `whisper-1` (verbose_json), so asking for `timestamps: 'word'` switches model.
+ */
+import type { ISpeechProvider } from '../interface';
+import { assertAudio, audioBlob, speechFetch } from '../http';
+import type { TranscribeRequest, TranscriptResult, TranscriptWord } from '../types';
+
+export interface OpenAISpeechOptions {
+  apiKey?: string;
+  baseUrl?: string;
+  fetchImpl?: typeof fetch;
+}
+
+export class OpenAISpeechProvider implements ISpeechProvider {
+  readonly name = 'openai';
+  readonly capabilities = { transcribe: true };
+  readonly defaultModel = 'gpt-4o-transcribe';
+  private readonly apiKey: string;
+  private readonly baseUrl: string;
+  private readonly fetchImpl?: typeof fetch;
+
+  constructor(options: OpenAISpeechOptions = {}) {
+    this.apiKey = options.apiKey ?? '';
+    this.baseUrl = (options.baseUrl ?? 'https://api.openai.com').replace(/\/+$/, '');
+    this.fetchImpl = options.fetchImpl;
+  }
+
+  isAvailable(): boolean {
+    return Boolean(this.apiKey);
+  }
+
+  async transcribe(request: TranscribeRequest): Promise<TranscriptResult> {
+    assertAudio(request, this.name);
+    const wantWords = request.timestamps === 'word';
+    const model = request.model ?? (wantWords ? 'whisper-1' : this.defaultModel);
+    const { blob, filename } = audioBlob(request);
+    const form = new FormData();
+    form.append('file', blob, filename);
+    form.append('model', model);
+    form.append('response_format', wantWords ? 'verbose_json' : 'json');
+    if (wantWords) form.append('timestamp_granularities[]', 'word');
+    if (request.language) form.append('language', request.language);
+    if (request.hints?.length) form.append('prompt', request.hints.join(', '));
+
+    const response = await speechFetch(
+      this.name, `${this.baseUrl}/v1/audio/transcriptions`,
+      { method: 'POST', headers: { authorization: `Bearer ${this.apiKey}` }, body: form },
+      request.timeoutMs, this.fetchImpl,
+    );
+    const json = (await response.json()) as {
+      text?: string; language?: string; duration?: number;
+      words?: Array<{ word: string; start: number; end: number }>;
+      usage?: { seconds?: number };
+    };
+    // On silence the model can hand the biasing prompt straight back as the
+    // "transcript". That is not speech, so report none rather than a confident lie.
+    const heard = (json.text ?? '').trim();
+    const echoed = isEcho(heard, request.hints);
+    const words: TranscriptWord[] | undefined = json.words?.map((w) => ({ text: w.word, start: w.start, end: w.end }));
+    const audioSeconds = json.usage?.seconds ?? json.duration;
+    return {
+      text: echoed ? '' : heard,
+      ...(json.language ? { language: json.language } : {}),
+      ...(words?.length ? { words } : {}),
+      model, provider: this.name,
+      usage: audioSeconds !== undefined ? { audioSeconds } : {},
+    };
+  }
+}
+
+export function createOpenAISpeechProvider(
+  env: Record<string, string | undefined> = typeof process !== 'undefined' ? process.env : {},
+): OpenAISpeechProvider | null {
+  return env.OPENAI_API_KEY ? new OpenAISpeechProvider({ apiKey: env.OPENAI_API_KEY }) : null;
+}
+
+const tokens = (text: string) => text.toLowerCase().split(/[^a-z0-9]+/).filter(Boolean);
+
+/** True when every word heard is one of the hint words — i.e. the prompt came back, not speech. */
+function isEcho(heard: string, hints?: string[]): boolean {
+  if (!hints?.length) return false;
+  const said = tokens(heard);
+  // Scaffold words a model attaches when it parrots the prompt ("context:", "projects:").
+  const allowed = new Set([...tokens(hints.join(' ')), 'context', 'prompt', 'vocabulary', 'glossary', 'terms', 'projects', 'names', 'keywords']);
+  return said.length > 0 && said.every((word) => allowed.has(word));
+}

package/src/speech/service.ts

@@ -0,0 +1,95 @@
+import { LLMProviderError } from '../errors';
+import { ProviderRegistry, pickProvider } from '../providers/registry';
+import type { ISpeechProvider } from './interface';
+import { createElevenLabsSpeechProvider } from './providers/elevenlabs';
+import { MockSpeechProvider } from './providers/mock';
+import { createOpenAISpeechProvider } from './providers/openai';
+import type { SpeechServiceConfig, TranscribeRequest, TranscriptResult } from './types';
+
+/**
+ * The speech registry. Built-ins register here; a new vendor is
+ * `registerSpeechProvider('name', factory)` — no edit to `SpeechService`.
+ */
+export const speechProviders = new ProviderRegistry<ISpeechProvider>()
+  .register('elevenlabs', () => createElevenLabsSpeechProvider())
+  .register('openai', () => createOpenAISpeechProvider());
+
+export function registerSpeechProvider(name: string, factory: () => ISpeechProvider | null): void {
+  speechProviders.register(name, factory);
+}
+
+const sleep = (ms: number) => new Promise<void>((resolve) => setTimeout(resolve, ms));
+
+/**
+ * Provider-agnostic speech. Pick a vendor with `preferredProvider` or the
+ * `SPEECH_PROVIDER` env var; with neither it prefers ElevenLabs and falls back
+ * to whatever else has a key. A retryable failure is retried with backoff, and
+ * one that survives the retries falls through to the next available vendor —
+ * unless the error is the caller's (bad audio), which no other vendor would fix.
+ */
+export class SpeechService {
+  private readonly config: Required<SpeechServiceConfig>;
+  private readonly providers: Map<string, ISpeechProvider>;
+
+  constructor(config: SpeechServiceConfig = {}) {
+    const env: Record<string, string | undefined> = typeof process !== 'undefined' ? process.env : {};
+    this.config = {
+      preferredProvider: config.preferredProvider ?? env.SPEECH_PROVIDER ?? 'elevenlabs',
+      maxRetries: config.maxRetries ?? 2,
+      fallback: config.fallback ?? true,
+      retryDelayMs: config.retryDelayMs ?? 500,
+    };
+    this.providers = env.SPEECH_PROVIDER === 'mock'
+      ? new Map([['mock', new MockSpeechProvider() as ISpeechProvider]])
+      : speechProviders.instantiate();
+  }
+
+  isAvailable(): boolean {
+    return this.getAvailableProviders().length > 0;
+  }
+
+  getAvailableProviders(): string[] {
+    return [...this.providers].filter(([, p]) => p.isAvailable() && p.capabilities.transcribe).map(([n]) => n);
+  }
+
+  /** Providers in the order they will be tried. */
+  private order(): ISpeechProvider[] {
+    const usable = new Map([...this.providers].filter(([, p]) => p.isAvailable() && p.capabilities.transcribe));
+    const first = pickProvider(usable, this.config.preferredProvider);
+    if (!first) return [];
+    return this.config.fallback ? [first, ...[...usable.values()].filter((p) => p !== first)] : [first];
+  }
+
+  async transcribe(input: TranscribeRequest): Promise<TranscriptResult> {
+    // A deployment-wide language default stops auto-detect drifting on short clips
+    // (an English note came back in French once). A per-call `language` wins.
+    const env: Record<string, string | undefined> = typeof process !== 'undefined' ? process.env : {};
+    const request = input.language || !env.SPEECH_LANGUAGE ? input : { ...input, language: env.SPEECH_LANGUAGE };
+    const order = this.order();
+    if (!order.length) {
+      throw new LLMProviderError(
+        'No speech providers available. Set ELEVENLABS_API_KEY or OPENAI_API_KEY.',
+        'AUTHENTICATION_ERROR', false,
+      );
+    }
+    let last: unknown;
+    for (const provider of order) {
+      for (let attempt = 0; ; attempt += 1) {
+        try {
+          return await provider.transcribe!(request);
+        } catch (error) {
+          last = error;
+          const typed = error instanceof LLMProviderError ? error : null;
+          if (typed?.retryable && attempt < this.config.maxRetries) {
+            await sleep(this.config.retryDelayMs * 2 ** attempt);
+            continue;
+          }
+          // The caller's input is bad — another vendor would say the same.
+          if (typed?.code === 'INVALID_REQUEST') throw error;
+          break;
+        }
+      }
+    }
+    throw last;
+  }
+}

package/src/speech/types.ts

@@ -0,0 +1,59 @@
+/**
+ * Speech types — provider-neutral. Nothing here names a vendor, an app, or a
+ * use case (voice notes, video narration): a provider maps its own API onto
+ * these shapes and a caller never sees the vendor's.
+ */
+export interface TranscriptWord {
+  text: string;
+  /** Seconds from the start of the audio. */
+  start: number;
+  end: number;
+  speaker?: string;
+}
+
+export interface SpeechUsage {
+  /** Billable input, when the provider reports it; else derived from word timings. */
+  audioSeconds?: number;
+  /** Characters synthesized (TTS). */
+  characters?: number;
+}
+
+export interface TranscribeRequest {
+  audio: Uint8Array;
+  /** The recorder's MIME type, parameters allowed (`audio/webm;codecs=opus`). */
+  mimeType: string;
+  /** ISO-639 hint. Omit to auto-detect. */
+  language?: string;
+  /**
+   * Words the speaker is likely to say (names, product terms). Providers that
+   * can bias recognition do; the rest ignore it.
+   */
+  hints?: string[];
+  /** `word` asks for per-word timings; providers that cannot return `words: undefined`. */
+  timestamps?: 'none' | 'word';
+  diarize?: boolean;
+  /** Provider-specific model override. */
+  model?: string;
+  timeoutMs?: number;
+}
+
+export interface TranscriptResult {
+  text: string;
+  language?: string;
+  words?: TranscriptWord[];
+  /** The model that answered, as the vendor reports it. */
+  model: string;
+  provider: string;
+  usage: SpeechUsage;
+}
+
+export interface SpeechServiceConfig {
+  /** Provider tried first. Default: `SPEECH_PROVIDER`, else `elevenlabs`. */
+  preferredProvider?: string;
+  /** Retries on a retryable error, per provider. Default 2. */
+  maxRetries?: number;
+  /** On a failure that survives retries, try the next available provider. Default true. */
+  fallback?: boolean;
+  /** Base delay for exponential backoff, ms. Default 500. */
+  retryDelayMs?: number;
+}

package/src/types.ts

@@ -38,6 +38,18 @@ export interface LLMImage {
   mediaType: 'image/png' | 'image/jpeg' | 'image/gif' | 'image/webp';
 }
 
+/**
+ * A document attached to a prompt (e.g. a PDF a user uploaded): base64 data
+ * (no data: prefix). Providers send it as their native document/file part, so a
+ * text-layer or scanned PDF reaches the model without app-side extraction.
+ */
+export interface LLMDocument {
+  data: string;
+  mediaType: 'application/pdf';
+  /** Original filename; OpenAI's file part requires one (a default is used when absent). */
+  filename?: string;
+}
+
 export interface GenerateOptions {
   /** Model to use (provider-specific) */
   model?: string;
@@ -58,6 +70,11 @@ export interface GenerateOptions {
    * a model look at a rendering of what it designed).
    */
   images?: LLMImage[];
+  /**
+   * Documents (PDFs) sent with the user prompt, as the provider's native
+   * document part (Anthropic `document` block, OpenAI `file` part).
+   */
+  documents?: LLMDocument[];
   /**
    * Prompt caching (OpenAI). Requests that share a long identical opening
    * (instructions first, per-request data last) are billed at the cached rate