parallel embedding cap

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`
+2. For speaker diarization: `pip install pyannote.audio`
-3. Run unit tests: `npm test`
+3. Build library: `npm build`
+4. Run unit tests: `npm test`
 
 </details>
 

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.1.13",
+	"version": "0.7.2",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -27,11 +27,13 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.67.0",
 		"@tensorflow/tfjs": "^4.22.0",
+		"@xenova/transformers": "^2.17.2",
 		"@ztimson/node-utils": "^1.0.4",
 		"@ztimson/utils": "^0.27.9",
-		"ollama": "^0.6.0",
+		"cheerio": "^1.2.0",
 		"openai": "^6.6.0",
-		"tesseract.js": "^6.0.1"
+		"tesseract.js": "^6.0.1",
+		"wavefile": "^11.0.0"
 	},
 	"devDependencies": {
 		"@types/node": "^24.8.1",

src/ai.ts

@@ -1,127 +1,42 @@
-import {$} from '@ztimson/node-utils';
+import * as os from 'node:os';
-import {createWorker} from 'tesseract.js';
+import LLM, {AnthropicConfig, OllamaConfig, OpenAiConfig, LLMRequest} from './llm';
-import {LLM, LLMOptions} from './llm';
+import { Audio } from './audio.ts';
-import fs from 'node:fs/promises';
+import {Vision} from './vision.ts';
-import Path from 'node:path';
-import * as tf from '@tensorflow/tfjs';
 
-export type AiOptions = LLMOptions & {
+export type AbortablePromise<T> = Promise<T> & {
-	whisper?: {
+	abort: () => any
-		/** Whisper binary location */
+};
-		binary: string;
+
-		/** Model: `ggml-base.en.bin` */
+export type AiOptions = {
-		model: string;
+	/** Token to pull models from hugging face */
-		/** Path to models */
+	hfToken?: string;
-		path: string;
+	/** Path to models */
-		/** Path to storage location for temporary files */
+	path?: string;
-		temp?: 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};
 	}
+	/** OCR model: eng, eng_best, eng_fast */
+	ocr?: string;
 }
 
 export class Ai {
-	private downloads: {[key: string]: Promise<string>} = {};
+	/** Audio processing AI */
-	private whisperModel!: string;
+	audio!: Audio;
-
+	/** Language processing AI */
-	/** Large Language Models */
+	language!: LLM;
-	llm!: LLM;
+	/** Vision processing AI */
+	vision!: Vision;
 
 	constructor(public readonly options: AiOptions) {
-		this.llm = new LLM(this, options);
+		if(!options.path) options.path = os.tmpdir();
-		if(this.options.whisper?.binary) {
+		process.env.TRANSFORMERS_CACHE = options.path;
-			this.whisperModel = this.options.whisper?.model.endsWith('.bin') ? this.options.whisper?.model : this.options.whisper?.model + '.bin';
+		this.audio = new Audio(this);
-			console.log('constructor: ' + this.options.whisper.model + ' -> ' + this.whisperModel);
+		this.language = new LLM(this);
-			this.downloadAsrModel();
+		this.vision = new Vision(this);
-		}
-	}
-
-	/**
-	 * Convert audio to text using Auditory Speech Recognition
-	 * @param {string} path Path to audio
-	 * @param model Whisper model
-	 * @returns {Promise<any>} Extracted text
-	 */
-	async asr(path: string, model: string = this.whisperModel): Promise<string | null> {
-		if(!this.options.whisper?.binary) throw new Error('Whisper not configured');
-		const m = await this.downloadAsrModel(model);
-		const name = Math.random().toString(36).substring(2, 10) + '-' + path.split('/').pop() + '.txt';
-		const output = Path.join(this.options.whisper.temp || '/tmp', name);
-		console.log('ASR: ' + this.options.whisper.model + ' -> ' + this.whisperModel);
-		console.log(`rm -f ${output} && ${this.options.whisper.binary} -nt -np -m ${m} -f ${path} -otxt -of ${output}`);
-		await $`rm -f ${output} && ${this.options.whisper.binary} -nt -np -m ${m} -f ${path} -otxt -of ${output}`;
-		return fs.readFile(output, 'utf-8').then(text => text?.trim() || null)
-			.finally(() => fs.rm(output, {force: true}).catch(() => {}));
-	}
-
-	/**
-	 * Downloads the specified Whisper model if it is not already present locally.
-	 *
-	 * @param {string} model Whisper model that will be downloaded
-	 * @return {Promise<string>} Absolute path to model file, resolves once downloaded
-	 */
-	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
-		if(!this.options.whisper?.binary) throw new Error('Whisper not configured');
-		if(!model.endsWith('.bin')) model += '.bin';
-		const p = Path.join(this.options.whisper.path, model);
-		console.log('Download: ' + p);
-		if(await fs.stat(p).then(() => true).catch(() => false)) {
-			console.log('Exists!');
-			return p;
-		}
-		if(!!this.downloads[model]) return this.downloads[model];
-		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${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];
-	}
-
-	/**
-	 * 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
-	 */
-	ocr(path: string): {abort: () => void, response: Promise<string | null>} {
-		let worker: any;
-		return {
-			abort: () => { worker?.terminate(); },
-			response: new Promise(async res => {
-				worker = await createWorker('eng');
-				const {data} = await worker.recognize(path);
-				await worker.terminate();
-				res(data.text.trim() || null);
-			})
-		}
-	}
-
-	/**
-	 * Compare the difference between two strings using tensor math
-	 * @param target Text that will 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
-	 */
-	semanticSimilarity(target: string, ...searchTerms: string[]) {
-		if(searchTerms.length < 2) throw new Error('Requires at least 2 strings to compare');
-
-		const vector = (text: string, dimensions: number = 10): number[] => {
-			return text.toLowerCase().split('').map((char, index) =>
-				(char.charCodeAt(0) * (index + 1)) % dimensions / dimensions).slice(0, dimensions);
-		}
-
-		const cosineSimilarity = (v1: number[], v2: number[]): number => {
-			if (v1.length !== v2.length) throw new Error('Vectors must be same length');
-			const tensor1 = tf.tensor1d(v1), tensor2 = tf.tensor1d(v2)
-			const dotProduct = tf.dot(tensor1, tensor2)
-			const magnitude1 = tf.norm(tensor1)
-			const magnitude2 = tf.norm(tensor2)
-			if(magnitude1.dataSync()[0] === 0 || magnitude2.dataSync()[0] === 0) return 0
-			return dotProduct.dataSync()[0] / (magnitude1.dataSync()[0] * magnitude2.dataSync()[0])
-		}
-
-		const v = vector(target);
-		const similarities = searchTerms.map(t => vector(t)).map(refVector => cosineSimilarity(v, refVector))
-		return {avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length, max: Math.max(...similarities), similarities}
 	}
 }

src/antrhopic.ts

@@ -1,8 +1,8 @@
 import {Anthropic as anthropic} from '@anthropic-ai/sdk';
 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 Anthropic extends LLMProvider {
 	client!: anthropic;
@@ -13,24 +13,25 @@ export class Anthropic extends LLMProvider {
 	}
 
 	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
+		const timestamp = Date.now();
-			const orgI = i;
+		const messages: LLMMessage[] = [];
-			if(typeof history[orgI].content != 'string') {
+		for(let h of history) {
-				if(history[orgI].role == 'assistant') {
+			if(typeof h.content == 'string') {
-					history[orgI].content.filter((c: any) => c.type =='tool_use').forEach((c: any) => {
+				messages.push(<any>{timestamp, ...h});
-						i++;
+			} else {
-						history.splice(i, 0, {role: 'tool', id: c.id, name: c.name, args: c.input});
+				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});
-				} else if(history[orgI].role == 'user') {
+				h.content.forEach((c: any) => {
-					history[orgI].content.filter((c: any) => c.type =='tool_result').forEach((c: any) => {
+					if(c.type == 'tool_use') {
-						const h = history.find((h: any) => h.id == c.tool_use_id);
+						messages.push({timestamp, role: 'tool', id: c.id, name: c.name, args: c.input, content: undefined});
-						h[c.is_error ? 'error' : 'content'] = c.content;
+					} 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;
-				history[orgI].content = history[orgI].content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
+					}
+				});
 			}
 		}
-		return history.filter(h => !!h.content);
+		return messages;
 	}
 
 	private fromStandard(history: LLMMessage[]): any[] {
@@ -44,20 +45,20 @@ export class Anthropic extends LLMProvider {
 				i++;
 			}
 		}
-		return history;
+		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) => {
+		return Object.assign(new Promise<any>(async (res) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
+			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
-			if(options.compress) history = await this.ai.llm.compress(<any>history, options.compress.max, options.compress.min, options);
+			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
-				max_tokens: options.max_tokens || this.ai.options.max_tokens || 4096,
+				max_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || 4096,
-				system: options.system || this.ai.options.system || '',
+				system: options.system || this.ai.options.llm?.system || '',
-				temperature: options.temperature || this.ai.options.temperature || 0.7,
+				temperature: options.temperature || this.ai.options.llm?.temperature || 0.7,
-				tools: (options.tools || this.ai.options.tools || []).map(t => ({
+				tools: tools.map(t => ({
 					name: t.name,
 					description: t.description,
 					input_schema: {
@@ -71,13 +72,17 @@ export class Anthropic extends LLMProvider {
 				stream: !!options.stream,
 			};
 
-			// Run tool changes
+			let resp: any, isFirstMessage = true;
-			let resp: any;
 			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) {
+					if(!isFirstMessage) options.stream({text: '\n\n'});
+					else isFirstMessage = false;
 					resp.content = [];
 					for await (const chunk of resp) {
 						if(controller.signal.aborted) break;
@@ -109,10 +114,11 @@ export class Anthropic extends LLMProvider {
 				if(toolCalls.length && !controller.signal.aborted) {
 					history.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);
+							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'};
@@ -122,12 +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, {
+			if(options.history) options.history.splice(0, options.history.length, ...history);
-				role: 'assistant',
+			res(history.at(-1)?.content);
-				content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')
+		}), {abort: () => controller.abort()});
-			}]));
-		});
-		return Object.assign(response, {abort: () => controller.abort()});
 	}
 }

src/asr.ts

@@ -0,0 +1,134 @@
+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<boolean> {
+	return new Promise((resolve) => {
+		const proc = spawn('python', ['-c', 'import pyannote.audio']);
+		proc.on('close', (code: number) => resolve(code === 0));
+		proc.on('error', () => resolve(false));
+	});
+}
+
+async function runDiarization(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('python', ['-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 = speaker && await canDiarization();
+		const [transcript, speakers] = await Promise.all([
+			whisperPipeline(buffer, {return_timestamps: speaker ? 'word' : false}),
+			(!speaker || !token || !hasDiarization) ? Promise.resolve(): runDiarization(f, modelDir, token),
+		]);
+		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

@@ -0,0 +1,60 @@
+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 {
+	constructor(private ai: Ai) {}
+
+	asr(file: string, options: { model?: string; speaker?: boolean | 'id' } = {}): AbortablePromise<string | null> {
+		const { model = this.ai.options.asr || 'whisper-base', speaker = false } = options;
+		let aborted = false;
+		const abort = () => { aborted = true; };
+
+		let p = new Promise<string | null>((resolve, reject) => {
+			const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'asr.js'));
+			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});
+		});
+
+		// 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"}', {
+					system: 'Use this following transcript to identify speakers. Only identify speakers you are sure about',
+					temperature: 0.1,
+				});
+				Object.entries(names).forEach(([speaker, name]) => {
+					transcript = (<string>transcript).replaceAll(`[Speaker ${speaker}]`, `[${name}]`);
+				});
+				return transcript;
+			})
+		}
+
+		return Object.assign(p, { abort });
+	}
+
+	canDiarization = canDiarization;
+}

src/embedder.ts

@@ -0,0 +1,11 @@
+import { pipeline } from '@xenova/transformers';
+import { parentPort } from 'worker_threads';
+
+let embedder: any;
+
+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({embedding});
+});

src/index.ts

@@ -1,4 +1,10 @@
 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';
 export * from './tools';
+export * from './vision';

src/llm.ts

@@ -1,16 +1,24 @@
 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 {Worker} from 'worker_threads';
+import {fileURLToPath} from 'url';
+import {dirname, join} from 'path';
+
+export type AnthropicConfig = {proto: 'anthropic', token: string};
+export type OllamaConfig = {proto: 'ollama', host: string};
+export type OpenAiConfig = {proto: 'openai', host?: string, token: string};
 
 export type LLMMessage = {
 	/** Message originator */
 	role: 'assistant' | 'system' | 'user';
 	/** Message content */
 	content: string | any;
+	/** Timestamp */
+	timestamp?: number;
 } | {
 	/** Tool call */
 	role: 'tool';
@@ -23,34 +31,22 @@ export type LLMMessage = {
 	/** Tool result */
 	content: undefined | string;
 	/** Tool error */
-	error: undefined | string;
+	error?: undefined | string;
+	/** Timestamp */
+	timestamp?: number;
 }
 
-export type LLMOptions = {
+/** Background information the AI will be fed */
-	/** Anthropic settings */
+export type LLMMemory = {
-	anthropic?: {
+	/** What entity is this fact about */
-		/** API Token */
+	owner: string;
-		token: string;
+	/** The information that will be remembered */
-		/** Default model */
+	fact: string;
-		model: string;
+	/** Owner and fact embedding vector */
-	},
+	embeddings: [number[], number[]];
-	/** Ollama settings */
+	/** Creation time */
-	ollama?: {
+	timestamp: Date;
-		/** connection URL */
+}
-		host: string;
-		/** Default model */
-		model: string;
-	},
-	/** Open AI settings */
-	openAi?: {
-		/** API Token */
-		token: string;
-		/** Default model */
-		model: string;
-	},
-	/** Default provider & model */
-	model: string | [string, string];
-} & Omit<LLMRequest, 'model'>;
 
 export type LLMRequest = {
 	/** System prompt */
@@ -64,68 +60,239 @@ export type LLMRequest = {
 	/** Available tools */
 	tools?: AiTool[];
 	/** LLM  model */
-	model?: string | [string, string];
+	model?: 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 */
 		max: number;
 		/** Compress chat until context size smaller than */
 		min: number
-	}
+	},
+	/** Background information the AI will be fed */
+	memory?: LLMMemory[],
 }
 
-export class LLM {
+class LLM {
-	private providers: {[key: string]: LLMProvider} = {};
+	defaultModel!: string;
+	models: {[model: string]: LLMProvider} = {};
 
-	constructor(public readonly ai: Ai, public readonly options: LLMOptions) {
+	constructor(public readonly ai: Ai) {
-		if(options.anthropic?.token) this.providers.anthropic = new Anthropic(this.ai, options.anthropic.token, options.anthropic.model);
+		if(!ai.options.llm?.models) return;
-		if(options.ollama?.host) this.providers.ollama = new Ollama(this.ai, options.ollama.host, options.ollama.model);
+		Object.entries(ai.options.llm.models).forEach(([model, config]) => {
-		if(options.openAi?.token) this.providers.openAi = new OpenAi(this.ai, options.openAi.token, options.openAi.model);
+			if(!this.defaultModel) this.defaultModel = model;
+			if(config.proto == 'anthropic') this.models[model] = new Anthropic(this.ai, config.token, model);
+			else if(config.proto == 'ollama') this.models[model] = new OpenAi(this.ai, config.host, 'not-needed', model);
+			else if(config.proto == 'openai') this.models[model] = new OpenAi(this.ai, config.host || null, config.token, model);
+		});
 	}
 
 	/**
 	 * 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> {
-		let model: any = [null, null];
+		const m = options.model || this.defaultModel;
-		if(options.model) {
+		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
-			if(typeof options.model == 'object') model = options.model;
+		let abort = () => {};
-			else model = [options.model, (<any>this.options)[options.model]?.model];
+		return Object.assign(new Promise<string>(async res => {
-		}
+			if(!options.history) options.history = [];
-		if(!options.model || model[1] == null) {
+			// If memories were passed, find any relivant ones and add a tool for ADHOC lookups
-			if(typeof this.options.model == 'object') model = this.options.model;
+			if(options.memory) {
-			else model = [this.options.model, (<any>this.options)[this.options.model]?.model];
+				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) => {
-		if(!model[0] || !model[1]) throw new Error(`Unknown LLM provider or model: ${model[0]} / ${model[1]}`);
+					const [o, q] = await Promise.all([
-		return this.providers[model[0]].ask(message, {...options, model: model[1]});
+						subject ? this.embedding(subject) : Promise.resolve(null),
+						query ? this.embedding(query) : Promise.resolve(null),
+					]);
+					return (options.memory || [])
+						.map(m => ({...m, score: o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 1}))
+						.filter((m: any) => m.score >= 0.8)
+						.map((m: any) => ({...m, score: q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : m.score}))
+						.filter((m: any) => m.score >= 0.2)
+						.toSorted((a: any, b: any) => a.score - b.score)
+						.slice(0, limit);
+				}
+
+				const relevant = await search(message);
+				if(relevant.length) options.history.push({role: 'assistant', content: 'Things I remembered:\n' + relevant.map(m => `${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 compress(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<LLMMessage[]> {
+	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<{history: LLMMessage[], memory: LLMMemory[]}> {
-		if(this.estimateTokens(history) < max) return history;
+		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;
+		if(history.length <= keep) return {history, memory: []};
-		const recent = keep == 0 ? [] : history.slice(-keep),
+		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: ${summary}`}, ...recent];
+		const summary: any = await this.json(process.map(m => `${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: ${summary?.summary}`, timestamp: Date.now()}, ...recent];
+		if(system) h.splice(0, 0, system);
+		return {history: <any>h, memory};
+	}
+
+	/**
+	 * Compare the difference between embeddings (calculates the angle between two vectors)
+	 * @param {number[]} v1 First embedding / vector comparison
+	 * @param {number[]} v2 Second embedding / vector for comparison
+	 * @returns {number} Similarity values 0-1: 0 = unique, 1 = identical
+	 */
+	cosineSimilarity(v1: number[], v2: number[]): number {
+		if (v1.length !== v2.length) throw new Error('Vectors must be same length');
+		let dotProduct = 0, normA = 0, normB = 0;
+		for (let i = 0; i < v1.length; i++) {
+			dotProduct += v1[i] * v2[i];
+			normA += v1[i] * v1[i];
+			normB += v2[i] * v2[i];
+		}
+		const denominator = Math.sqrt(normA) * Math.sqrt(normB);
+		return denominator === 0 ? 0 : dotProduct / denominator;
+	}
+
+	/**
+	 * Chunk text into parts for AI digestion
+	 * @param {object | string} target Item that will be chunked (objects get converted)
+	 * @param {number} maxTokens Chunking size. More = better context, less = more specific (Search by paragraphs or lines)
+	 * @param {number} overlapTokens Includes previous X tokens to provide continuity to AI (In addition to max tokens)
+	 * @returns {string[]} Chunked strings
+	 */
+	chunk(target: object | string, maxTokens = 500, overlapTokens = 50): string[] {
+		const objString = (obj: any, path = ''): string[] => {
+			if(!obj) return [];
+			return Object.entries(obj).flatMap(([key, value]) => {
+				const p = path ? `${path}${isNaN(+key) ? `.${key}` : `[${key}]`}` : key;
+				if(typeof value === 'object' && !Array.isArray(value)) return objString(value, p);
+				return `${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[] = [];
+		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;
+	}
+
+	/**
+	 * Create a vector representation of a string
+	 * @param {object | string} target Item that will be embedded (objects get converted)
+	 * @param {maxTokens?: number, overlapTokens?: number, parellel?: number} opts Options for embedding such as chunk sizes and parallel processing
+	 * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
+	 */
+	async embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number, parallel?: number} = {}) {
+		let {maxTokens = 500, overlapTokens = 50, parallel = 1} = opts;
+
+		const embed = (text: string): Promise<number[]> => {
+			return new Promise((resolve, reject) => {
+				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});
+			});
+		};
+		let i = 0, chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
+		const next: Function = () => {
+			const index = i++;
+			if(index >= chunks.length) return;
+			const text = chunks[index];
+			return embed(text).then(embedding => {
+				results.push({index, embedding, text, tokens: this.estimateTokens(text)});
+				return next();
+			})
+		}
+
+		await Promise.all(Array(parallel).fill(null).map(() => next()));
+		return results.toSorted((a, b) => a.index - b.index);
 	}
 
 	/**
@@ -138,19 +305,39 @@ export class LLM {
 		return Math.ceil((text.length / 4) * 1.2);
 	}
 
+	/**
+	 * Compare the difference between two strings using tensor math
+	 * @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
+	 */
+	fuzzyMatch(target: string, ...searchTerms: string[]) {
+		if(searchTerms.length < 2) throw new Error('Requires at least 2 strings to compare');
+		const vector = (text: string, dimensions: number = 10): number[] => {
+			return text.toLowerCase().split('').map((char, index) =>
+				(char.charCodeAt(0) * (index + 1)) % dimensions / dimensions).slice(0, dimensions);
+		}
+		const v = vector(target);
+		const similarities = searchTerms.map(t => vector(t)).map(refVector => this.cosineSimilarity(v, refVector))
+		return {avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length, max: Math.max(...similarities), similarities}
+	}
+
 	/**
 	 * 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) {
+	async json(text: string, schema: string, options?: LLMRequest): Promise<any> {
-		let resp = await this.ask(message, {
+		let resp = await this.ask(text, {...options, system: (options?.system ? `${options.system}\n` : '') + `Only respond using a JSON code block matching this schema:
-			system: 'Respond using a JSON blob',
+\`\`\`json
-			...options
+${schema}
-		});
+\`\`\``});
-		if(!resp?.[0]?.content) return {};
+		if(!resp) return {};
-		return JSONAttemptParse(new RegExp('\{[\s\S]*\}').exec(resp[0].content), {});
+		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
+		const jsonStr = codeBlock ? codeBlock[1].trim() : resp;
+		return JSONAttemptParse(jsonStr, {});
 	}
 
 	/**
@@ -161,7 +348,8 @@ 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})
+		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);
 	}
 }
+
+export default LLM;

src/ollama.ts

@@ -1,113 +0,0 @@
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
-import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
-import {Ollama as ollama} from 'ollama';
-
-export class Ollama extends LLMProvider {
-	client!: ollama;
-
-	constructor(public readonly ai: Ai, public host: string, public model: string) {
-		super();
-		this.client = new ollama({host});
-	}
-
-	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
-			if(history[i].role == 'assistant' && history[i].tool_calls) {
-				if(history[i].content) delete history[i].tool_calls;
-				else {
-					history.splice(i, 1);
-					i--;
-				}
-			} else if(history[i].role == 'tool') {
-				const error = history[i].content.startsWith('{"error":');
-				history[i] = {role: 'tool', name: history[i].tool_name, args: history[i].args, [error ? 'error' : 'content']: history[i].content};
-			}
-		}
-		return history;
-	}
-
-	private fromStandard(history: LLMMessage[]): any[] {
-		return history.map((h: any) => {
-			if(h.role != 'tool') return h;
-			return {role: 'tool', tool_name: h.name, content: h.error || h.content}
-		});
-	}
-
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
-		const controller = new AbortController();
-		const response = new Promise<any>(async (res, rej) => {
-			let system = options.system || this.ai.options.system;
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
-			if(history[0].roll == 'system') {
-				if(!system) system = history.shift();
-				else history.shift();
-			}
-			if(options.compress) history = await this.ai.llm.compress(<any>history, options.compress.max, options.compress.min);
-			if(options.system) history.unshift({role: 'system', content: system})
-
-			const requestParams: any = {
-				model: options.model || this.model,
-				messages: history,
-				stream: !!options.stream,
-				signal: controller.signal,
-				options: {
-					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 => ({
-					type: 'function',
-					function: {
-						name: t.name,
-						description: t.description,
-						parameters: {
-							type: 'object',
-							properties: t.args ? objectMap(t.args, (key, value) => ({...value, required: undefined})) : {},
-							required: t.args ? Object.entries(t.args).filter(t => t[1].required).map(t => t[0]) : []
-						}
-					}
-				}))
-			}
-
-			// Run tool chains
-			let resp: any;
-			do {
-				resp = await this.client.chat(requestParams);
-				if(options.stream) {
-					resp.message = {role: 'assistant', content: '', tool_calls: []};
-					for await (const chunk of resp) {
-						if(controller.signal.aborted) break;
-						if(chunk.message?.content) {
-							resp.message.content += chunk.message.content;
-							options.stream({text: chunk.message.content});
-						}
-						if(chunk.message?.tool_calls) resp.message.tool_calls = chunk.message.tool_calls;
-						if(chunk.done) break;
-					}
-				}
-
-				// Run tools
-				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));
-						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 {
-							const result = await tool.fn(args, this.ai);
-							return {role: 'tool', tool_name: toolCall.function.name, args, content: JSONSanitize(result)};
-						} catch (err: any) {
-							return {role: 'tool', tool_name: toolCall.function.name, args, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'})};
-						}
-					}));
-					history.push(...results);
-					requestParams.messages = history;
-				}
-			} while (!controller.signal.aborted && resp.message?.tool_calls?.length);
-			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {role: 'assistant', content: resp.message?.content}]));
-		});
-		return Object.assign(response, {abort: () => controller.abort()});
-	}
-}

src/open-ai.ts

@@ -1,15 +1,18 @@
 import {OpenAI as openAI} from 'openai';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean} 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;
 
-	constructor(public readonly ai: Ai, public readonly apiToken: string, public model: string) {
+	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string, public model: string) {
 		super();
-		this.client = new openAI({apiKey: apiToken});
+		this.client = new openAI(clean({
+			baseURL: host,
+			apiKey: token
+		}));
 	}
 
 	private toStandard(history: any[]): LLMMessage[] {
@@ -20,7 +23,8 @@ export class OpenAi extends LLMProvider {
 					role: 'tool',
 					id: tc.id,
 					name: tc.function.name,
-					args: JSONAttemptParse(tc.function.arguments, {})
+					args: JSONAttemptParse(tc.function.arguments, {}),
+					timestamp: h.timestamp
 				}));
 				history.splice(i, 1, ...tools);
 				i += tools.length - 1;
@@ -33,7 +37,7 @@ export class OpenAi extends LLMProvider {
 				history.splice(i, 1);
 				i--;
 			}
-
+			if(!history[i]?.timestamp) history[i].timestamp = Date.now();
 		}
 		return history;
 	}
@@ -46,32 +50,33 @@ export class OpenAi extends LLMProvider {
 					content: null,
 					tool_calls: [{ id: h.id, type: 'function', function: { name: h.name, arguments: JSON.stringify(h.args) } }],
 					refusal: null,
-					annotations: [],
+					annotations: []
 				}, {
 					role: 'tool',
 					tool_call_id: h.id,
 					content: h.error || h.content
 				});
 			} else {
-				result.push(h);
+				const {timestamp, ...rest} = h;
+				result.push(rest);
 			}
 			return result;
 		}, [] 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) => {
+		return Object.assign(new Promise<any>(async (res, rej) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
+			if(options.system && options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
-			if(options.compress) history = await this.ai.llm.compress(<any>history, options.compress.max, options.compress.min, options);
+			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,
 				messages: history,
 				stream: !!options.stream,
-				max_tokens: options.max_tokens || this.ai.options.max_tokens || 4096,
+				max_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || 4096,
-				temperature: options.temperature || this.ai.options.temperature || 0.7,
+				temperature: options.temperature || this.ai.options.llm?.temperature || 0.7,
-				tools: (options.tools || this.ai.options.tools || []).map(t => ({
+				tools: tools.map(t => ({
 					type: 'function',
 					function: {
 						name: t.name,
@@ -85,32 +90,39 @@ export class OpenAi extends LLMProvider {
 				}))
 			};
 
-			// Tool call and streaming logic similar to other providers
+			let resp: any, isFirstMessage = true;
-			let resp: any;
 			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;
+				});
 
-				// Implement streaming and tool call handling
 				if(options.stream) {
-					resp.choices = [];
+					if(!isFirstMessage) options.stream({text: '\n\n'});
+					else isFirstMessage = false;
+					resp.choices = [{message: {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;
+						}
 					}
 				}
 
-				// Run tools
 				const toolCalls = resp.choices[0].message.tool_calls || [];
 				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, {});
-							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'})};
@@ -120,11 +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 || ''}]));
+			if(options.history) options.history.splice(0, options.history.length, ...history);
-		});
+			res(history.at(-1)?.content);
-
+		}), {abort: () => controller.abort()});
-		return Object.assign(response, {abort: () => controller.abort()});
 	}
 }

src/provider.ts

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

src/tools.ts

@@ -1,6 +1,8 @@
+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 */
@@ -31,7 +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 = {
@@ -43,9 +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 = {
@@ -55,15 +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) {
@@ -111,9 +113,43 @@ export const PythonTool: AiTool = {
 	fn: async (args: {code: string}) => ({result: $Sync`python -c "${args.code}"`})
 }
 
-export const SearchTool: AiTool = {
+export const ReadWebpageTool: AiTool = {
-	name: 'search',
+	name: 'read_webpage',
-	description: 'Use a search engine to find relevant URLs, should be changed with fetch to scrape sources',
+	description: 'Extract clean, structured content from a webpage. Use after web_search to read specific URLs',
+	args: {
+		url: {type: 'string', description: 'URL to extract content from', required: true},
+		focus: {type: 'string', description: 'Optional: What aspect to focus on (e.g., "pricing", "features", "contact info")'}
+	},
+	fn: async (args: {url: string; focus?: string}) => {
+		const html = await fetch(args.url, {headers: {"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64)"}})
+			.then(r => r.text()).catch(err => {throw new Error(`Failed to fetch: ${err.message}`)});
+
+		const $ = cheerio.load(html);
+		$('script, style, nav, footer, header, aside, iframe, noscript, [role="navigation"], [role="banner"], .ad, .ads, .cookie, .popup').remove();
+		const metadata = {
+			title: $('meta[property="og:title"]').attr('content') || $('title').text() || '',
+			description: $('meta[name="description"]').attr('content') || $('meta[property="og:description"]').attr('content') || '',
+		};
+
+		let content = '';
+		const contentSelectors = ['article', 'main', '[role="main"]', '.content', '.post', '.entry', 'body'];
+		for (const selector of contentSelectors) {
+			const el = $(selector).first();
+			if (el.length && el.text().trim().length > 200) {
+				content = el.text();
+				break;
+			}
+		}
+		if (!content) content = $('body').text();
+		content = content.replace(/\s+/g, ' ').trim().slice(0, 8000);
+
+		return {url: args.url, title: metadata.title.trim(), description: metadata.description.trim(), content, focus: args.focus};
+	}
+}
+
+export const WebSearchTool: AiTool = {
+	name: 'web_search',
+	description: 'Use duckduckgo (anonymous) to find find relevant online resources. Returns a list of URLs that works great with the `read_webpage` tool',
 	args: {
 		query: {type: 'string', description: 'Search string', required: true},
 		length: {type: 'string', description: 'Number of results to return', default: 5},

src/vision.ts

@@ -0,0 +1,23 @@
+import {createWorker} from 'tesseract.js';
+import {AbortablePromise, Ai} from './ai.ts';
+
+export class Vision {
+
+	constructor(private ai: Ai) { }
+
+	/**
+	 * Convert image to text using Optical Character Recognition
+	 * @param {string} path Path to image
+	 * @returns {AbortablePromise<string | null>} Promise of extracted text with abort method
+	 */
+	ocr(path: string): AbortablePromise<string | null> {
+		let worker: any;
+		const p = new Promise<string | null>(async res => {
+			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);
+		});
+		return Object.assign(p, {abort: () => worker?.terminate()});
+	}
+}

vite.config.ts

@@ -1,12 +1,20 @@
 import {defineConfig} from 'vite';
 import dts from 'vite-plugin-dts';
+import {resolve} from 'path';
 
 export default defineConfig({
 	build: {
 		lib: {
-			entry: './src/index.ts',
+			entry: {
+				asr: './src/asr.ts',
+				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,