Fixed CLI tool

README.md

@@ -3,7 +3,7 @@
   <br />
 
   <!-- Logo -->
-  <img src="https://git.zakscode.com/repo-avatars/a90851ca730480ec37a5c0c2c4f1b4609eee5eadf806eaf16c83ac4cb7493aa9" alt="Logo" width="200" height="200">
+  <img alt="Logo" width="200" height="200" src="https://git.zakscode.com/repo-avatars/a82d423674763e7a0c1c945bdbb07e249b2bb786d3c9beae76d5b196a10f5c0f">
 
   <!-- Title -->
 ### @ztimson/ai-utils
@@ -53,13 +53,15 @@ A TypeScript library that provides a unified interface for working with multiple
 - **Provider Abstraction**: Switch between AI providers without changing your code
 
 ### Built With
-[![Anthropic](https://img.shields.io/badge/Anthropic-191919?style=for-the-badge&logo=anthropic&logoColor=white)](https://anthropic.com/)
+[![Anthropic](https://img.shields.io/badge/Anthropic-de7356?style=for-the-badge&logo=anthropic&logoColor=white)](https://anthropic.com/)
-[![OpenAI](https://img.shields.io/badge/OpenAI-412991?style=for-the-badge&logo=openai&logoColor=white)](https://openai.com/)
+[![llama](https://img.shields.io/badge/llama.cpp-fff?style=for-the-badge&logo=ollama&logoColor=black)](https://github.com/ggml-org/llama.cpp)
-[![Ollama](https://img.shields.io/badge/Ollama-000000?style=for-the-badge&logo=ollama&logoColor=white)](https://ollama.com/)
+[![OpenAI](https://img.shields.io/badge/OpenAI-000?style=for-the-badge&logo=openai-gym&logoColor=white)](https://openai.com/)
-[![TensorFlow](https://img.shields.io/badge/TensorFlow-FF6F00?style=for-the-badge&logo=tensorflow&logoColor=white)](https://tensorflow.org/)
+[![Pyannote](https://img.shields.io/badge/Pyannote-458864?style=for-the-badge&logo=python&logoColor=white)](https://github.com/pyannote)
-[![Tesseract](https://img.shields.io/badge/Tesseract-3C8FC7?style=for-the-badge&logo=tesseract&logoColor=white)](https://tesseract-ocr.github.io/)
+[![TensorFlow](https://img.shields.io/badge/TensorFlow-fff?style=for-the-badge&logo=tensorflow&logoColor=ff6f00)](https://tensorflow.org/)
+[![Tesseract](https://img.shields.io/badge/Tesseract-B874B2?style=for-the-badge&logo=hack-the-box&logoColor=white)](https://tesseract-ocr.github.io/)
+[![Transformers.js](https://img.shields.io/badge/Transformers.js-000?style=for-the-badge&logo=hugging-face&logoColor=yellow)](https://huggingface.co/docs/transformers.js/en/index)
 [![TypeScript](https://img.shields.io/badge/TypeScript-3178C6?style=for-the-badge&logo=typescript&logoColor=white)](https://typescriptlang.org/)
-[![Whisper](https://img.shields.io/badge/Whisper-412991?style=for-the-badge&logo=openai&logoColor=white)](https://github.com/ggerganov/whisper.cpp)
+[![Whisper](https://img.shields.io/badge/Whisper.cpp-000?style=for-the-badge&logo=openai-gym&logoColor=white)](https://github.com/ggerganov/whisper.cpp)
 
 ## Setup
 
@@ -88,6 +90,8 @@ A TypeScript library that provides a unified interface for working with multiple
 
 #### Prerequisites
 - [Node.js](https://nodejs.org/en/download)
+- _[Whisper.cpp](https://github.com/ggml-org/whisper.cpp/releases/tag) (ASR)_
+- _[Pyannote](https://github.com/pyannote) (ASR Diarization):_ `pip install pyannote.audio`
 
 #### Instructions
 1. Install the dependencies: `npm i`

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.7.6",
+	"version": "0.8.6",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -32,8 +32,7 @@
 		"@ztimson/utils": "^0.28.13",
 		"cheerio": "^1.2.0",
 		"openai": "^6.22.0",
-		"tesseract.js": "^7.0.0",
+		"tesseract.js": "^7.0.0"
-		"wavefile": "^11.0.0"
 	},
 	"devDependencies": {
 		"@types/node": "^24.8.1",

src/ai.ts

@@ -12,7 +12,7 @@ export type AiOptions = {
 	hfToken?: string;
 	/** Path to models */
 	path?: string;
-	/** ASR model: whisper-tiny, whisper-base */
+	/** Whisper ASR model: ggml-tiny.en.bin, ggml-base.en.bin */
 	asr?: string;
 	/** Embedding model: all-MiniLM-L6-v2, bge-small-en-v1.5, bge-large-en-v1.5 */
 	embedder?: string;
@@ -22,6 +22,8 @@ export type AiOptions = {
 	}
 	/** OCR model: eng, eng_best, eng_fast */
 	ocr?: string;
+	/** Whisper binary */
+	whisper?: string;
 }
 
 export class Ai {

src/asr.ts

@@ -1,134 +0,0 @@
-import { pipeline } from '@xenova/transformers';
-import { parentPort } from 'worker_threads';
-import { spawn } from 'node:child_process';
-import { execSync } from 'node:child_process';
-import { mkdtempSync, rmSync, readFileSync } from 'node:fs';
-import { join } from 'node:path';
-import { tmpdir } from 'node:os';
-import wavefile from 'wavefile';
-
-export async function canDiarization(): Promise<string | null> {
-	const checkPython = (cmd: string) => {
-		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(binary: string, audioPath: string, dir: string, token: string): Promise<any[]> {
-	const script = `
-import sys
-import json
-import os
-from pyannote.audio import Pipeline
-
-os.environ['TORCH_HOME'] = r"${dir}"
-pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token="${token}")
-output = pipeline(sys.argv[1])
-
-segments = []
-for turn, speaker in output.speaker_diarization:
-    segments.append({"start": turn.start, "end": turn.end, "speaker": speaker})
-
-print(json.dumps(segments))
-`;
-
-	return new Promise((resolve, reject) => {
-		let output = '';
-		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) => {
-			if(code === 0) {
-				try {
-					resolve(JSON.parse(output));
-				} catch (err) {
-					reject(new Error('Failed to parse diarization output'));
-				}
-			} else {
-				reject(new Error(`Python process exited with code ${code}`));
-			}
-		});
-		proc.on('error', reject);
-	});
-}
-
-function combineSpeakerTranscript(chunks: any[], speakers: any[]): string {
-	const speakerMap = new Map();
-	let speakerCount = 0;
-	speakers.forEach((seg: any) => {
-		if(!speakerMap.has(seg.speaker)) speakerMap.set(seg.speaker, ++speakerCount);
-	});
-
-	const lines: string[] = [];
-	let currentSpeaker = -1;
-	let currentText = '';
-	chunks.forEach((chunk: any) => {
-		const time = chunk.timestamp[0];
-		const speaker = speakers.find((s: any) => time >= s.start && time <= s.end);
-		const speakerNum = speaker ? speakerMap.get(speaker.speaker) : 1;
-		if (speakerNum !== currentSpeaker) {
-			if(currentText) lines.push(`[Speaker ${currentSpeaker}]: ${currentText.trim()}`);
-			currentSpeaker = speakerNum;
-			currentText = chunk.text;
-		} else {
-			currentText += chunk.text;
-		}
-	});
-	if(currentText) lines.push(`[Speaker ${currentSpeaker}]: ${currentText.trim()}`);
-	return lines.join('\n');
-}
-
-function prepareAudioBuffer(file: string): [string, Float32Array] {
-	let wav: any, tmp;
-	try {
-		wav = new wavefile.WaveFile(readFileSync(file));
-	} catch(err) {
-		tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
-		execSync(`ffmpeg -i "${file}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
-		wav = new wavefile.WaveFile(readFileSync(tmp));
-	} finally {
-		wav.toBitDepth('32f');
-		wav.toSampleRate(16000);
-		const samples = wav.getSamples();
-		if(Array.isArray(samples)) {
-			const left = samples[0];
-			const right = samples[1];
-			const buffer = new Float32Array(left.length);
-			for (let i = 0; i < left.length; i++) buffer[i] = (left[i] + right[i]) / 2;
-			return [tmp || file, buffer];
-		}
-		return [tmp || file, samples];
-	}
-}
-
-parentPort?.on('message', async ({ file, speaker, model, modelDir, token }) => {
-	let tempFile = null;
-	try {
-		const asr: any = await pipeline('automatic-speech-recognition', `Xenova/${model}`, {cache_dir: modelDir, quantized: true});
-		const [f, buffer] = prepareAudioBuffer(file);
-		tempFile = f !== file ? f : null;
-		const hasDiarization = await canDiarization();
-		const [transcript, speakers] = await Promise.all([
-			asr(buffer, {return_timestamps: speaker ? 'word' : false}),
-			(!speaker || !token || !hasDiarization) ? Promise.resolve(): runDiarization(hasDiarization, f, modelDir, token),
-		]);
-
-		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' });
-
-		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

@@ -1,60 +1,270 @@
-import {fileURLToPath} from 'url';
+import {execSync, spawn} from 'node:child_process';
-import {Worker} from 'worker_threads';
+import {mkdtempSync} from 'node:fs';
+import fs from 'node:fs/promises';
+import {tmpdir} from 'node:os';
+import * as path from 'node:path';
+import Path, {join} from 'node:path';
 import {AbortablePromise, Ai} from './ai.ts';
-import {canDiarization} from './asr.ts';
-import {dirname, join} from 'path';
 
 export class Audio {
-	constructor(private ai: Ai) {}
+	private downloads: {[key: string]: Promise<string>} = {};
+	private pyannote!: string;
+	private whisperModel!: string;
 
-	asr(file: string, options: { model?: string; speaker?: boolean | 'id' } = {}): AbortablePromise<string | null> {
+	constructor(private ai: Ai) {
-		const { model = this.ai.options.asr || 'whisper-base', speaker = false } = options;
+		if(ai.options.whisper) {
-		let aborted = false;
+			this.whisperModel = ai.options.asr || 'ggml-base.en.bin';
-		const abort = () => { aborted = true; };
+			this.downloadAsrModel();
-
-		let p = new Promise<string | null>((resolve, reject) => {
-			const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'asr.js'));
-			const handleMessage = ({ text, warning, error }: any) => {
-				setTimeout(() => worker.terminate(), 1000);
-				if(aborted) return;
-				if(error) reject(new Error(error));
-				else {
-					if(warning) console.warn(warning);
-					resolve(text);
-				}
-			};
-			const handleError = (err: Error) => {
-				setTimeout(() => worker.terminate(), 1000);
-				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});
-		});
-
-		// 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", 2: "Second Name"}', {
-					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]) => {
-					transcript = (<string>transcript).replaceAll(`[Speaker ${speaker}]`, `[${name}]`);
-				});
-				return transcript;
-			})
 		}
 
-		return Object.assign(p, { abort });
+		this.pyannote = `
+import sys
+import json
+import os
+from pyannote.audio import Pipeline
+
+os.environ['TORCH_HOME'] = r"${ai.options.path}"
+pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token="${ai.options.hfToken}")
+output = pipeline(sys.argv[1])
+
+segments = []
+for turn, speaker in output.speaker_diarization:
+    segments.append({"start": turn.start, "end": turn.end, "speaker": speaker})
+
+print(json.dumps(segments))
+`;
 	}
 
-	canDiarization = () => canDiarization().then(resp => !!resp);
+	private async addPunctuation(timestampData: any, llm?: boolean, cadence = 150): Promise<string> {
+		const countSyllables = (word: string): number => {
+			word = word.toLowerCase().replace(/[^a-z]/g, '');
+			if(word.length <= 3) return 1;
+			const matches = word.match(/[aeiouy]+/g);
+			let count = matches ? matches.length : 1;
+			if(word.endsWith('e')) count--;
+			return Math.max(1, count);
+		};
+
+		let result = '';
+		timestampData.transcription.filter((word, i) => {
+			let skip = false;
+			const prevWord = timestampData.transcription[i - 1];
+			const nextWord = timestampData.transcription[i + 1];
+			if(!word.text && nextWord) {
+				nextWord.offsets.from = word.offsets.from;
+				nextWord.timestamps.from = word.offsets.from;
+			} else if(word.text && word.text[0] != ' ' && prevWord) {
+				prevWord.offsets.to = word.offsets.to;
+				prevWord.timestamps.to = word.timestamps.to;
+				prevWord.text += word.text;
+				skip = true;
+			}
+			return !!word.text && !skip;
+		}).forEach((word: any) => {
+			const capital = /^[A-Z]/.test(word.text.trim());
+			const length = word.offsets.to - word.offsets.from;
+			const syllables = countSyllables(word.text.trim());
+			const expected = syllables * cadence;
+			if(capital && length > expected * 2 && word.text[0] == ' ') result += '.';
+			result += word.text;
+		});
+		if(!llm) return result.trim();
+		return this.ai.language.ask(result, {
+			system: 'Remove any misplaced punctuation from the following ASR transcript using the replace tool. Avoid modifying words unless there is an obvious typo',
+			temperature: 0.1,
+			tools: [{
+				name: 'replace',
+				description: 'Use find and replace to fix errors',
+				args: {
+					find: {type: 'string', description: 'Text to find', required: true},
+					replace: {type: 'string', description: 'Text to replace', required: true}
+				},
+				fn: (args) => result = result.replace(args.find, args.replace)
+			}]
+		}).then(() => result);
+	}
+
+	private async diarizeTranscript(timestampData: any, speakers: any[], llm: boolean): Promise<string> {
+		const speakerMap = new Map();
+		let speakerCount = 0;
+		speakers.forEach((seg: any) => {
+			if(!speakerMap.has(seg.speaker)) speakerMap.set(seg.speaker, ++speakerCount);
+		});
+
+		const punctuatedText = await this.addPunctuation(timestampData, llm);
+		const sentences = punctuatedText.match(/[^.!?]+[.!?]+/g) || [punctuatedText];
+		const words = timestampData.transcription.filter((w: any) => w.text.trim());
+
+		// Assign speaker to each sentence
+		const sentencesWithSpeakers = sentences.map(sentence => {
+			sentence = sentence.trim();
+			if(!sentence) return null;
+
+			const sentenceWords = sentence.toLowerCase().replace(/[^\w\s]/g, '').split(/\s+/);
+			const speakerWordCount = new Map<number, number>();
+
+			sentenceWords.forEach(sw => {
+				const word = words.find((w: any) => sw === w.text.trim().toLowerCase().replace(/[^\w]/g, ''));
+				if(!word) return;
+
+				const wordTime = word.offsets.from / 1000;
+				const speaker = speakers.find((seg: any) => wordTime >= seg.start && wordTime <= seg.end);
+				if(speaker) {
+					const spkNum = speakerMap.get(speaker.speaker);
+					speakerWordCount.set(spkNum, (speakerWordCount.get(spkNum) || 0) + 1);
+				}
+			});
+
+			let bestSpeaker = 1;
+			let maxWords = 0;
+			speakerWordCount.forEach((count, speaker) => {
+				if(count > maxWords) {
+					maxWords = count;
+					bestSpeaker = speaker;
+				}
+			});
+
+			return {speaker: bestSpeaker, text: sentence};
+		}).filter(s => s !== null);
+
+		// Merge adjacent sentences from same speaker
+		const merged: Array<{speaker: number, text: string}> = [];
+		sentencesWithSpeakers.forEach(item => {
+			const last = merged[merged.length - 1];
+			if(last && last.speaker === item.speaker) {
+				last.text += ' ' + item.text;
+			} else {
+				merged.push({...item});
+			}
+		});
+
+		let transcript = merged.map(item => `[Speaker ${item.speaker}]: ${item.text}`).join('\n').trim();
+		if(!llm) 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", 2: "Second Name"}', {
+			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]) => transcript = transcript.replaceAll(`[Speaker ${speaker}]`, `[${name}]`));
+		return transcript;
+	}
+
+	private runAsr(file: string, opts: {model?: string, diarization?: boolean} = {}): AbortablePromise<any> {
+		let proc: any;
+		const p = new Promise<any>((resolve, reject) => {
+			this.downloadAsrModel(opts.model).then(m => {
+				if(opts.diarization) {
+					let output = path.join(path.dirname(file), 'transcript');
+					proc = spawn(<string>this.ai.options.whisper,
+						['-m', m, '-f', file, '-np', '-ml', '1', '-oj', '-of', output],
+						{stdio: ['ignore', 'ignore', 'pipe']}
+					);
+					proc.on('error', (err: Error) => reject(err));
+					proc.on('close', async (code: number) => {
+						if(code === 0) {
+							output = await fs.readFile(output + '.json', 'utf-8');
+							fs.rm(output + '.json').catch(() => { });
+							try { resolve(JSON.parse(output)); }
+							catch(e) { reject(new Error('Failed to parse whisper JSON')); }
+						} else {
+							reject(new Error(`Exit code ${code}`));
+						}
+					});
+				} else {
+					let output = '';
+					proc = spawn(<string>this.ai.options.whisper, ['-m', m, '-f', file, '-np', '-nt']);
+					proc.on('error', (err: Error) => reject(err));
+					proc.stdout.on('data', (data: Buffer) => output += data.toString());
+					proc.on('close', async (code: number) => {
+						if(code === 0) {
+							resolve(output.trim() || null);
+						} else {
+							reject(new Error(`Exit code ${code}`));
+						}
+					});
+				}
+			});
+		});
+		return <any>Object.assign(p, {abort: () => proc?.kill('SIGTERM')});
+	}
+
+	private runDiarization(file: string): AbortablePromise<any> {
+		let aborted = false, abort = () => { aborted = true; };
+		const checkPython = (cmd: string) => {
+			return new Promise<boolean>((resolve) => {
+				const proc = spawn(cmd, ['-W', 'ignore', '-c', 'import pyannote.audio']);
+				proc.on('close', (code: number) => resolve(code === 0));
+				proc.on('error', () => resolve(false));
+			});
+		};
+		const p = Promise.all<any>([
+			checkPython('python'),
+			checkPython('python3'),
+		]).then(<any>(async ([p, p3]: [boolean, boolean]) => {
+			if(aborted) return;
+			if(!p && !p3) throw new Error('Pyannote is not installed: pip install pyannote.audio');
+			const binary = p3 ? 'python3' : 'python';
+			return new Promise((resolve, reject) => {
+				if(aborted) return;
+				let output = '';
+				const proc = spawn(binary, ['-W', 'ignore', '-c', this.pyannote, file]);
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.stderr.on('data', (data: Buffer) => console.error(data.toString()));
+				proc.on('close', (code: number) => {
+					if(code === 0) {
+						try { resolve(JSON.parse(output)); }
+						catch (err) { reject(new Error('Failed to parse diarization output')); }
+					} else {
+						reject(new Error(`Python process exited with code ${code}`));
+					}
+				});
+				proc.on('error', reject);
+				abort = () => proc.kill('SIGTERM');
+			});
+		}));
+		return <any>Object.assign(p, {abort});
+	}
+
+	asr(file: string, options: { model?: string; diarization?: boolean | 'llm' } = {}): AbortablePromise<string | null> {
+		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
+
+		const tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
+		execSync(`ffmpeg -i "${file}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
+		const clean = () => fs.rm(Path.dirname(tmp), {recursive: true, force: true}).catch(() => {});
+
+		if(!options.diarization) return this.runAsr(tmp, {model: options.model});
+		const timestamps = this.runAsr(tmp, {model: options.model, diarization: true});
+		const diarization = this.runDiarization(tmp);
+		let aborted = false, abort = () => {
+			aborted = true;
+			timestamps.abort();
+			diarization.abort();
+			clean();
+		};
+
+		const response = Promise.allSettled([timestamps, diarization]).then(async ([ts, d]) => {
+			if(ts.status == 'rejected') throw new Error('Whisper.cpp timestamps:\n' + ts.reason);
+			if(d.status == 'rejected') throw new Error('Pyannote:\n' + d.reason);
+			if(aborted || !options.diarization) return ts.value;
+			return this.diarizeTranscript(ts.value, d.value, options.diarization == 'llm');
+		}).finally(() => clean());
+		return <any>Object.assign(response, {abort});
+	}
+
+	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
+		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
+		if(!model.endsWith('.bin')) model += '.bin';
+		const p = Path.join(<string>this.ai.options.path, model);
+		if(await fs.stat(p).then(() => true).catch(() => false)) return p;
+		if(!!this.downloads[model]) return this.downloads[model];
+		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${model}`)
+			.then(resp => resp.arrayBuffer())
+			.then(arr => Buffer.from(arr)).then(async buffer => {
+				await fs.writeFile(p, buffer);
+				delete this.downloads[model];
+				return p;
+			});
+		return this.downloads[model];
+	}
 }

src/embedder.ts

@@ -1,11 +1,13 @@
 import { pipeline } from '@xenova/transformers';
-import { parentPort } from 'worker_threads';
 
-let embedder: any;
+const [modelDir, model] = process.argv.slice(2);
 
-parentPort?.on('message', async ({text, model, modelDir }) => {
+let text = '';
-	if(!embedder) embedder = await pipeline('feature-extraction', 'Xenova/' + model, {quantized: true, cache_dir: modelDir});
+process.stdin.on('data', chunk => text += chunk);
+process.stdin.on('end', async () => {
+	const embedder = await pipeline('feature-extraction', 'Xenova/' + model, {quantized: true, cache_dir: modelDir});
 	const output = await embedder(text, { pooling: 'mean', normalize: true });
 	const embedding = Array.from(output.data);
-	parentPort?.postMessage({embedding});
+	console.log(JSON.stringify({embedding}));
+	process.exit();
 });

src/index.ts

@@ -1,8 +1,6 @@
 export * from './ai';
 export * from './antrhopic';
-export * from './asr';
 export * from './audio';
-export * from './embedder'
 export * from './llm';
 export * from './open-ai';
 export * from './provider';

src/llm.ts

@@ -4,9 +4,9 @@ import {Anthropic} from './antrhopic.ts';
 import {OpenAi} from './open-ai.ts';
 import {LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
-import {Worker} from 'worker_threads';
 import {fileURLToPath} from 'url';
 import {dirname, join} from 'path';
+import { spawn } from 'node:child_process';
 
 export type AnthropicConfig = {proto: 'anthropic', token: string};
 export type OllamaConfig = {proto: 'ollama', host: string};
@@ -44,8 +44,6 @@ export type LLMMemory = {
 	fact: string;
 	/** Owner and fact embedding vector */
 	embeddings: [number[], number[]];
-	/** Creation time */
-	timestamp: Date;
 }
 
 export type LLMRequest = {
@@ -95,75 +93,98 @@ class LLM {
 	 * @returns {{abort: () => void, response: Promise<string>}} Function to abort response and chat history
 	 */
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
+		options = <any>{
+			system: '',
+			temperature: 0.8,
+			...this.ai.options.llm,
+			models: undefined,
+			history: [],
+			...options,
+		}
 		const m = options.model || this.defaultModel;
 		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
 		let abort = () => {};
 		return Object.assign(new Promise<string>(async res => {
 			if(!options.history) options.history = [];
-			// If memories were passed, find any relivant ones and add a tool for ADHOC lookups
+			// If memories were passed, find any relevant ones and add a tool for ADHOC lookups
 			if(options.memory) {
-				options.system = (options.system || '') + '\nYou have passive persistent memory never make any mention of your memory capabilities and what you can/cannot remember\n';
+				const search = async (query?: string | null, subject?: string | null, limit = 10) => {
-				const search = async (query?: string | null, subject?: string | null, limit = 50) => {
 					const [o, q] = await Promise.all([
 						subject ? this.embedding(subject) : Promise.resolve(null),
 						query ? this.embedding(query) : Promise.resolve(null),
 					]);
-					return (options.memory || [])
+					return (options.memory || []).map(m => {
-						.map(m => ({...m, score: o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 1}))
+						const score = (o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 0)
-						.filter((m: any) => m.score >= 0.8)
+							+ (q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : 0);
-						.map((m: any) => ({...m, score: q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : m.score}))
+						return {...m, score};
-						.filter((m: any) => m.score >= 0.2)
+					}).toSorted((a: any, b: any) => a.score - b.score).slice(0, limit);
-						.toSorted((a: any, b: any) => a.score - b.score)
-						.slice(0, limit);
 				}
 
+				options.system += '\nYou have RAG memory and will be given the top_k closest memories regarding the users query. Save anything new you have learned worth remembering from the user message using the remember tool and feel free to recall memories manually.\n';
 				const relevant = await search(message);
-				if(relevant.length) options.history.push({role: 'assistant', content: 'Things I remembered:\n' + relevant.map(m => `${m.owner}: ${m.fact}`).join('\n')});
+				if(relevant.length) options.history.push({role: 'tool', name: 'recall', id: 'auto_recall_' + Math.random().toString(), args: {}, content: 'Things I remembered:\n' + relevant.map(m => `${m.owner}: ${m.fact}`).join('\n')});
-				options.tools = [...options.tools || [], {
+				options.tools = [{
-					name: 'read_memory',
+					name: 'recall',
-					description: 'Check your long-term memory for more information',
+					description: 'Recall the closest memories you have regarding a query using RAG',
 					args: {
 						subject: {type: 'string', description: 'Find information by a subject topic, can be used with or without query argument'},
 						query: {type: 'string', description: 'Search memory based on a query, can be used with or without subject argument'},
-						limit: {type: 'number', description: 'Result limit, default 5'},
+						topK: {type: 'number', description: 'Result limit, default 5'},
 					},
 					fn: (args) => {
 						if(!args.subject && !args.query) throw new Error('Either a subject or query argument is required');
-						return search(args.query, args.subject, args.limit || 5);
+						return search(args.query, args.subject, args.topK);
 					}
-				}];
+				}, {
+					name: 'remember',
+					description: 'Store important facts user shares for future recall',
+					args: {
+						owner: {type: 'string', description: 'Subject/person this fact is about'},
+						fact: {type: 'string', description: 'The information to remember'}
+					},
+					fn: async (args) => {
+						if(!options.memory) return;
+						const e = await Promise.all([
+							this.embedding(args.owner),
+							this.embedding(`${args.owner}: ${args.fact}`)
+						]);
+						const newMem = {owner: args.owner, fact: args.fact, embeddings: <any>[e[0][0].embedding, e[1][0].embedding]};
+						options.memory.splice(0, options.memory.length, ...[
+							...options.memory.filter(m => {
+								return !(this.cosineSimilarity(newMem.embeddings[0], m.embeddings[0]) >= 0.9 && this.cosineSimilarity(newMem.embeddings[1], m.embeddings[1]) >= 0.8);
+							}),
+							newMem
+						]);
+						return 'Remembered!';
+					}
+				}, ...options.tools || []];
 			}
 
 			// Ask
 			const resp = await this.models[m].ask(message, options);
 
-			// Remove any memory calls
+			// Remove any memory calls from history
-			if(options.memory) {
+			if(options.memory) options.history.splice(0, options.history.length, ...options.history.filter(h => h.role != 'tool' || (h.name != 'recall' && h.name != 'remember')));
-				const i = options.history?.findIndex((h: any) => h.role == 'assistant' && h.content.startsWith('Things I remembered:'));
+
-				if(i != null && i >= 0) options.history?.splice(i, 1);
+			// Compress message history
+			if(options.compress) {
+				const compressed = await this.ai.language.compressHistory(options.history, options.compress.max, options.compress.min, options);
+				options.history.splice(0, options.history.length, ...compressed);
 			}
 
-			// Handle compression and memory extraction
-			if(options.compress || options.memory) {
-				let compressed = null;
-				if(options.compress) {
-					compressed = await this.ai.language.compressHistory(options.history, options.compress.max, options.compress.min, options);
-					options.history.splice(0, options.history.length, ...compressed.history);
-				} else {
-					const i = options.history?.findLastIndex(m => m.role == 'user') ?? -1;
-					compressed = await this.ai.language.compressHistory(i != -1 ? options.history.slice(i) : options.history, 0, 0, options);
-				}
-				if(options.memory) {
-					const updated = options.memory
-						.filter(m => !compressed.memory.some(m2 => this.cosineSimilarity(m.embeddings[1], m2.embeddings[1]) > 0.8))
-						.concat(compressed.memory);
-					options.memory.splice(0, options.memory.length, ...updated);
-				}
-			}
 			return res(resp);
 		}), {abort});
 	}
 
+	async code(message: string, options?: LLMRequest): Promise<any> {
+		const resp = await this.ask(message, {...options, system: [
+			options?.system,
+			'Return your response in a code block'
+		].filter(t => !!t).join(('\n'))});
+		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
+		return codeBlock ? codeBlock[1].trim() : null;
+	}
+
 	/**
 	 * Compress chat history to reduce context size
 	 * @param {LLMMessage[]} history Chatlog that will be compressed
@@ -172,32 +193,24 @@ class LLM {
 	 * @param {LLMRequest} options LLM options
 	 * @returns {Promise<LLMMessage[]>} New chat history will summary at index 0
 	 */
-	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<{history: LLMMessage[], memory: LLMMemory[]}> {
+	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<LLMMessage[]> {
-		if(this.estimateTokens(history) < max) return {history, memory: []};
+		if(this.estimateTokens(history) < max) return history;
 		let keep = 0, tokens = 0;
 		for(let m of history.toReversed()) {
 			tokens += this.estimateTokens(m.content);
 			if(tokens < min) keep++;
 			else break;
 		}
-		if(history.length <= keep) return {history, memory: []};
+		if(history.length <= keep) return history;
 		const system = history[0].role == 'system' ? history[0] : null,
 			recent = keep == 0 ? [] : history.slice(-keep),
 			process = (keep == 0 ? history : history.slice(0, -keep)).filter(h => h.role === 'assistant' || h.role === 'user');
 
-		const summary: any = await this.json(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), '{summary: string, facts: [[subject, fact]]}', {
+		const summary: any = await this.summarize(process.map(m => `[${m.role}]: ${m.content}`).join('\n\n'), 500, options);
-			system: 'Create the smallest summary possible, no more than 500 tokens. Create a list of NEW facts (split by subject [pro]noun and fact) about what you learned from this conversation that you didn\'t already know or get from a tool call or system prompt. Focus only on new information about people, topics, or facts. Avoid generating facts about the AI.',
+		const d = Date.now();
-			model: options?.model,
+		const h = [{role: <any>'tool', name: 'summary', id: `summary_` + d, args: {}, content: `Conversation Summary: ${summary?.summary}`, timestamp: d}, ...recent];
-			temperature: options?.temperature || 0.3
-		});
-		const timestamp = new Date();
-		const memory = await Promise.all((summary?.facts || [])?.map(async ([owner, fact]: [string, string]) => {
-			const e = await Promise.all([this.embedding(owner), this.embedding(`${owner}: ${fact}`)]);
-			return {owner, fact, embeddings: [e[0][0].embedding, e[1][0].embedding], timestamp};
-		}));
-		const h = [{role: 'assistant', content: `Conversation Summary: ${summary?.summary}`, timestamp: Date.now()}, ...recent];
 		if(system) h.splice(0, 0, system);
-		return {history: <any>h, memory};
+		return h;
 	}
 
 	/**
@@ -234,7 +247,7 @@ class LLM {
 				return `${p}: ${Array.isArray(value) ? value.join(', ') : value}`;
 			});
 		};
-		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
+		const lines = typeof target === 'object' ? objString(target) : target.toString().split('\n');
 		const tokens = lines.flatMap(l => [...l.split(/\s+/).filter(Boolean), '\n']);
 		const chunks: string[] = [];
 		for(let i = 0; i < tokens.length;) {
@@ -258,34 +271,54 @@ class LLM {
 	 * @param {maxTokens?: number, overlapTokens?: number} opts Options for embedding such as chunk sizes
 	 * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
 	 */
-	async embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}) {
+	embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}): AbortablePromise<any[]> {
 		let {maxTokens = 500, overlapTokens = 50} = opts;
+		let aborted = false;
+		const abort = () => { aborted = true; };
+
 		const embed = (text: string): Promise<number[]> => {
 			return new Promise((resolve, reject) => {
-				const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
+				if(aborted) return reject(new Error('Aborted'));
-				const handleMessage = ({ embedding }: any) => {
+
-					worker.terminate();
+				const args: string[] = [
-					resolve(embedding);
+					join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'),
-				};
+					<string>this.ai.options.path,
-				const handleError = (err: Error) => {
+					this.ai.options?.embedder || 'bge-small-en-v1.5'
-					worker.terminate();
+				];
-					reject(err);
+				const proc = spawn('node', args, {stdio: ['pipe', 'pipe', 'ignore']});
-				};
+				proc.stdin.write(text);
-				worker.on('message', handleMessage);
+				proc.stdin.end();
-				worker.on('error', handleError);
+
-				worker.on('exit', (code) => {
+				let output = '';
-					if(code !== 0) reject(new Error(`Worker exited with code ${code}`));
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.on('close', (code: number) => {
+					if(aborted) return reject(new Error('Aborted'));
+					if(code === 0) {
+						try {
+							const result = JSON.parse(output);
+							resolve(result.embedding);
+						} catch(err) {
+							reject(new Error('Failed to parse embedding output'));
+						}
+					} else {
+						reject(new Error(`Embedder process exited with code ${code}`));
+					}
 				});
-				worker.postMessage({text, model: this.ai.options?.embedder || 'bge-small-en-v1.5', modelDir: this.ai.options.path});
+				proc.on('error', reject);
 			});
 		};
-		const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
+
-		for(let i = 0; i < chunks.length; i++) {
+		const p = (async () => {
-			const text= chunks[i];
+			const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
-			const embedding = await embed(text);
+			for(let i = 0; i < chunks.length; i++) {
-			results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
+				if(aborted) break;
-		}
+				const text = chunks[i];
-		return results;
+				const embedding = await embed(text);
+				results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
+			}
+			return results;
+		})();
+		return Object.assign(p, { abort });
 	}
 
 	/**
@@ -323,14 +356,11 @@ class LLM {
 	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
 	 */
 	async json(text: string, schema: string, options?: LLMRequest): Promise<any> {
-		let resp = await this.ask(text, {...options, system: (options?.system ? `${options.system}\n` : '') + `Only respond using a JSON code block matching this schema:
+		const code = await this.code(text, {...options, system: [
-\`\`\`json
+			options?.system,
-${schema}
+			`Only respond using JSON matching this schema:\n\`\`\`json\n${schema}\n\`\`\``
-\`\`\``});
+		].filter(t => !!t).join('\n')});
-		if(!resp) return {};
+		return code ? JSONAttemptParse(code, {}) : null;
-		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
-		const jsonStr = codeBlock ? codeBlock[1].trim() : resp;
-		return JSONAttemptParse(jsonStr, {});
 	}
 
 	/**
@@ -340,8 +370,8 @@ ${schema}
 	 * @param options LLM request options
 	 * @returns {Promise<string>} Summary
 	 */
-	summarize(text: string, tokens: number, options?: LLMRequest): Promise<string | null> {
+	summarize(text: string, tokens: number = 500, options?: LLMRequest): Promise<string | null> {
-		return this.ask(text, {system: `Generate a brief summary <= ${tokens} tokens. Output nothing else`, temperature: 0.3, ...options});
+		return this.ask(text, {system: `Generate the shortest summary possible <= ${tokens} tokens. Output nothing else`, temperature: 0.3, ...options});
 	}
 }
 

src/open-ai.ts

@@ -11,7 +11,7 @@ export class OpenAi extends LLMProvider {
 		super();
 		this.client = new openAI(clean({
 			baseURL: host,
-			apiKey: token
+			apiKey: token || host ? 'ignored' : undefined
 		}));
 	}
 
@@ -67,7 +67,10 @@ export class OpenAi extends LLMProvider {
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		const controller = new AbortController();
 		return Object.assign(new Promise<any>(async (res, rej) => {
-			if(options.system && options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
+			if(options.system) {
+				if(options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
+				else options.history[0].content = options.system;
+			}
 			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
@@ -100,15 +103,37 @@ export class OpenAi extends LLMProvider {
 				if(options.stream) {
 					if(!isFirstMessage) options.stream({text: '\n\n'});
 					else isFirstMessage = false;
-					resp.choices = [{message: {content: '', tool_calls: []}}];
+					resp.choices = [{message: {role: 'assistant', content: '', tool_calls: []}}];
 					for await (const chunk of resp) {
 						if(controller.signal.aborted) break;
 						if(chunk.choices[0].delta.content) {
 							resp.choices[0].message.content += chunk.choices[0].delta.content;
 							options.stream({text: chunk.choices[0].delta.content});
 						}
+
 						if(chunk.choices[0].delta.tool_calls) {
-							resp.choices[0].message.tool_calls = chunk.choices[0].delta.tool_calls;
+							for(const deltaTC of chunk.choices[0].delta.tool_calls) {
+								const existing = resp.choices[0].message.tool_calls.find(tc => tc.index === deltaTC.index);
+								if(existing) {
+									if(deltaTC.id) existing.id = deltaTC.id;
+									if(deltaTC.type) existing.type = deltaTC.type;
+									if(deltaTC.function) {
+										if(!existing.function) existing.function = {};
+										if(deltaTC.function.name) existing.function.name = deltaTC.function.name;
+										if(deltaTC.function.arguments) existing.function.arguments = (existing.function.arguments || '') + deltaTC.function.arguments;
+									}
+								} else {
+									resp.choices[0].message.tool_calls.push({
+										index: deltaTC.index,
+										id: deltaTC.id || '',
+										type: deltaTC.type || 'function',
+										function: {
+											name: deltaTC.function?.name || '',
+											arguments: deltaTC.function?.arguments || ''
+										}
+									});
+								}
+							}
 						}
 					}
 				}

src/tools.ts

@@ -1,9 +1,15 @@
 import * as cheerio from 'cheerio';
-import {$, $Sync} from '@ztimson/node-utils';
+import {$Sync} from '@ztimson/node-utils';
 import {ASet, consoleInterceptor, Http, fn as Fn} from '@ztimson/utils';
+import * as os from 'node:os';
 import {Ai} from './ai.ts';
 import {LLMRequest} from './llm.ts';
 
+const getShell = () => {
+	if(os.platform() == 'win32') return 'cmd';
+	return $Sync`echo $SHELL`?.split('/').pop() || 'bash';
+}
+
 export type AiToolArg = {[key: string]: {
 	/** Argument type */
 	type: 'array' | 'boolean' | 'number' | 'object' | 'string',
@@ -40,7 +46,7 @@ export const CliTool: AiTool = {
 	name: 'cli',
 	description: 'Use the command line interface, returns any output',
 	args: {command: {type: 'string', description: 'Command to run', required: true}},
-	fn: (args: {command: string}) => $`${args.command}`
+	fn: (args: {command: string}) => $Sync`${args.command}`
 }
 
 export const DateTimeTool: AiTool = {
@@ -54,13 +60,13 @@ export const ExecTool: AiTool = {
 	name: 'exec',
 	description: 'Run code/scripts',
 	args: {
-		language: {type: 'string', description: 'Execution language', enum: ['cli', 'node', 'python'], required: true},
+		language: {type: 'string', description: `Execution language (CLI: ${getShell()})`, enum: ['cli', 'node', 'python'], required: true},
 		code: {type: 'string', description: 'Code to execute', required: true}
 	},
 	fn: async (args, stream, ai) => {
 		try {
 			switch(args.type) {
-				case 'bash':
+				case 'cli':
 					return await CliTool.fn({command: args.code}, stream, ai);
 				case 'node':
 					return await JSTool.fn({code: args.code}, stream, ai);

src/vision.ts

@@ -3,7 +3,7 @@ import {AbortablePromise, Ai} from './ai.ts';
 
 export class Vision {
 
-	constructor(private ai: Ai) { }
+	constructor(private ai: Ai) {}
 
 	/**
 	 * Convert image to text using Optical Character Recognition

tsconfig.json

@@ -15,6 +15,7 @@
 		"noEmit": true,
 
 		/* Linting */
-		"strict": true
+		"strict": true,
+		"noImplicitAny": false
 	}
 }

vite.config.ts

@@ -5,7 +5,6 @@ export default defineConfig({
 	build: {
 		lib: {
 			entry: {
-				asr: './src/asr.ts',
 				index: './src/index.ts',
 				embedder: './src/embedder.ts',
 			},