Queue OCR & ASR work

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.7.1",
+	"version": "0.7.7",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -25,14 +25,14 @@
 		"watch": "npx vite build --watch"
 	},
 	"dependencies": {
-		"@anthropic-ai/sdk": "^0.67.0",
+		"@anthropic-ai/sdk": "^0.78.0",
 		"@tensorflow/tfjs": "^4.22.0",
 		"@xenova/transformers": "^2.17.2",
-		"@ztimson/node-utils": "^1.0.4",
+		"@ztimson/node-utils": "^1.0.7",
-		"@ztimson/utils": "^0.27.9",
+		"@ztimson/utils": "^0.28.13",
 		"cheerio": "^1.2.0",
-		"openai": "^6.6.0",
+		"openai": "^6.22.0",
-		"tesseract.js": "^6.0.1",
+		"tesseract.js": "^7.0.0",
 		"wavefile": "^11.0.0"
 	},
 	"devDependencies": {

src/asr.ts

@@ -9,15 +9,20 @@ import wavefile from 'wavefile';
 
 let whisperPipeline: any;
 
-export async function canDiarization(): Promise<boolean> {
+export async function canDiarization(): Promise<string | null> {
-	return new Promise((resolve) => {
+	const checkPython = (cmd: string) => {
-		const proc = spawn('python', ['-c', 'import pyannote.audio']);
+		return new Promise<boolean>((resolve) => {
+			const proc = spawn(cmd, ['-c', 'import pyannote.audio']);
 			proc.on('close', (code: number) => resolve(code === 0));
 			proc.on('error', () => resolve(false));
 		});
+	};
+	if(await checkPython('python3')) return 'python3';
+	if(await checkPython('python')) return 'python';
+	return null;
 }
 
-async function runDiarization(audioPath: string, dir: string, token: string): Promise<any[]> {
+async function runDiarization(binary: string, audioPath: string, dir: string, token: string): Promise<any[]> {
 	const script = `
 import sys
 import json
@@ -37,7 +42,7 @@ print(json.dumps(segments))
 
 	return new Promise((resolve, reject) => {
 		let output = '';
-		const proc = spawn('python', ['-c', script, audioPath]);
+		const proc = spawn(binary, ['-c', script, audioPath]);
 		proc.stdout.on('data', (data: Buffer) => output += data.toString());
 		proc.stderr.on('data', (data: Buffer) => console.error(data.toString()));
 		proc.on('close', (code: number) => {
@@ -105,30 +110,28 @@ function prepareAudioBuffer(file: string): [string, Float32Array] {
 }
 
 parentPort?.on('message', async ({ file, speaker, model, modelDir, token }) => {
+	let tempFile = null;
 	try {
 		if(!whisperPipeline) whisperPipeline = await pipeline('automatic-speech-recognition', `Xenova/${model}`, {cache_dir: modelDir, quantized: true});
 
-		// Prepare audio file
 		const [f, buffer] = prepareAudioBuffer(file);
-
+		tempFile = f !== file ? f : null;
-		// Fetch transcript and speakers
+		const hasDiarization = await canDiarization();
-		const hasDiarization = speaker && await canDiarization();
 		const [transcript, speakers] = await Promise.all([
 			whisperPipeline(buffer, {return_timestamps: speaker ? 'word' : false}),
-			(!speaker || !token || !hasDiarization) ? Promise.resolve(): runDiarization(f, modelDir, token),
+			(!speaker || !token || !hasDiarization) ? Promise.resolve(): runDiarization(hasDiarization, f, modelDir, token),
 		]);
-		if(file != f) rmSync(f, { recursive: true, force: true });
 
-		// Return any results / errors if no more processing required
 		const text = transcript.text?.trim() || null;
 		if(!speaker) return parentPort?.postMessage({ text });
 		if(!token) return parentPort?.postMessage({ text, error: 'HuggingFace token required' });
 		if(!hasDiarization) return parentPort?.postMessage({ text, error: 'Speaker diarization unavailable' });
 
-		// Combine transcript and speakers
 		const combined = combineSpeakerTranscript(transcript.chunks || [], speakers || []);
 		parentPort?.postMessage({ text: combined });
 	} catch (err: any) {
 		parentPort?.postMessage({ error: err.stack || err.message });
+	} finally {
+		if(tempFile) rmSync(tempFile, { recursive: true, force: true });
 	}
 });

src/audio.ts

@@ -5,45 +5,67 @@ import {canDiarization} from './asr.ts';
 import {dirname, join} from 'path';
 
 export class Audio {
+	private busy = false;
+	private currentJob: any;
+	private queue: Array<{file: string, model: string, speaker: boolean | 'id', modelDir: string, token: string, resolve: any, reject: any}> = [];
+	private worker: Worker | null = null;
+
 	constructor(private ai: Ai) {}
 
+	private processQueue() {
+		if(this.busy || !this.queue.length) return;
+
+		this.busy = true;
+		const job = this.queue.shift()!;
+		if(!this.worker) {
+			this.worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'asr.js'));
+			this.worker.on('message', this.handleMessage.bind(this));
+			this.worker.on('error', this.handleError.bind(this));
+		}
+
+		this.currentJob = job;
+		this.worker.postMessage({file: job.file, model: job.model, speaker: job.speaker, modelDir: job.modelDir, token: job.token});
+	}
+
+	private handleMessage({text, warning, error}: any) {
+		const job = this.currentJob!;
+		this.busy = false;
+		if(error) job.reject(new Error(error));
+		else {
+			if(warning) console.warn(warning);
+			job.resolve(text);
+		}
+		this.processQueue();
+	}
+
+	private handleError(err: Error) {
+		if(this.currentJob) {
+			this.currentJob.reject(err);
+			this.busy = false;
+			this.processQueue();
+		}
+	}
+
 	asr(file: string, options: { model?: string; speaker?: boolean | 'id' } = {}): AbortablePromise<string | null> {
 		const { model = this.ai.options.asr || 'whisper-base', speaker = false } = options;
 		let aborted = false;
 		const abort = () => { aborted = true; };
-
 		let p = new Promise<string | null>((resolve, reject) => {
-			const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'asr.js'));
+			this.queue.push({file, model, speaker, modelDir: <string>this.ai.options.path, token: <string>this.ai.options.hfToken,
-			const handleMessage = ({ text, warning, error }: any) => {
+				resolve: (text: string | null) => !aborted && resolve(text),
-				worker.terminate();
+				reject: (err: Error) => !aborted && reject(err)
-				if(aborted) return;
-				if(error) reject(new Error(error));
-				else {
-					if(warning) console.warn(warning);
-					resolve(text);
-				}
-			};
-			const handleError = (err: Error) => {
-				worker.terminate();
-				if(!aborted) reject(err);
-			};
-			worker.on('message', handleMessage);
-			worker.on('error', handleError);
-			worker.on('exit', (code) => {
-				if(code !== 0 && !aborted) reject(new Error(`Worker exited with code ${code}`));
 			});
-			worker.postMessage({file, model, speaker, modelDir: this.ai.options.path, token: this.ai.options.hfToken});
+			this.processQueue();
 		});
 
-		// Name speakers using AI
 		if(options.speaker == 'id') {
 			if(!this.ai.language.defaultModel) throw new Error('Configure an LLM for advanced ASR speaker detection');
 			p = p.then(async transcript => {
 				if(!transcript) return transcript;
 				let chunks = this.ai.language.chunk(transcript, 500, 0);
 				if(chunks.length > 4) chunks = [...chunks.slice(0, 3), <string>chunks.at(-1)];
-				const names = await this.ai.language.json(chunks.join('\n'), '{1: "Detected Name"}', {
+				const names = await this.ai.language.json(chunks.join('\n'), '{1: "Detected Name", 2: "Second Name"}', {
-					system: 'Use this following transcript to identify speakers. Only identify speakers you are sure about',
+					system: 'Use the following transcript to identify speakers. Only identify speakers you are positive about, dont mention speakers you are unsure about in your response',
 					temperature: 0.1,
 				});
 				Object.entries(names).forEach(([speaker, name]) => {
@@ -56,5 +78,5 @@ export class Audio {
 		return Object.assign(p, { abort });
 	}
 
-	canDiarization = canDiarization;
+	canDiarization = () => canDiarization().then(resp => !!resp);
 }

src/llm.ts

@@ -255,11 +255,11 @@ class LLM {
 	/**
 	 * Create a vector representation of a string
 	 * @param {object | string} target Item that will be embedded (objects get converted)
-	 * @param {number} maxTokens Chunking size. More = better context, less = more specific (Search by paragraphs or lines)
+	 * @param {maxTokens?: number, overlapTokens?: number} opts Options for embedding such as chunk sizes
-	 * @param {number} overlapTokens Includes previous X tokens to provide continuity to AI (In addition to max tokens)
 	 * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
 	 */
-	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+	async embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}) {
+		let {maxTokens = 500, overlapTokens = 50} = opts;
 		const embed = (text: string): Promise<number[]> => {
 			return new Promise((resolve, reject) => {
 				const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
@@ -279,13 +279,13 @@ class LLM {
 				worker.postMessage({text, model: this.ai.options?.embedder || 'bge-small-en-v1.5', modelDir: this.ai.options.path});
 			});
 		};
-		const chunks = this.chunk(target, maxTokens, overlapTokens);
+		const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
-		return Promise.all(chunks.map(async (text, index) => ({
+		for(let i = 0; i < chunks.length; i++) {
-			index,
+			const text= chunks[i];
-			embedding: await embed(text),
+			const embedding = await embed(text);
-			text,
+			results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
-			tokens: this.estimateTokens(text),
+		}
-		})));
+		return results;
 	}
 
 	/**

src/vision.ts

@@ -2,8 +2,26 @@ import {createWorker} from 'tesseract.js';
 import {AbortablePromise, Ai} from './ai.ts';
 
 export class Vision {
+	private worker: any = null;
+	private queue: Array<{ path: string, resolve: any, reject: any }> = [];
+	private busy = false;
 
-	constructor(private ai: Ai) { }
+	constructor(private ai: Ai) {}
+
+	private async processQueue() {
+		if(this.busy || !this.queue.length) return;
+		this.busy = true;
+		const job = this.queue.shift()!;
+		if(!this.worker) this.worker = await createWorker(this.ai.options.ocr || 'eng', 2, {cachePath: this.ai.options.path});
+		try {
+			const {data} = await this.worker.recognize(job.path);
+			job.resolve(data.text.trim() || null);
+		} catch(err) {
+			job.reject(err);
+		}
+		this.busy = false;
+		this.processQueue();
+	}
 
 	/**
 	 * Convert image to text using Optical Character Recognition
@@ -11,13 +29,16 @@ export class Vision {
 	 * @returns {AbortablePromise<string | null>} Promise of extracted text with abort method
 	 */
 	ocr(path: string): AbortablePromise<string | null> {
-		let worker: any;
+		let aborted = false;
-		const p = new Promise<string | null>(async res => {
+		const abort = () => { aborted = true; };
-			worker = await createWorker(this.ai.options.ocr || 'eng', 2, {cachePath: this.ai.options.path});
+		const p = new Promise<string | null>((resolve, reject) => {
-			const {data} = await worker.recognize(path);
+			this.queue.push({
-			await worker.terminate();
+				path,
-			res(data.text.trim() || null);
+				resolve: (text: string | null) => !aborted && resolve(text),
+				reject: (err: Error) => !aborted && reject(err)
 			});
-		return Object.assign(p, {abort: () => worker?.terminate()});
+			this.processQueue();
+		});
+		return Object.assign(p, {abort});
 	}
 }

vite.config.ts

@@ -1,6 +1,5 @@
 import {defineConfig} from 'vite';
 import dts from 'vite-plugin-dts';
-import {resolve} from 'path';
 
 export default defineConfig({
 	build: {