Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.
JavaScript aianthropiccross-platformgeminigroqlinuxmacosnvidiaopen-sourceopenaiparakeetspeech-to-texttranscribewhisperwindows
import { getCleanupSystemPrompt } from "../config/prompts";
import { getSettings } from "../stores/settingsStore";
import { resolveCleanupLanguage } from "../utils/chineseScript";
import { getDictionaryHintWords } from "../utils/snippets";
import type { InferenceScope } from "../config/inferenceScopes";
import type { DictationToneStyle } from "../config/toneStyle";
import type { ScreenContextImage } from "../types/electron";
export interface ReasoningConfig {
maxTokens?: number;
temperature?: number;
contextSize?: number;
systemPrompt?: string;
lanUrl?: string;
baseUrl?: string;
customApiKey?: string;
provider?: string;
disableThinking?: boolean;
reasoningEffort?: string;
/** Screenshot attached to voice-agent requests when screen context is on. */
screenContext?: ScreenContextImage;
/** Dictation cleanup only: the Tone & Style tone for the app being dictated into. */
toneStyle?: DictationToneStyle | null;
/** Suffix-free prompt used when a screenshot-carrying request is retried text-only. */
textOnlySystemPrompt?: string;
language?: string;
requireCompleteOutput?: boolean;
/** Bundled local server only; callers must validate the decoded response. */
responseFormat?: { type: "json_object"; schema: Record<string, unknown> };
/**
* Local models only: when the prompt leaves less than `maxTokens` of room and
* the reply fills what is left, fail as CONTEXT_TOO_LARGE instead of
* returning a reply the context window clipped. With `requireCompleteOutput`,
* it still lets the allowance shrink to fit, where that flag alone refuses.
*/
refuseClippedByWindow?: boolean;
/**
* Tags the request so a cancel can abort it: `cancelLocalReasoning` for local
* models, `cancelRequest` (ai/requestCancellation) for the renderer's HTTP routes.
*/
requestId?: string;
requiresAgent?: boolean;
inferenceScope?: InferenceScope;
}
export abstract class BaseReasoningService {
protected isProcessing = false;
protected getCustomDictionary(): string[] {
return getDictionaryHintWords(getSettings());
}
// Auto must remain auto here: zh-CN/zh-TW instructions make cleanup write its
// entire response in Chinese before the transcription language is known. The
// final deterministic script pass handles likely-Chinese output instead. See #975.
protected getPreferredLanguage(): string {
return resolveCleanupLanguage(getSettings().preferredLanguage);
}
protected getUiLanguage(): string {
return getSettings().uiLanguage || "en";
}
protected getSystemPrompt(
agentName: string | null,
toneStyle?: DictationToneStyle | null
): string {
return getCleanupSystemPrompt(
agentName,
this.getCustomDictionary(),
this.getPreferredLanguage(),
this.getUiLanguage(),
toneStyle
);
}
protected calculateMaxTokens(
textLength: number,
minTokens = 100,
maxTokens = 2048,
multiplier = 2
): number {
return Math.max(minTokens, Math.min(textLength * multiplier, maxTokens));
}
abstract isAvailable(): Promise<boolean>;
abstract processText(
text: string,
modelId: string,
agentName?: string | null,
config?: ReasoningConfig
): Promise<string>;
}