TTS

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.2.1",
+	"version": "0.3.0",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -42,6 +42,7 @@
 		"vite-plugin-dts": "^4.5.3"
 	},
 	"files": [
+		"bin",
 		"dist"
 	]
 }

src/ai.ts

@@ -1,22 +1,33 @@
+import * as os from 'node:os';
 import {LLM, LLMOptions} from './llm';
 import { Audio } from './audio.ts';
 import {Vision} from './vision.ts';
 
+export type AbortablePromise<T> = Promise<T> & {abort: () => any};
+
 export type AiOptions = LLMOptions & {
+	/** Path to models */
+	path?: string;
+	/** Piper TTS configuratoin */
+	piper?: {
+		/** Model URL: `https://huggingface.co/rhasspy/piper-voices/tree/main/.../model.onnx` */
+		model: string;
+	},
+	/** Tesseract OCR configuration */
+	tesseract?: {
+		/** Model: eng, eng_best, eng_fast */
+		model?: string;
+	}
+	/** Whisper ASR configuration */
 	whisper?: {
 		/** Whisper binary location */
 		binary: string;
 		/** Model: `ggml-base.en.bin` */
 		model: string;
-		/** Path to models */
-		path: string;
 	}
 }
 
 export class Ai {
-	private downloads: {[key: string]: Promise<string>} = {};
-	private whisperModel!: string;
-
 	/** Audio processing AI */
 	audio!: Audio;
 	/** Language processing AI */
@@ -25,6 +36,8 @@ export class Ai {
 	vision!: Vision;
 
 	constructor(public readonly options: AiOptions) {
+		if(!options.path) options.path = os.tmpdir();
+		process.env.TRANSFORMERS_CACHE = options.path;
 		this.audio = new Audio(this);
 		this.language = new LLM(this);
 		this.vision = new Vision(this);

src/antrhopic.ts

@@ -1,8 +1,8 @@
 import {Anthropic as anthropic} from '@anthropic-ai/sdk';
 import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, deepCopy} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 
 export class Anthropic extends LLMProvider {
 	client!: anthropic;
@@ -54,12 +54,14 @@ export class Anthropic extends LLMProvider {
 			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			const original = deepCopy(history);
 			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
+
+			const tools = options.tools || this.ai.options.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
 				max_tokens: options.max_tokens || this.ai.options.max_tokens || 4096,
 				system: options.system || this.ai.options.system || '',
 				temperature: options.temperature || this.ai.options.temperature || 0.7,
-				tools: (options.tools || this.ai.options.tools || []).map(t => ({
+				tools: tools.map(t => ({
 					name: t.name,
 					description: t.description,
 					input_schema: {
@@ -76,7 +78,10 @@ export class Anthropic extends LLMProvider {
 			let resp: any, isFirstMessage = true;
 			const assistantMessages: string[] = [];
 			do {
-				resp = await this.client.messages.create(requestParams);
+				resp = await this.client.messages.create(requestParams).catch(err => {
+					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
+					throw err;
+				});
 
 				// Streaming mode
 				if(options.stream) {
@@ -114,7 +119,8 @@ export class Anthropic extends LLMProvider {
 					history.push({role: 'assistant', content: resp.content});
 					original.push({role: 'assistant', content: resp.content});
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
-						const tool = options.tools?.find(findByProp('name', toolCall.name));
+						const tool = tools.find(findByProp('name', toolCall.name));
+						if(options.stream) options.stream({tool: toolCall.name});
 						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
 						try {
 							const result = await tool.fn(toolCall.input, this.ai);

src/audio.ts

@@ -1,30 +1,35 @@
 import {spawn} from 'node:child_process';
+import * as os from 'node:os';
+import {platform, arch} from 'node:os';
 import fs from 'node:fs/promises';
 import Path from 'node:path';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 
 export class Audio {
 	private downloads: {[key: string]: Promise<string>} = {};
 	private whisperModel!: string;
+	private piperBinary?: string;
 
 	constructor(private ai: Ai) {
 		if(ai.options.whisper?.binary) {
 			this.whisperModel = ai.options.whisper?.model.endsWith('.bin') ? ai.options.whisper?.model : ai.options.whisper?.model + '.bin';
 			this.downloadAsrModel();
 		}
+		if(ai.options.piper?.model) {
+			if(!ai.options.piper.model.startsWith('http') || !ai.options.piper.model.endsWith('.onnx'))
+				throw new Error('Piper model should be a URL to an onnx file to download');
+			if(platform() != 'linux' || (arch() != 'x64' && arch() != 'arm64'))
+				throw new Error('Piper TTS only supported on Linux x64/arm64');
+			this.piperBinary = Path.join(import.meta.dirname, '../bin/piper');
+			this.downloadTtsModel();
+		}
 	}
 
-	/**
+	asr(path: string, model: string = this.whisperModel): AbortablePromise<string | null> {
-	 * Convert audio to text using Auditory Speech Recognition
-	 * @param {string} path Path to audio
-	 * @param model Whisper model
-	 * @returns {Promise<any>} Extracted text
-	 */
-	asr(path: string, model: string = this.whisperModel): {abort: () => void, response: Promise<string | null>} {
 		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
 		let abort: any = () => {};
-		const response = new Promise<string | null>((resolve, reject) => {
+		const p = new Promise<string | null>(async (resolve, reject) => {
-			this.downloadAsrModel(model).then(m => {
+			const m = await this.downloadAsrModel(model);
 			let output = '';
 			const proc = spawn(<string>this.ai.options.whisper?.binary, ['-nt', '-np', '-m', m, '-f', path], {stdio: ['ignore', 'pipe', 'ignore']});
 			abort = () => proc.kill('SIGTERM');
@@ -35,20 +40,45 @@ export class Audio {
 				else reject(new Error(`Exit code ${code}`));
 			});
 		});
-		});
+		return Object.assign(p, {abort});
-		return {response, abort};
+	}
+
+	tts(text: string, outputPath?: string, model: string = <string>this.ai.options.piper?.model): AbortablePromise<Buffer | string> {
+		if(!this.piperBinary) throw new Error('Piper not configured');
+		if(!model) throw new Error('Invalid Piper model');
+		let abort: any = () => {};
+		const p = new Promise<Buffer | string>(async (resolve, reject) => {
+			const modelPath = await this.downloadTtsModel(model);
+			const tmpFile = outputPath || Path.join(os.tmpdir(), `piper_${Date.now()}.wav`);
+			const proc = spawn(<string>this.piperBinary, ['--model', modelPath, '--output_file', tmpFile], {
+				stdio: ['pipe', 'ignore', 'ignore'],
+				env: {...process.env, LD_LIBRARY_PATH: Path.dirname(<string>this.piperBinary)}
+			});
+			abort = () => proc.kill('SIGTERM');
+			proc.stdin.write(text);
+			proc.stdin.end();
+			proc.on('error', (err: Error) => reject(err));
+			proc.on('close', async (code: number) => {
+				if(code === 0) {
+					if(outputPath) {
+						resolve(outputPath);
+					} else {
+						const buffer = await fs.readFile(tmpFile);
+						await fs.unlink(tmpFile).catch(() => {});
+						resolve(buffer);
+					}
+				} else {
+					reject(new Error(`Exit code ${code}`));
+				}
+			});
+		});
+		return Object.assign(p, {abort});
 	}
 
-	/**
-	 * Downloads the specified Whisper model if it is not already present locally.
-	 *
-	 * @param {string} model Whisper model that will be downloaded
-	 * @return {Promise<string>} Absolute path to model file, resolves once downloaded
-	 */
 	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
 		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
 		if(!model.endsWith('.bin')) model += '.bin';
-		const p = Path.join(this.ai.options.whisper.path, model);
+		const p = Path.join(<string>this.ai.options.path, model);
 		if(await fs.stat(p).then(() => true).catch(() => false)) return p;
 		if(!!this.downloads[model]) return this.downloads[model];
 		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${model}`)
@@ -60,4 +90,24 @@ export class Audio {
 			});
 		return this.downloads[model];
 	}
+
+	async downloadTtsModel(model: string = <string>this.ai.options.piper?.model): Promise<string> {
+		if(!model) throw new Error('Invalid Piper model');
+		const m = <string>model.split('/').pop();
+		const p = Path.join(<string>this.ai.options.path, m);
+		const [onnxExists, jsonExists] = await Promise.all([
+			fs.stat(p).then(() => true).catch(() => false),
+			fs.stat(p + '.json').then(() => true).catch(() => false)
+		]);
+		if(onnxExists && jsonExists) return p;
+		if(!!this.downloads[m]) return this.downloads[m];
+		this.downloads[m] = Promise.all([
+			onnxExists ? Promise.resolve() : fetch(model).then(r => r.arrayBuffer()).then(b => fs.writeFile(p, Buffer.from(b))),
+			jsonExists ? Promise.resolve() : fetch(model + '.json').then(r => r.arrayBuffer()).then(b => fs.writeFile(p + '.json', Buffer.from(b)))
+		]).then(() => {
+			delete this.downloads[m];
+			return p;
+		});
+		return this.downloads[m];
+	}
 }

src/embedder.ts

@@ -0,0 +1,11 @@
+import { pipeline } from '@xenova/transformers';
+import { parentPort } from 'worker_threads';
+
+let model: any;
+
+parentPort?.on('message', async ({ id, text }) => {
+	if(!model) model = await pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2');
+	const output = await model(text, { pooling: 'mean', normalize: true });
+	const embedding = Array.from(output.data);
+	parentPort?.postMessage({ id, embedding });
+});

src/llm.ts

@@ -1,12 +1,13 @@
-import {pipeline} from '@xenova/transformers';
 import {JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
 import {Ollama} from './ollama.ts';
 import {OpenAi} from './open-ai.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
-import * as tf from '@tensorflow/tfjs';
+import {Worker} from 'worker_threads';
+import {fileURLToPath} from 'url';
+import {dirname, join} from 'path';
 
 export type LLMMessage = {
 	/** Message originator */
@@ -72,7 +73,7 @@ export type LLMRequest = {
 	/** LLM  model */
 	model?: string | [string, string];
 	/** Stream response */
-	stream?: (chunk: {text?: string, done?: true}) => any;
+	stream?: (chunk: {text?: string, tool?: string, done?: true}) => any;
 	/** Compress old messages in the chat to free up context */
 	compress?: {
 		/** Trigger chat compression once context exceeds the token count */
@@ -83,11 +84,22 @@ export type LLMRequest = {
 }
 
 export class LLM {
-	private embedModel: any;
+	private embedWorker: Worker | null = null;
+	private embedQueue = new Map<number, { resolve: (value: number[]) => void; reject: (error: any) => void }>();
+	private embedId = 0;
 	private providers: {[key: string]: LLMProvider} = {};
 
+
 	constructor(public readonly ai: Ai) {
-		this.embedModel = pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2');
+		this.embedWorker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
+		this.embedWorker.on('message', ({ id, embedding }) => {
+			const pending = this.embedQueue.get(id);
+			if (pending) {
+				pending.resolve(embedding);
+				this.embedQueue.delete(id);
+			}
+		});
+
 		if(ai.options.anthropic?.token) this.providers.anthropic = new Anthropic(this.ai, ai.options.anthropic.token, ai.options.anthropic.model);
 		if(ai.options.ollama?.host) this.providers.ollama = new Ollama(this.ai, ai.options.ollama.host, ai.options.ollama.model);
 		if(ai.options.openAi?.token) this.providers.openAi = new OpenAi(this.ai, ai.options.openAi.token, ai.options.openAi.model);
@@ -148,49 +160,44 @@ export class LLM {
 		return denominator === 0 ? 0 : dotProduct / denominator;
 	}
 
-	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+	chunk(target: object | string, maxTokens = 500, overlapTokens = 50): string[] {
 		const objString = (obj: any, path = ''): string[] => {
-			if(obj === null || obj === undefined) return [];
+			if(!obj) return [];
 			return Object.entries(obj).flatMap(([key, value]) => {
 				const p = path ? `${path}${isNaN(+key) ? `.${key}` : `[${key}]`}` : key;
-				if(typeof value === 'object' && value !== null && !Array.isArray(value)) return objString(value, p);
+				if(typeof value === 'object' && !Array.isArray(value)) return objString(value, p);
-				const valueStr = Array.isArray(value) ? value.join(', ') : String(value);
+				return `${p}: ${Array.isArray(value) ? value.join(', ') : value}`;
-				return `${p}: ${valueStr}`;
 			});
 		};
 
-		const embed = async (text: string): Promise<number[]> => {
+		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
-			const model = await this.embedModel;
+		const tokens = lines.flatMap(l => [...l.split(/\s+/).filter(Boolean), '\n']);
-			const output = await model(text, {pooling: 'mean', normalize: true});
+		const chunks: string[] = [];
-			return Array.from(output.data);
+		for(let i = 0; i < tokens.length;) {
+			let text = '', j = i;
+			while(j < tokens.length) {
+				const next = text + (text ? ' ' : '') + tokens[j];
+				if(this.estimateTokens(next.replace(/\s*\n\s*/g, '\n')) > maxTokens && text) break;
+				text = next;
+				j++;
+			}
+			const clean = text.replace(/\s*\n\s*/g, '\n').trim();
+			if(clean) chunks.push(clean);
+			i = Math.max(j - overlapTokens, j === i ? i + 1 : j);
+		}
+		return chunks;
+	}
+
+	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+		const embed = (text: string): Promise<number[]> => {
+			return new Promise((resolve, reject) => {
+				const id = this.embedId++;
+				this.embedQueue.set(id, { resolve, reject });
+				this.embedWorker?.postMessage({ id, text });
+			});
 		};
 
-		// Tokenize
+		const chunks = this.chunk(target, maxTokens, overlapTokens);
-		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
-		const tokens = lines.flatMap(line => [...line.split(/\s+/).filter(w => w.trim()), '\n']);
-
-		// Chunk
-		const chunks: string[] = [];
-		let start = 0;
-		while (start < tokens.length) {
-			let end = start;
-			let text = '';
-			// Build chunk
-			while (end < tokens.length) {
-				const nextToken = tokens[end];
-				const testText = text + (text ? ' ' : '') + nextToken;
-				const testTokens = this.estimateTokens(testText.replace(/\s*\n\s*/g, '\n'));
-				if (testTokens > maxTokens && text) break;
-				text = testText;
-				end++;
-			}
-			// Save chunk
-			const cleanText = text.replace(/\s*\n\s*/g, '\n').trim();
-			if(cleanText) chunks.push(cleanText);
-			start = end - overlapTokens;
-			if (start <= end - tokens.length + end) start = end; // Safety: prevent infinite loop
-		}
-
 		return Promise.all(chunks.map(async (text, index) => ({
 			index,
 			embedding: await embed(text),

src/ollama.ts

@@ -1,7 +1,7 @@
 import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 import {Ollama as ollama} from 'ollama';
 
 export class Ollama extends LLMProvider {
@@ -49,6 +49,7 @@ export class Ollama extends LLMProvider {
 			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min);
 			if(options.system) history.unshift({role: 'system', content: system})
 
+			const tools = options.tools || this.ai.options.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
 				messages: history,
@@ -58,7 +59,7 @@ export class Ollama extends LLMProvider {
 					temperature: options.temperature || this.ai.options.temperature || 0.7,
 					num_predict: options.max_tokens || this.ai.options.max_tokens || 4096,
 				},
-				tools: (options.tools || this.ai.options.tools || []).map(t => ({
+				tools: tools.map(t => ({
 					type: 'function',
 					function: {
 						name: t.name,
@@ -74,7 +75,11 @@ export class Ollama extends LLMProvider {
 
 			let resp: any, isFirstMessage = true;
 			do {
-				resp = await this.client.chat(requestParams);
+				resp = await this.client.chat(requestParams).catch(err => {
+					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
+					throw err;
+				});
+
 				if(options.stream) {
 					if(!isFirstMessage) options.stream({text: '\n\n'});
 					else isFirstMessage = false;
@@ -93,7 +98,8 @@ export class Ollama extends LLMProvider {
 				if(resp.message?.tool_calls?.length && !controller.signal.aborted) {
 					history.push(resp.message);
 					const results = await Promise.all(resp.message.tool_calls.map(async (toolCall: any) => {
-						const tool = (options.tools || this.ai.options.tools)?.find(findByProp('name', toolCall.function.name));
+						const tool = tools.find(findByProp('name', toolCall.function.name));
+						if(options.stream) options.stream({tool: toolCall.function.name});
 						if(!tool) return {role: 'tool', tool_name: toolCall.function.name, content: '{"error": "Tool not found"}'};
 						const args = typeof toolCall.function.arguments === 'string' ? JSONAttemptParse(toolCall.function.arguments, {}) : toolCall.function.arguments;
 						try {

src/open-ai.ts

@@ -1,8 +1,8 @@
 import {OpenAI as openAI} from 'openai';
 import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 
 export class OpenAi extends LLMProvider {
 	client!: openAI;
@@ -67,13 +67,14 @@ export class OpenAi extends LLMProvider {
 			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
 
+			const tools = options.tools || this.ai.options.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
 				messages: history,
 				stream: !!options.stream,
 				max_tokens: options.max_tokens || this.ai.options.max_tokens || 4096,
 				temperature: options.temperature || this.ai.options.temperature || 0.7,
-				tools: (options.tools || this.ai.options.tools || []).map(t => ({
+				tools: tools.map(t => ({
 					type: 'function',
 					function: {
 						name: t.name,
@@ -89,7 +90,11 @@ export class OpenAi extends LLMProvider {
 
 			let resp: any, isFirstMessage = true;
 			do {
-				resp = await this.client.chat.completions.create(requestParams);
+				resp = await this.client.chat.completions.create(requestParams).catch(err => {
+					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
+					throw err;
+				});
+
 				if(options.stream) {
 					if(!isFirstMessage) options.stream({text: '\n\n'});
 					else isFirstMessage = false;
@@ -110,7 +115,8 @@ export class OpenAi extends LLMProvider {
 				if(toolCalls.length && !controller.signal.aborted) {
 					history.push(resp.choices[0].message);
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
-						const tool = options.tools?.find(findByProp('name', toolCall.function.name));
+						const tool = tools?.find(findByProp('name', toolCall.function.name));
+						if(options.stream) options.stream({tool: toolCall.function.name});
 						if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}'};
 						try {
 							const args = JSONAttemptParse(toolCall.function.arguments, {});

src/provider.ts

@@ -1,7 +1,6 @@
+import {AbortablePromise} from './ai.ts';
 import {LLMMessage, LLMOptions, LLMRequest} from './llm.ts';
 
-export type AbortablePromise<T> = Promise<T> & {abort: () => void};
-
 export abstract class LLMProvider {
 	abstract ask(message: string, options: LLMRequest): AbortablePromise<LLMMessage[]>;
 }

src/vision.ts

@@ -1,5 +1,5 @@
 import {createWorker} from 'tesseract.js';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 
 export class Vision {
 
@@ -8,18 +8,16 @@ export class Vision {
 	/**
 	 * Convert image to text using Optical Character Recognition
 	 * @param {string} path Path to image
-	 * @returns {{abort: Function, response: Promise<string | null>}} Abort function & Promise of extracted text
+	 * @returns {AbortablePromise<string | null>} Promise of extracted text with abort method
 	 */
-	ocr(path: string): {abort: () => void, response: Promise<string | null>} {
+	ocr(path: string): AbortablePromise<string | null> {
 		let worker: any;
-		return {
+		const p = new Promise<string | null>(async res => {
-			abort: () => { worker?.terminate(); },
+			worker = await createWorker(this.ai.options.tesseract?.model || 'eng', 2, {cachePath: this.ai.options.path});
-			response: new Promise(async res => {
-				worker = await createWorker('eng');
 			const {data} = await worker.recognize(path);
 			await worker.terminate();
 			res(data.text.trim() || null);
-			})
+		});
-		}
+		return Object.assign(p, {abort: () => worker?.terminate()});
 	}
 }

vite.config.ts

@@ -1,12 +1,19 @@
 import {defineConfig} from 'vite';
 import dts from 'vite-plugin-dts';
+import {resolve} from 'path';
 
 export default defineConfig({
 	build: {
 		lib: {
-			entry: './src/index.ts',
+			entry: {
+				index: './src/index.ts',
+				embedder: './src/embedder.ts',
+			},
 			name: 'utils',
-			fileName: (format) => (format === 'es' ? 'index.mjs' : 'index.js'),
+			fileName: (format, entryName) => {
+				if (entryName === 'embedder') return 'embedder.js';
+				return format === 'es' ? 'index.mjs' : 'index.js';
+			},
 		},
 		ssr: true,
 		emptyOutDir: true,