Better ASR IDing

README.md

@@ -75,6 +75,7 @@ A TypeScript library that provides a unified interface for working with multiple
 
 #### Instructions
 1. Install the package: `npm i @ztimson/ai-utils`
+2. For speaker diarization: `pip install pyannote.audio`
 
 </details>
 
@@ -90,8 +91,9 @@ A TypeScript library that provides a unified interface for working with multiple
 
 #### Instructions
 1. Install the dependencies: `npm i`
-2. Build library: `npm build`
-3. Run unit tests: `npm test`
+2. For speaker diarization: `pip install pyannote.audio`
+3. Build library: `npm build`
+4. Run unit tests: `npm test`
 
 </details>
 

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.4.0",
+	"version": "0.7.5",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -25,14 +25,15 @@
 		"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/utils": "^0.27.9",
+		"@ztimson/node-utils": "^1.0.7",
+		"@ztimson/utils": "^0.28.13",
 		"cheerio": "^1.2.0",
-		"openai": "^6.6.0",
-		"tesseract.js": "^6.0.1"
+		"openai": "^6.22.0",
+		"tesseract.js": "^7.0.0",
+		"wavefile": "^11.0.0"
 	},
 	"devDependencies": {
 		"@types/node": "^24.8.1",

src/ai.ts

@@ -1,29 +1,27 @@
 import * as os from 'node:os';
-import {LLM, AnthropicConfig, OllamaConfig, OpenAiConfig, LLMRequest} from './llm';
+import LLM, {AnthropicConfig, OllamaConfig, OpenAiConfig, LLMRequest} from './llm';
 import { Audio } from './audio.ts';
 import {Vision} from './vision.ts';
 
-export type AbortablePromise<T> = Promise<T> & {abort: () => any};
+export type AbortablePromise<T> = Promise<T> & {
+	abort: () => any
+};
 
 export type AiOptions = {
+	/** Token to pull models from hugging face */
+	hfToken?: string;
 	/** Path to models */
 	path?: string;
+	/** ASR model: whisper-tiny, whisper-base */
+	asr?: string;
+	/** Embedding model: all-MiniLM-L6-v2, bge-small-en-v1.5, bge-large-en-v1.5 */
+	embedder?: string;
 	/** Large language models, first is default */
 	llm?: Omit<LLMRequest, 'model'> & {
 		models: {[model: string]: AnthropicConfig | OllamaConfig | OpenAiConfig};
 	}
-	/** 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;
-	}
+	/** OCR model: eng, eng_best, eng_fast */
+	ocr?: string;
 }
 
 export class Ai {

src/antrhopic.ts

@@ -13,25 +13,25 @@ export class Anthropic extends LLMProvider {
 	}
 
 	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
-			const orgI = i;
-			if(typeof history[orgI].content != 'string') {
-				if(history[orgI].role == 'assistant') {
-					history[orgI].content.filter((c: any) => c.type =='tool_use').forEach((c: any) => {
-						i++;
-						history.splice(i, 0, {role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: Date.now()});
-					});
-				} else if(history[orgI].role == 'user') {
-					history[orgI].content.filter((c: any) => c.type =='tool_result').forEach((c: any) => {
-						const h = history.find((h: any) => h.id == c.tool_use_id);
-						h[c.is_error ? 'error' : 'content'] = c.content;
-					});
-				}
-				history[orgI].content = history[orgI].content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
+		const timestamp = Date.now();
+		const messages: LLMMessage[] = [];
+		for(let h of history) {
+			if(typeof h.content == 'string') {
+				messages.push(<any>{timestamp, ...h});
+			} else {
+				const textContent = h.content?.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
+				if(textContent) messages.push({timestamp, role: h.role, content: textContent});
+				h.content.forEach((c: any) => {
+					if(c.type == 'tool_use') {
+						messages.push({timestamp, role: 'tool', id: c.id, name: c.name, args: c.input, content: undefined});
+					} else if(c.type == 'tool_result') {
+						const m: any = messages.findLast(m => (<any>m).id == c.tool_use_id);
+						if(m) m[c.is_error ? 'error' : 'content'] = c.content;
+					}
+				});
 			}
-			if(!history[orgI].timestamp) history[orgI].timestamp = Date.now();
 		}
-		return history.filter(h => !!h.content);
+		return messages;
 	}
 
 	private fromStandard(history: LLMMessage[]): any[] {
@@ -48,13 +48,10 @@ export class Anthropic extends LLMProvider {
 		return history.map(({timestamp, ...h}) => h);
 	}
 
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
+	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		const controller = new AbortController();
-		const response = new Promise<any>(async (res, rej) => {
-			let history = [...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);
-			history = this.fromStandard(<any>history);
-
+		return Object.assign(new Promise<any>(async (res) => {
+			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 = {
 				model: options.model || this.model,
@@ -76,7 +73,6 @@ export class Anthropic extends LLMProvider {
 			};
 
 			let resp: any, isFirstMessage = true;
-			const assistantMessages: string[] = [];
 			do {
 				resp = await this.client.messages.create(requestParams).catch(err => {
 					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
@@ -122,7 +118,7 @@ export class Anthropic extends LLMProvider {
 						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);
+							const result = await tool.fn(toolCall.input, options?.stream, this.ai);
 							return {type: 'tool_result', tool_use_id: toolCall.id, content: JSONSanitize(result)};
 						} catch (err: any) {
 							return {type: 'tool_result', tool_use_id: toolCall.id, is_error: true, content: err?.message || err?.toString() || 'Unknown'};
@@ -132,11 +128,12 @@ export class Anthropic extends LLMProvider {
 					requestParams.messages = history;
 				}
 			} while (!controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
+			history.push({role: 'assistant', content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')});
+			history = this.toStandard(history);
 
 			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {role: 'assistant', content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')}]));
-		});
-
-		return Object.assign(response, {abort: () => controller.abort()});
+			if(options.history) options.history.splice(0, options.history.length, ...history);
+			res(history.at(-1)?.content);
+		}), {abort: () => controller.abort()});
 	}
 }

src/asr.ts

@@ -0,0 +1,139 @@
+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';
+
+let whisperPipeline: any;
+
+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 "${tmp}"`, { stdio: 'ignore' });
+		wav = new wavefile.WaveFile(readFileSync(tmp));
+	} finally {
+		wav.toBitDepth('32f');
+		wav.toSampleRate(16000);
+		const samples = wav.getSamples();
+		if(Array.isArray(samples)) {
+			const left = samples[0];
+			const right = samples[1];
+			const buffer = new Float32Array(left.length);
+			for (let i = 0; i < left.length; i++) buffer[i] = (left[i] + right[i]) / 2;
+			return [tmp || file, buffer];
+		}
+		return [tmp || file, samples];
+	}
+}
+
+parentPort?.on('message', async ({ file, speaker, model, modelDir, token }) => {
+	try {
+		if(!whisperPipeline) whisperPipeline = await pipeline('automatic-speech-recognition', `Xenova/${model}`, {cache_dir: modelDir, quantized: true});
+
+		// Prepare audio file
+		const [f, buffer] = prepareAudioBuffer(file);
+
+		// Fetch transcript and speakers
+		const hasDiarization = await canDiarization();
+		const [transcript, speakers] = await Promise.all([
+			whisperPipeline(buffer, {return_timestamps: speaker ? 'word' : false}),
+			(!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 });
+	}
+});

src/audio.ts

@@ -1,50 +1,60 @@
-import {spawn} from 'node:child_process';
-import fs from 'node:fs/promises';
-import Path from 'node:path';
+import {fileURLToPath} from 'url';
+import {Worker} from 'worker_threads';
 import {AbortablePromise, Ai} from './ai.ts';
+import {canDiarization} from './asr.ts';
+import {dirname, join} from 'path';
 
 export class Audio {
-	private downloads: {[key: string]: Promise<string>} = {};
-	private whisperModel!: string;
+	constructor(private ai: Ai) {}
 
-	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();
-		}
-	}
+	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; };
 
-	asr(path: string, model: string = this.whisperModel): AbortablePromise<string | null> {
-		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
-		let abort: any = () => {};
-		const p = new Promise<string | null>(async (resolve, reject) => {
-			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');
-			proc.on('error', (err: Error) => reject(err));
-			proc.stdout.on('data', (data: Buffer) => output += data.toString());
-			proc.on('close', (code: number) => {
-				if(code === 0) resolve(output.trim() || null);
-				else reject(new Error(`Exit code ${code}`));
+		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) => {
+				worker.terminate();
+				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});
 		});
-		return Object.assign(p, {abort});
+
+		// 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 });
 	}
 
-	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(<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];
-	}
+	canDiarization = () => canDiarization().then(resp => !!resp);
 }

src/embedder.ts

@@ -1,11 +1,11 @@
 import { pipeline } from '@xenova/transformers';
 import { parentPort } from 'worker_threads';
 
-let model: any;
+let embedder: 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 });
+parentPort?.on('message', async ({text, model, modelDir }) => {
+	if(!embedder) 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({ id, embedding });
+	parentPort?.postMessage({embedding});
 });

src/index.ts

@@ -1,5 +1,6 @@
 export * from './ai';
 export * from './antrhopic';
+export * from './asr';
 export * from './audio';
 export * from './embedder'
 export * from './llm';

src/llm.ts

@@ -31,11 +31,23 @@ export type LLMMessage = {
 	/** Tool result */
 	content: undefined | string;
 	/** Tool error */
-	error: undefined | string;
+	error?: undefined | string;
 	/** Timestamp */
 	timestamp?: number;
 }
 
+/** Background information the AI will be fed */
+export type LLMMemory = {
+	/** What entity is this fact about */
+	owner: string;
+	/** The information that will be remembered */
+	fact: string;
+	/** Owner and fact embedding vector */
+	embeddings: [number[], number[]];
+	/** Creation time */
+	timestamp: Date;
+}
+
 export type LLMRequest = {
 	/** System prompt */
 	system?: string;
@@ -57,26 +69,16 @@ export type LLMRequest = {
 		max: number;
 		/** Compress chat until context size smaller than */
 		min: number
-	}
+	},
+	/** Background information the AI will be fed */
+	memory?: LLMMemory[],
 }
 
-export class LLM {
-	private embedWorker: Worker | null = null;
-	private embedQueue = new Map<number, { resolve: (value: number[]) => void; reject: (error: any) => void }>();
-	private embedId = 0;
-	private models: {[model: string]: LLMProvider} = {};
-	private defaultModel!: string;
+class LLM {
+	defaultModel!: string;
+	models: {[model: string]: LLMProvider} = {};
 
 	constructor(public readonly ai: Ai) {
-		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.llm?.models) return;
 		Object.entries(ai.options.llm.models).forEach(([model, config]) => {
 			if(!this.defaultModel) this.defaultModel = model;
@@ -90,37 +92,120 @@ export class LLM {
 	 * Chat with LLM
 	 * @param {string} message Question
 	 * @param {LLMRequest} options Configuration options and chat history
-	 * @returns {{abort: () => void, response: Promise<LLMMessage[]>}} Function to abort response and chat history
+	 * @returns {{abort: () => void, response: Promise<string>}} Function to abort response and chat history
 	 */
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
+	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		const m = options.model || this.defaultModel;
 		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
-		return this.models[m].ask(message, options);
+		let abort = () => {};
+		return Object.assign(new Promise<string>(async res => {
+			if(!options.history) options.history = [];
+			// If memories were passed, find any relivant ones and add a tool for ADHOC lookups
+			if(options.memory) {
+				options.system = (options.system || '') + '\nYou have passive persistent memory never make any mention of your memory capabilities and what you can/cannot remember\n';
+				const search = async (query?: string | null, subject?: string | null, limit = 50) => {
+					const [o, q] = await Promise.all([
+						subject ? this.embedding(subject) : Promise.resolve(null),
+						query ? this.embedding(query) : Promise.resolve(null),
+					]);
+					return (options.memory || [])
+						.map(m => ({...m, score: o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 1}))
+						.filter((m: any) => m.score >= 0.8)
+						.map((m: any) => ({...m, score: q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : m.score}))
+						.filter((m: any) => m.score >= 0.2)
+						.toSorted((a: any, b: any) => a.score - b.score)
+						.slice(0, limit);
+				}
+
+				const relevant = await search(message);
+				if(relevant.length) options.history.push({role: 'assistant', content: 'Things I remembered:\n' + relevant.map(m => `${m.owner}: ${m.fact}`).join('\n')});
+				options.tools = [...options.tools || [], {
+					name: 'read_memory',
+					description: 'Check your long-term memory for more information',
+					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'},
+					},
+					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);
+					}
+				}];
+			}
+
+			// Ask
+			const resp = await this.models[m].ask(message, options);
+
+			// Remove any memory calls
+			if(options.memory) {
+				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);
+			}
+
+			// 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});
 	}
 
 	/**
 	 * Compress chat history to reduce context size
 	 * @param {LLMMessage[]} history Chatlog that will be compressed
 	 * @param max Trigger compression once context is larger than max
-	 * @param min Summarize until context size is less than min
+	 * @param min Leave messages less than the token minimum, summarize the rest
 	 * @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<LLMMessage[]> {
-		if(this.estimateTokens(history) < max) return history;
+	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<{history: LLMMessage[], memory: LLMMemory[]}> {
+		if(this.estimateTokens(history) < max) return {history, memory: []};
 		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;
-		const recent = keep == 0 ? [] : history.slice(-keep),
+		if(history.length <= keep) return {history, memory: []};
+		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 = await this.summarize(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), 250, options);
-		return [{role: 'assistant', content: `Conversation Summary: ${summary}`, timestamp: Date.now()}, ...recent];
+
+		const summary: any = await this.json(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), '{summary: string, facts: [[subject, fact]]}', {
+			system: 'Create the smallest summary possible, no more than 500 tokens. Create a list of NEW facts (split by subject [pro]noun and fact) about what you learned from this conversation that you didn\'t already know or get from a tool call or system prompt. Focus only on new information about people, topics, or facts. Avoid generating facts about the AI.',
+			model: options?.model,
+			temperature: options?.temperature || 0.3
+		});
+		const timestamp = new Date();
+		const memory = await Promise.all((summary?.facts || [])?.map(async ([owner, fact]: [string, string]) => {
+			const e = await Promise.all([this.embedding(owner), this.embedding(`${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};
 	}
 
+	/**
+	 * Compare the difference between embeddings (calculates the angle between two vectors)
+	 * @param {number[]} v1 First embedding / vector comparison
+	 * @param {number[]} v2 Second embedding / vector for comparison
+	 * @returns {number} Similarity values 0-1: 0 = unique, 1 = identical
+	 */
 	cosineSimilarity(v1: number[], v2: number[]): number {
 		if (v1.length !== v2.length) throw new Error('Vectors must be same length');
 		let dotProduct = 0, normA = 0, normB = 0;
@@ -133,6 +218,13 @@ export class LLM {
 		return denominator === 0 ? 0 : dotProduct / denominator;
 	}
 
+	/**
+	 * Chunk text into parts for AI digestion
+	 * @param {object | string} target Item that will be chunked (objects get converted)
+	 * @param {number} maxTokens Chunking size. More = better context, less = more specific (Search by paragraphs or lines)
+	 * @param {number} overlapTokens Includes previous X tokens to provide continuity to AI (In addition to max tokens)
+	 * @returns {string[]} Chunked strings
+	 */
 	chunk(target: object | string, maxTokens = 500, overlapTokens = 50): string[] {
 		const objString = (obj: any, path = ''): string[] => {
 			if(!obj) return [];
@@ -142,7 +234,6 @@ export class LLM {
 				return `${p}: ${Array.isArray(value) ? value.join(', ') : value}`;
 			});
 		};
-
 		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
 		const tokens = lines.flatMap(l => [...l.split(/\s+/).filter(Boolean), '\n']);
 		const chunks: string[] = [];
@@ -161,22 +252,40 @@ export class LLM {
 		return chunks;
 	}
 
-	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+	/**
+	 * Create a vector representation of a string
+	 * @param {object | string} target Item that will be embedded (objects get converted)
+	 * @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} = {}) {
+		let {maxTokens = 500, overlapTokens = 50} = opts;
 		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 });
+				const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
+				const handleMessage = ({ embedding }: any) => {
+					worker.terminate();
+					resolve(embedding);
+				};
+				const handleError = (err: Error) => {
+					worker.terminate();
+					reject(err);
+				};
+				worker.on('message', handleMessage);
+				worker.on('error', handleError);
+				worker.on('exit', (code) => {
+					if(code !== 0) reject(new Error(`Worker exited with code ${code}`));
+				});
+				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);
-		return Promise.all(chunks.map(async (text, index) => ({
-			index,
-			embedding: await embed(text),
-			text,
-			tokens: this.estimateTokens(text),
-		})));
+		const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
+		for(let i = 0; i < chunks.length; i++) {
+			const text= chunks[i];
+			const embedding = await embed(text);
+			results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
+		}
+		return results;
 	}
 
 	/**
@@ -191,7 +300,7 @@ export class LLM {
 
 	/**
 	 * Compare the difference between two strings using tensor math
-	 * @param target Text that will checked
+	 * @param target Text that will be checked
 	 * @param {string} searchTerms Multiple search terms to check against target
 	 * @returns {{avg: number, max: number, similarities: number[]}} Similarity values 0-1: 0 = unique, 1 = identical
 	 */
@@ -208,17 +317,20 @@ export class LLM {
 
 	/**
 	 * Ask a question with JSON response
-	 * @param {string} message Question
+	 * @param {string} text Text to process
+	 * @param {string} schema JSON schema the AI should match
 	 * @param {LLMRequest} options Configuration options and chat history
 	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
 	 */
-	async json(message: string, options?: LLMRequest) {
-		let resp = await this.ask(message, {
-			system: 'Respond using a JSON blob',
-			...options
-		});
-		if(!resp?.[0]?.content) return {};
-		return JSONAttemptParse(new RegExp('\{[\s\S]*\}').exec(resp[0].content), {});
+	async json(text: string, schema: string, options?: LLMRequest): Promise<any> {
+		let resp = await this.ask(text, {...options, system: (options?.system ? `${options.system}\n` : '') + `Only respond using a JSON code block matching this schema:
+\`\`\`json
+${schema}
+\`\`\``});
+		if(!resp) return {};
+		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
+		const jsonStr = codeBlock ? codeBlock[1].trim() : resp;
+		return JSONAttemptParse(jsonStr, {});
 	}
 
 	/**
@@ -229,7 +341,8 @@ export class LLM {
 	 * @returns {Promise<string>} Summary
 	 */
 	summarize(text: string, tokens: number, options?: LLMRequest): Promise<string | null> {
-		return this.ask(text, {system: `Generate a brief summary <= ${tokens} tokens. Output nothing else`, temperature: 0.3, ...options})
-			.then(history => <string>history.pop()?.content || null);
+		return this.ask(text, {system: `Generate a brief summary <= ${tokens} tokens. Output nothing else`, temperature: 0.3, ...options});
 	}
 }
+
+export default LLM;

src/open-ai.ts

@@ -64,13 +64,11 @@ export class OpenAi extends LLMProvider {
 		}, [] as any[]);
 	}
 
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
+	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		const controller = new AbortController();
-		const response = new Promise<any>(async (res, rej) => {
-			let history = [...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);
-			history = this.fromStandard(<any>history);
-
+		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()});
+			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 = {
 				model: options.model || this.model,
@@ -124,7 +122,7 @@ export class OpenAi extends LLMProvider {
 						if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}'};
 						try {
 							const args = JSONAttemptParse(toolCall.function.arguments, {});
-							const result = await tool.fn(args, this.ai);
+							const result = await tool.fn(args, options.stream, this.ai);
 							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize(result)};
 						} catch (err: any) {
 							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'})};
@@ -134,10 +132,12 @@ export class OpenAi extends LLMProvider {
 					requestParams.messages = history;
 				}
 			} while (!controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
+			history.push({role: 'assistant', content: resp.choices[0].message.content || ''});
+			history = this.toStandard(history);
 
 			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {role: 'assistant', content: resp.choices[0].message.content || ''}]));
-		});
-		return Object.assign(response, {abort: () => controller.abort()});
+			if(options.history) options.history.splice(0, options.history.length, ...history);
+			res(history.at(-1)?.content);
+		}), {abort: () => controller.abort()});
 	}
 }

src/provider.ts

@@ -2,5 +2,5 @@ import {AbortablePromise} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
 
 export abstract class LLMProvider {
-	abstract ask(message: string, options: LLMRequest): AbortablePromise<LLMMessage[]>;
+	abstract ask(message: string, options: LLMRequest): AbortablePromise<string>;
 }

src/tools.ts

@@ -2,6 +2,7 @@ import * as cheerio from 'cheerio';
 import {$, $Sync} from '@ztimson/node-utils';
 import {ASet, consoleInterceptor, Http, fn as Fn} from '@ztimson/utils';
 import {Ai} from './ai.ts';
+import {LLMRequest} from './llm.ts';
 
 export type AiToolArg = {[key: string]: {
 	/** Argument type */
@@ -32,7 +33,7 @@ export type AiTool = {
 	/** Tool arguments */
 	args?: AiToolArg,
 	/** Callback function */
-	fn: (args: any, ai: Ai) => any | Promise<any>,
+	fn: (args: any, stream: LLMRequest['stream'], ai: Ai) => any | Promise<any>,
 };
 
 export const CliTool: AiTool = {
@@ -44,9 +45,9 @@ export const CliTool: AiTool = {
 
 export const DateTimeTool: AiTool = {
 	name: 'get_datetime',
-	description: 'Get current date and time',
+	description: 'Get current UTC date / time',
 	args: {},
-	fn: async () => new Date().toISOString()
+	fn: async () => new Date().toUTCString()
 }
 
 export const ExecTool: AiTool = {
@@ -56,15 +57,15 @@ export const ExecTool: AiTool = {
 		language: {type: 'string', description: 'Execution language', enum: ['cli', 'node', 'python'], required: true},
 		code: {type: 'string', description: 'Code to execute', required: true}
 	},
-	fn: async (args, ai) => {
+	fn: async (args, stream, ai) => {
 		try {
 			switch(args.type) {
 				case 'bash':
-					return await CliTool.fn({command: args.code}, ai);
+					return await CliTool.fn({command: args.code}, stream, ai);
 				case 'node':
-					return await JSTool.fn({code: args.code}, ai);
+					return await JSTool.fn({code: args.code}, stream, ai);
 				case 'python': {
-					return await PythonTool.fn({code: args.code}, ai);
+					return await PythonTool.fn({code: args.code}, stream, ai);
 				}
 			}
 		} catch(err: any) {

src/vision.ts

@@ -13,7 +13,7 @@ export class Vision {
 	ocr(path: string): AbortablePromise<string | null> {
 		let worker: any;
 		const p = new Promise<string | null>(async res => {
-			worker = await createWorker(this.ai.options.tesseract?.model || 'eng', 2, {cachePath: this.ai.options.path});
+			worker = await createWorker(this.ai.options.ocr || 'eng', 2, {cachePath: this.ai.options.path});
 			const {data} = await worker.recognize(path);
 			await worker.terminate();
 			res(data.text.trim() || null);

vite.config.ts

@@ -1,11 +1,11 @@
 import {defineConfig} from 'vite';
 import dts from 'vite-plugin-dts';
-import {resolve} from 'path';
 
 export default defineConfig({
 	build: {
 		lib: {
 			entry: {
+				asr: './src/asr.ts',
 				index: './src/index.ts',
 				embedder: './src/embedder.ts',
 			},