@juspay/neurolink 12.31.1 → 12.32.1

package/dist/cli/loop/optionsSchema.d.ts

@@ -4,4 +4,4 @@ import type { OptionSchema, TextGenerationOptions } from "../../types/index.js";
  * This object provides metadata for validation and help text in the CLI loop.
  * It is derived from the main TextGenerationOptions interface to ensure consistency.
  */
-export declare const textGenerationOptionsSchema: Record<keyof Omit<TextGenerationOptions, "prompt" | "input" | "schema" | "tools" | "context" | "conversationHistory" | "conversationMessages" | "conversationMemoryConfig" | "originalPrompt" | "middleware" | "expectedOutcome" | "evaluationCriteria" | "region" | "csvOptions" | "pdfOptions" | "imageOptions" | "videoOptions" | "officeOptions" | "tts" | "stt" | "thinkingConfig" | "requestId" | "fileRegistry" | "abortSignal" | "toolFilter" | "excludeTools" | "toolChoice" | "prepareStep" | "credentials" | "onFinish" | "onError" | "processors" | "piiDetection" | "responseValidation" | "inputValidation" | "toolExecutionCapture" | "toolExecutionRecorder" | "toolRoots">, OptionSchema>;
+export declare const textGenerationOptionsSchema: Record<keyof Omit<TextGenerationOptions, "prompt" | "input" | "schema" | "tools" | "context" | "conversationHistory" | "conversationMessages" | "conversationMemoryConfig" | "originalPrompt" | "middleware" | "expectedOutcome" | "evaluationCriteria" | "region" | "csvOptions" | "pdfOptions" | "imageOptions" | "audioOptions" | "videoOptions" | "officeOptions" | "tts" | "stt" | "thinkingConfig" | "requestId" | "fileRegistry" | "abortSignal" | "toolFilter" | "excludeTools" | "toolChoice" | "prepareStep" | "credentials" | "onFinish" | "onError" | "processors" | "piiDetection" | "responseValidation" | "inputValidation" | "toolExecutionCapture" | "toolExecutionRecorder" | "toolRoots">, OptionSchema>;

package/dist/core/modules/MessageBuilder.js

@@ -105,6 +105,10 @@ export class MessageBuilder {
                     csvOptions: options.csvOptions,
                     pdfOptions: options.pdfOptions,
                     imageOptions: options.imageOptions,
+                    // The audio option bag is carried the same way: omitted here, a
+                    // caller's transcription backend, model and language never reach
+                    // AudioProcessor and it silently falls back to auto-selection.
+                    audioOptions: options.audioOptions,
                     // Same reason as csvOptions: the keyframe budget and the
                     // transcription flag are read downstream by the detector, and
                     // a reconstruction that omits them silently restores defaults.
@@ -245,6 +249,10 @@ export class MessageBuilder {
                     csvOptions: options.csvOptions,
                     pdfOptions: options.pdfOptions,
                     imageOptions: options.imageOptions,
+                    // The audio option bag is carried the same way: omitted here, a
+                    // caller's transcription backend, model and language never reach
+                    // AudioProcessor and it silently falls back to auto-selection.
+                    audioOptions: options.audioOptions,
                     // Same reason as csvOptions: the keyframe budget and the
                     // transcription flag are read downstream by the detector, and
                     // a reconstruction that omits them silently restores defaults.

package/dist/neurolink.js

@@ -4280,6 +4280,10 @@ Current user's request: ${currentInput}`;
             csvOptions: options.csvOptions,
             pdfOptions: options.pdfOptions,
             imageOptions: options.imageOptions,
+            // #413/#440: the audio option bag belongs in the same list. Left out
+            // here it was accepted by the public type and then silently discarded
+            // before the message builder, like every field above.
+            audioOptions: options.audioOptions,
             // videoOptions belongs in that same list and was never in it, so no SDK
             // caller's video settings have ever reached a provider -- frames and
             // quality included, not just the native-video flags added alongside.

package/dist/processors/media/AudioProcessor.d.ts

@@ -37,7 +37,7 @@
  * ```
  */
 import { BaseFileProcessor } from "../base/BaseFileProcessor.js";
-import type { FileInfo, ProcessedAudio, ProcessorFileProcessingResult, ProcessOptions } from "../../types/index.js";
+import type { AudioProcessorOptions, FileInfo, ProcessedAudio, ProcessorFileProcessingResult, ProcessOptions } from "../../types/index.js";
 /**
  * Audio Processor - extracts metadata and tags from audio files for LLM consumption.
  *
@@ -83,17 +83,46 @@ export declare class AudioProcessor extends BaseFileProcessor<ProcessedAudio> {
      *
      * @param fileInfo - File information (can include URL or buffer)
      * @param options - Optional processing options (auth headers, timeout, etc.)
+     *   widened with the transcription knobs (#413/#440) so a caller-chosen
+     *   backend, language or prompt reaches the transcriber instead of being
+     *   discarded at the detector boundary.
      * @returns Processing result with audio metadata or error
      */
-    processFile(fileInfo: FileInfo, options?: ProcessOptions): Promise<ProcessorFileProcessingResult<ProcessedAudio>>;
+    processFile(fileInfo: FileInfo, options?: ProcessOptions & AudioProcessorOptions): Promise<ProcessorFileProcessingResult<ProcessedAudio>>;
     /**
-     * Attempt to transcribe audio using the Vercel AI SDK's `transcribe()` function
-     * with the OpenAI Whisper model.
+     * Whether a transcription backend has the credentials it needs (#413).
+     *
+     * Availability is decided from the environment only — no network call — so
+     * selection stays cheap and cannot itself fail. A key that is present but
+     * rejected upstream surfaces later, as a transcription failure with the
+     * provider's own message, not as "unavailable".
+     *
+     * @param provider - Backend to check
+     * @returns True when at least one of the backend's credential vars is set
+     */
+    private isProviderAvailable;
+    /**
+     * Choose the transcription backend for this file (#413).
+     *
+     * With no `requested` backend, the first configured entry of
+     * {@link TRANSCRIPTION_PROVIDER_ORDER} wins (OpenAI, then Google, then
+     * Azure). With one, it is normalised through the alias table and validated
+     * for availability — a backend the caller explicitly asked for is never
+     * silently swapped for a different one, because a caller who pinned Azure
+     * for a data-residency reason would not want OpenAI chosen behind their
+     * back. The mismatch is reported instead.
+     *
+     * @param requested - Caller's `AudioProcessorOptions.provider`, if any
+     * @returns The chosen backend, or the reason no backend could be chosen
+     */
+    private selectProvider;
+    /**
+     * Attempt to transcribe audio with whichever backend is configured (#413).
      *
      * Transcription is attempted when:
-     * 1. `OPENAI_API_KEY` environment variable is set
-     * 2. File size is within Whisper's 25MB limit
-     * 3. File format is supported by Whisper
+     * 1. A backend's credentials are present (see {@link selectProvider})
+     * 2. File size is within the 25MB ceiling
+     * 3. The file format is one the backend accepts
      *
      * Gracefully degrades: if transcription fails for any reason, metadata-only
      * output is returned (transcription is additive, never blocks processing).
@@ -101,9 +130,49 @@ export declare class AudioProcessor extends BaseFileProcessor<ProcessedAudio> {
      * @param buffer - Audio file content
      * @param filename - Original filename (used for format detection)
      * @param mimetype - MIME type of the audio file
-     * @returns Transcription result with transcript text, or empty result
+     * @param options - Caller's transcription knobs (provider, language, prompt)
+     * @returns Transcription result with transcript text, or the reason there is none
      */
     private attemptTranscription;
+    /**
+     * Transcribe via OpenAI Whisper (#416).
+     *
+     * A native multipart POST to OpenAI's transcription endpoint. This used to
+     * go through @ai-sdk/openai's createOpenAI().transcription() plus the ai
+     * package's experimental_transcribe; both were dropped, and this is the
+     * only wire behaviour of theirs the processor ever depended on. The same
+     * request is already made natively by voice/providers/OpenAISTT.ts.
+     *
+     * `verbose_json` is requested for `language` and `duration` alongside the
+     * text — the response was previously parsed for `text` alone, so both were
+     * received and discarded, and `FileProcessingResult.metadata.language` had
+     * nothing to report (#409).
+     *
+     * @param buffer - Audio file content
+     * @param filename - Original filename, sent as the multipart part name
+     * @param mimetype - MIME type used for the upload blob
+     * @param options - Caller's language / model / prompt overrides
+     * @param skipped - Builds the "no transcript, and here is why" outcome
+     * @returns Transcription result, or the reason there is none
+     */
+    private transcribeWithOpenAI;
+    /**
+     * Transcribe via one of the non-OpenAI STT handlers (#413).
+     *
+     * Google and Azure already have working, tested implementations under
+     * `src/lib/voice/providers/`, reached here through a dynamic import so a
+     * file-processing run that never transcribes does not pay to load them —
+     * the same reason `music-metadata` is loaded lazily above.
+     *
+     * @param chosen - Backend and display label from {@link selectProvider}
+     * @param buffer - Audio file content
+     * @param filename - Original filename; its extension picks the wire format
+     * @param mimetype - MIME type, used as the fallback format signal
+     * @param options - Caller's language / model overrides
+     * @param skipped - Builds the "no transcript, and here is why" outcome
+     * @returns Transcription result, or the reason there is none
+     */
+    private transcribeWithHandler;
     /**
      * Stub implementation required by BaseFileProcessor.
      * Not used because processFile is fully overridden.
@@ -161,6 +230,11 @@ export declare class AudioProcessor extends BaseFileProcessor<ProcessedAudio> {
      * @param metadata - Extracted audio metadata
      * @param tags - Extracted audio tags
      * @param transcript - Optional transcribed text from Whisper
+     * @param skippedReason - When `transcript` is absent, why transcription was
+     *   skipped (no backend configured, unavailable pinned backend, format/size
+     *   limit, or a backend call that failed) — inlined so the model is told the
+     *   reason instead of silently receiving metadata with no transcript, which
+     *   `AUDIO_TRANSCRIPTION_INSTRUCTIONS` (messageBuilder.ts) tells it to expect.
      * @returns Formatted text content string
      *
      * @example Output:
@@ -239,7 +313,10 @@ export declare function isAudioFile(mimetype: string, filename: string): boolean
  * Convenience function that uses the singleton processor.
  *
  * @param fileInfo - File information (can include URL or buffer)
- * @param options - Optional processing options (auth headers, timeout, etc.)
+ * @param options - Optional processing options (auth headers, timeout, etc.),
+ *   widened with the transcription knobs (#413/#440, matching
+ *   {@link AudioProcessor.processFile}) so a caller can pin a provider,
+ *   language, model or prompt through this convenience export too.
  * @returns Processing result with audio metadata or error
  *
  * @example
@@ -263,4 +340,4 @@ export declare function isAudioFile(mimetype: string, filename: string): boolean
  * }
  * ```
  */
-export declare function processAudio(fileInfo: FileInfo, options?: ProcessOptions): Promise<ProcessorFileProcessingResult<ProcessedAudio>>;
+export declare function processAudio(fileInfo: FileInfo, options?: ProcessOptions & AudioProcessorOptions): Promise<ProcessorFileProcessingResult<ProcessedAudio>>;