@qvac/sdk 0.18.2 → 0.19.0

package/dist/schemas/transcription.js

@@ -1,145 +0,0 @@
-import { z } from 'zod';
-export const audioInputSchema = z.discriminatedUnion('type', [
-    z.object({
-        type: z.literal('base64'),
-        value: z.string()
-    }),
-    z.object({
-        type: z.literal('filePath'),
-        value: z.string()
-    })
-]);
-const transcribeBaseSchema = z.object({
-    modelId: z.string(),
-    /**
-     * Initial transcription prompt. Whisper engine only — silently
-     * ignored by the parakeet engine, which has no equivalent prompting
-     * surface in `qvac-parakeet.cpp`.
-     */
-    prompt: z.string().optional(),
-    metadata: z.boolean().optional()
-});
-export const transcribeParamsSchema = transcribeBaseSchema.extend({
-    audioChunk: audioInputSchema
-});
-export const transcribeStatsSchema = z.object({
-    audioDuration: z.number().optional(),
-    realTimeFactor: z.number().optional(),
-    tokensPerSecond: z.number().optional(),
-    totalTokens: z.number().optional(),
-    totalSegments: z.number().optional(),
-    whisperEncodeTime: z.number().optional(),
-    whisperDecodeTime: z.number().optional(),
-    encoderTime: z.number().optional(),
-    decoderTime: z.number().optional(),
-    melSpecTime: z.number().optional(),
-    // Backend selection captured once at model load.
-    // `0` CPU / `1` GPU; `backendId` codes the GPU family (see the addon's
-    // BackendId enum). gpu memory fields are `-1` when the backend exposes
-    // no accounting, and absent entirely for the parakeet engine.
-    backendDevice: z.number().optional(),
-    backendId: z.number().optional(),
-    gpuUnsupported: z.number().optional(),
-    gpuMemTotalMb: z.number().optional(),
-    gpuMemFreeMb: z.number().optional()
-});
-export const transcribeSegmentSchema = z.object({
-    text: z.string(),
-    startMs: z.number(),
-    endMs: z.number(),
-    append: z.boolean(),
-    id: z.number()
-});
-export const vadStateEventSchema = z.object({
-    speaking: z.boolean(),
-    probability: z.number()
-});
-export const whisperEndOfTurnEventSchema = z.object({
-    source: z.literal('whisper'),
-    silenceDurationMs: z.number()
-});
-export const parakeetEndOfTurnEventSchema = z.object({
-    source: z.literal('parakeet')
-});
-const endOfTurnDiscriminatedSchema = z.discriminatedUnion('source', [
-    whisperEndOfTurnEventSchema,
-    parakeetEndOfTurnEventSchema
-]);
-/** Legacy whisper wire frames omitted `source` and only sent `silenceDurationMs`. */
-function normalizeEndOfTurnWire(value) {
-    if (typeof value === 'object' &&
-        value !== null &&
-        !('source' in value) &&
-        typeof value.silenceDurationMs === 'number') {
-        return {
-            source: 'whisper',
-            silenceDurationMs: value.silenceDurationMs
-        };
-    }
-    return value;
-}
-export const endOfTurnEventSchema = z.preprocess(normalizeEndOfTurnWire, endOfTurnDiscriminatedSchema);
-export const transcribeRequestSchema = transcribeParamsSchema.extend({
-    type: z.literal('transcribe'),
-    requestId: z
-        .string()
-        .min(1)
-        .optional()
-        .describe('Stable identifier for this in-flight transcription, generated by the client at call time. Surfaces on the registry so callers can target it with `cancel({ requestId })`. Optional on the wire — the server falls back to a server-generated id when the field is missing.')
-});
-const transcriptionResultBase = z.object({
-    text: z.string().optional(),
-    done: z.boolean().optional(),
-    stats: transcribeStatsSchema.optional(),
-    error: z.string().optional(),
-    segment: transcribeSegmentSchema.optional(),
-    vad: vadStateEventSchema.optional(),
-    endOfTurn: endOfTurnEventSchema.optional()
-});
-export const transcribeResponseSchema = transcriptionResultBase.extend({
-    type: z.literal('transcribe')
-});
-/**
- * Per-call overrides for parakeet's duplex streaming session.
- *
- * Each field maps to its `streaming*`-prefixed counterpart in
- * `parakeetConfig` (see `parakeetRuntimeConfigSchema`). The `streaming`
- * prefix is intentionally dropped here because every field on this
- * object is already namespaced under the `parakeetStreamingConfig`
- * field of `transcribeStream({ ... })`. Any field omitted falls back
- * to the load-time value.
- */
-export const parakeetStreamingRunConfigSchema = z.object({
-    chunkMs: z.number().int().positive().optional(),
-    historyMs: z.number().int().positive().optional(),
-    leftContextMs: z.number().int().nonnegative().optional(),
-    rightLookaheadMs: z.number().int().nonnegative().optional(),
-    emitPartials: z.boolean().optional(),
-    emitEnergyVad: z.boolean().optional(),
-    spkCacheEnable: z.boolean().optional(),
-    spkCacheLen: z.number().int().positive().optional(),
-    fifoLen: z.number().int().positive().optional(),
-    chunkLeftContextMs: z.number().int().nonnegative().optional(),
-    chunkRightContextMs: z.number().int().nonnegative().optional(),
-    spkCacheUpdatePeriod: z.number().int().positive().optional()
-});
-export const transcribeStreamRequestSchema = transcribeBaseSchema.extend({
-    type: z.literal('transcribeStream'),
-    emitVadEvents: z.boolean().optional(),
-    endOfTurnSilenceMs: z
-        .number()
-        .int()
-        .nonnegative()
-        .optional()
-        .describe('Silence (ms) before an endOfTurn event fires. 0 or unset disables end-of-turn detection entirely. Whisper engine only.'),
-    vadRunIntervalMs: z.number().int().positive().optional(),
-    parakeetStreamingConfig: parakeetStreamingRunConfigSchema.optional(),
-    requestId: z
-        .string()
-        .min(1)
-        .optional()
-        .describe('Stable identifier for this in-flight duplex transcription, generated by the client at call time. Surfaces on the registry so callers can target it with `cancel({ requestId })`. Optional on the wire — the server falls back to a server-generated id when the field is missing.')
-});
-export const transcribeStreamResponseSchema = transcriptionResultBase.extend({
-    type: z.literal('transcribeStream')
-});

package/dist/schemas/translate.d.ts

@@ -1,139 +0,0 @@
-import { z } from 'zod';
-import type { NmtLanguage } from './translation-config.js';
-import { type NmtModelTypeInput, type LlmModelTypeInput } from './model-types.js';
-declare const translateParamsSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
-    modelId: z.ZodString;
-    text: z.ZodUnion<readonly [z.ZodString, z.ZodArray<z.ZodString>]>;
-    stream: z.ZodBoolean;
-    modelType: z.ZodEnum<{
-        "nmtcpp-translation": "nmtcpp-translation";
-        nmt: "nmt";
-    }>;
-}, z.core.$strip>, z.ZodObject<{
-    modelId: z.ZodString;
-    text: z.ZodString;
-    stream: z.ZodBoolean;
-    modelType: z.ZodEnum<{
-        "llamacpp-completion": "llamacpp-completion";
-        llm: "llm";
-    }>;
-    from: z.ZodOptional<z.ZodString>;
-    to: z.ZodString;
-    context: z.ZodOptional<z.ZodString>;
-}, z.core.$strip>], "modelType">;
-export declare const translationStatsSchema: z.ZodObject<{
-    totalTime: z.ZodOptional<z.ZodNumber>;
-    totalTokens: z.ZodOptional<z.ZodNumber>;
-    tokensPerSecond: z.ZodOptional<z.ZodNumber>;
-    timeToFirstToken: z.ZodOptional<z.ZodNumber>;
-    decodeTime: z.ZodOptional<z.ZodNumber>;
-    encodeTime: z.ZodOptional<z.ZodNumber>;
-    cacheTokens: z.ZodOptional<z.ZodNumber>;
-}, z.core.$strip>;
-export declare const translateRequestSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
-    modelId: z.ZodString;
-    text: z.ZodUnion<readonly [z.ZodString, z.ZodArray<z.ZodString>]>;
-    stream: z.ZodBoolean;
-    modelType: z.ZodEnum<{
-        "nmtcpp-translation": "nmtcpp-translation";
-        nmt: "nmt";
-    }>;
-    type: z.ZodLiteral<"translate">;
-    requestId: z.ZodOptional<z.ZodString>;
-}, z.core.$strip>, z.ZodObject<{
-    modelId: z.ZodString;
-    text: z.ZodString;
-    stream: z.ZodBoolean;
-    modelType: z.ZodEnum<{
-        "llamacpp-completion": "llamacpp-completion";
-        llm: "llm";
-    }>;
-    from: z.ZodOptional<z.ZodString>;
-    to: z.ZodString;
-    context: z.ZodOptional<z.ZodString>;
-    type: z.ZodLiteral<"translate">;
-    requestId: z.ZodOptional<z.ZodString>;
-}, z.core.$strip>], "modelType">;
-export declare const translateServerParamsSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[z.ZodObject<{
-    modelId: z.ZodString;
-    text: z.ZodUnion<readonly [z.ZodString, z.ZodArray<z.ZodString>]>;
-    stream: z.ZodBoolean;
-    modelType: z.ZodEnum<{
-        "nmtcpp-translation": "nmtcpp-translation";
-        nmt: "nmt";
-    }>;
-}, z.core.$strip>, z.ZodObject<{
-    modelId: z.ZodString;
-    text: z.ZodString;
-    stream: z.ZodBoolean;
-    modelType: z.ZodEnum<{
-        "llamacpp-completion": "llamacpp-completion";
-        llm: "llm";
-    }>;
-    from: z.ZodOptional<z.ZodString>;
-    to: z.ZodString;
-    context: z.ZodOptional<z.ZodString>;
-}, z.core.$strip>], "modelType">, z.ZodTransform<{
-    modelType: string;
-    modelId: string;
-    text: string | string[];
-    stream: boolean;
-} | {
-    modelType: string;
-    modelId: string;
-    text: string;
-    stream: boolean;
-    to: string;
-    from?: string | undefined;
-    context?: string | undefined;
-}, {
-    modelId: string;
-    text: string | string[];
-    stream: boolean;
-    modelType: "nmtcpp-translation" | "nmt";
-} | {
-    modelId: string;
-    text: string;
-    stream: boolean;
-    modelType: "llamacpp-completion" | "llm";
-    to: string;
-    from?: string | undefined;
-    context?: string | undefined;
-}>>;
-export declare const translateResponseSchema: z.ZodObject<{
-    type: z.ZodLiteral<"translate">;
-    token: z.ZodString;
-    done: z.ZodOptional<z.ZodBoolean>;
-    stats: z.ZodOptional<z.ZodObject<{
-        totalTime: z.ZodOptional<z.ZodNumber>;
-        totalTokens: z.ZodOptional<z.ZodNumber>;
-        tokensPerSecond: z.ZodOptional<z.ZodNumber>;
-        timeToFirstToken: z.ZodOptional<z.ZodNumber>;
-        decodeTime: z.ZodOptional<z.ZodNumber>;
-        encodeTime: z.ZodOptional<z.ZodNumber>;
-        cacheTokens: z.ZodOptional<z.ZodNumber>;
-    }, z.core.$strip>>;
-    error: z.ZodOptional<z.ZodString>;
-}, z.core.$strip>;
-export type TranslateParams = z.infer<typeof translateParamsSchema>;
-export type TranslateRequest = z.infer<typeof translateRequestSchema>;
-export type TranslateResponse = z.infer<typeof translateResponseSchema>;
-export type TranslationStats = z.infer<typeof translationStatsSchema>;
-type TranslateParamsNmt = {
-    modelId: string;
-    text: string | string[];
-    stream: boolean;
-    modelType: NmtModelTypeInput;
-};
-type TranslateParamsLlm = {
-    modelId: string;
-    text: string;
-    stream: boolean;
-    modelType: LlmModelTypeInput;
-    from?: NmtLanguage | (string & {});
-    to: NmtLanguage | (string & {});
-    context?: string;
-};
-export type TranslateClientParams = TranslateParamsNmt | TranslateParamsLlm;
-export {};
-//# sourceMappingURL=translate.d.ts.map

package/dist/schemas/translate.js

@@ -1,103 +0,0 @@
-import { z } from 'zod';
-import { nmtModelTypeSchema, llmModelTypeSchema, normalizeModelType } from './model-types.js';
-const translateParamsNmtSchema = z.object({
-    modelId: z.string().describe('The identifier of the NMT translation model to use.'),
-    text: z
-        .union([
-        z.string().min(1, 'Text cannot be empty'),
-        z.array(z.string().min(1, 'Text cannot be empty')).min(1, 'Array cannot be empty')
-    ])
-        .describe('The input text(s) to translate. A single string returns a single translation; an array returns one translation per input.'),
-    stream: z
-        .boolean()
-        .describe('Whether to stream tokens (`true`) or resolve the complete translation once (`false`).'),
-    modelType: nmtModelTypeSchema.describe('NMT model-type variant identifier.')
-});
-const translateParamsLlmSchema = z.object({
-    modelId: z.string().describe('The identifier of the LLM model used for translation.'),
-    text: z.string().min(1, 'Text cannot be empty').describe('The input text to translate.'),
-    stream: z
-        .boolean()
-        .describe('Whether to stream tokens (`true`) or resolve the complete translation once (`false`).'),
-    modelType: llmModelTypeSchema.describe('LLM model-type variant identifier.'),
-    from: z
-        .string()
-        .optional()
-        .describe('Source language code. When omitted, the SDK attempts to auto-detect the source language.'),
-    to: z.string().describe('Target language code.'),
-    context: z
-        .string()
-        .optional()
-        .describe('Optional translation context passed to the LLM as a system hint.')
-});
-const translateParamsSchema = z.discriminatedUnion('modelType', [
-    translateParamsNmtSchema,
-    translateParamsLlmSchema
-]);
-export const translationStatsSchema = z.object({
-    // Common stats
-    totalTime: z.number().optional().describe('Total translation time in milliseconds.'),
-    totalTokens: z.number().optional().describe('Total tokens produced by the translation.'),
-    tokensPerSecond: z.number().optional().describe('Tokens generated per second.'),
-    timeToFirstToken: z
-        .number()
-        .optional()
-        .describe('Time to first token in milliseconds (LLM translation only).'),
-    // NMT-specific
-    decodeTime: z.number().optional().describe('Time spent in the NMT decoder in milliseconds.'),
-    encodeTime: z.number().optional().describe('Time spent in the NMT encoder in milliseconds.'),
-    // LLM-specific
-    cacheTokens: z
-        .number()
-        .optional()
-        .describe('Tokens served from the KV cache during LLM translation.')
-});
-const translateRequestIdField = z
-    .string()
-    .min(1)
-    .optional()
-    .describe('Stable identifier for this in-flight translation, generated by the client at call time. Surfaces on the registry so callers can target it with `cancel({ requestId })`. Optional on the wire — the server falls back to a server-generated id when the field is missing.');
-export const translateRequestSchema = z.discriminatedUnion('modelType', [
-    translateParamsNmtSchema
-        .extend({
-        type: z.literal('translate'),
-        requestId: translateRequestIdField
-    })
-        .meta({ title: 'TranslateNmtRequest' }),
-    translateParamsLlmSchema
-        .extend({
-        type: z.literal('translate'),
-        requestId: translateRequestIdField
-    })
-        .meta({ title: 'TranslateLlmRequest' })
-]);
-// Valid model types for translation (aliases and canonical)
-const validTranslationModelTypes = ['nmt', 'nmtcpp-translation', 'llm', 'llamacpp-completion'];
-const llmModelTypes = ['llm', 'llamacpp-completion'];
-// Validates the translate server args and returns the model info
-export const translateServerParamsSchema = translateParamsSchema
-    .refine((data) => data.modelType && validTranslationModelTypes.includes(data.modelType), {
-    message: 'Model type is not compatible with translation. Only LLM and NMT models are supported.'
-})
-    .refine((data) => {
-    if (!llmModelTypes.includes(data.modelType))
-        return true;
-    // For LLM, `to` is required; `from` is resolved by the worker
-    // (explicit, else auto-detected in server/bare/ops/translate.ts), so an
-    // absent `from` is valid here — it no longer has to arrive on the wire.
-    const llmData = data;
-    return !!llmData.to;
-}, {
-    message: "A 'to' language is required for LLM translation models"
-})
-    .transform((data) => ({
-    ...data,
-    modelType: normalizeModelType(data.modelType)
-}));
-export const translateResponseSchema = z.object({
-    type: z.literal('translate'),
-    token: z.string(),
-    done: z.boolean().optional(),
-    stats: translationStatsSchema.optional(),
-    error: z.string().optional()
-});