Updated LLM config and added read_webpage

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.1.12",
+	"version": "0.4.0",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -27,9 +27,10 @@
 	"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"
 	},

src/ai.ts

@@ -1,120 +1,44 @@
-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> & {abort: () => any};
+
+export type AiOptions = {
+	/** Path to models */
+	path?: 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;
-		/** Path to models */
-		path: string;
-		/** Path to storage location for temporary files */
-		temp?: 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);
-			this.downloadAsrModel();
+		this.language = new LLM(this);
-		}
+		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);
-		await $`rm -f ${output} && ${this.options.whisper.binary} -nt -np -m ${m} -f ${path} -otxt -of ${output}`;
-		return fs.readFile(output, 'utf-8').then(text => text?.trim() || null)
-			.finally(() => fs.rm(output, {force: true}).catch(() => {}));
-	}
-
-	/**
-	 * Downloads the specified Whisper model if it is not already present locally.
-	 *
-	 * @param {string} model Whisper model that will be downloaded
-	 * @return {Promise<string>} Absolute path to model file, resolves once downloaded
-	 */
-	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
-		if(!this.options.whisper?.binary) throw new Error('Whisper not configured');
-		if(!model.endsWith('.bin')) model += '.bin';
-		const p = Path.join(this.options.whisper.path, model);
-		if(await fs.stat(p).then(() => true).catch(() => false)) return p;
-		if(!!this.downloads[model]) return this.downloads[model];
-		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${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];
-	}
-
-	/**
-	 * 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;
@@ -19,7 +19,7 @@ export class Anthropic extends LLMProvider {
 				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});
+						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) => {
@@ -29,6 +29,7 @@ export class Anthropic extends LLMProvider {
 				}
 				history[orgI].content = history[orgI].content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
 			}
+			if(!history[orgI].timestamp) history[orgI].timestamp = Date.now();
 		}
 		return history.filter(h => !!h.content);
 	}
@@ -44,20 +45,23 @@ export class Anthropic extends LLMProvider {
 				i++;
 			}
 		}
-		return history;
+		return history.map(({timestamp, ...h}) => h);
 	}
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
 		const controller = new AbortController();
 		const response = new Promise<any>(async (res, rej) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
+			let history = [...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);
+			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
+			history = this.fromStandard(<any>history);
+
+			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 +75,18 @@ export class Anthropic extends LLMProvider {
 				stream: !!options.stream,
 			};
 
-			// Run tool changes
+			let resp: any, isFirstMessage = true;
-			let resp: any;
+			const assistantMessages: string[] = [];
 			do {
-				resp = await this.client.messages.create(requestParams);
+				resp = await this.client.messages.create(requestParams).catch(err => {
+					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
+					throw err;
+				});
 
 				// Streaming mode
 				if(options.stream) {
+					if(!isFirstMessage) options.stream({text: '\n\n'});
+					else isFirstMessage = false;
 					resp.content = [];
 					for await (const chunk of resp) {
 						if(controller.signal.aborted) break;
@@ -109,7 +118,8 @@ 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);
@@ -122,12 +132,11 @@ export class Anthropic extends LLMProvider {
 					requestParams.messages = history;
 				}
 			} while (!controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
+
 			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {
+			res(this.toStandard([...history, {role: 'assistant', content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')}]));
-				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()});
 	}
 }

src/audio.ts

@@ -0,0 +1,50 @@
+import {spawn} from 'node:child_process';
+import fs from 'node:fs/promises';
+import Path from 'node:path';
+import {AbortablePromise, Ai} from './ai.ts';
+
+export class Audio {
+	private downloads: {[key: string]: Promise<string>} = {};
+	private whisperModel!: string;
+
+	constructor(private ai: Ai) {
+		if(ai.options.whisper?.binary) {
+			this.whisperModel = ai.options.whisper?.model.endsWith('.bin') ? ai.options.whisper?.model : ai.options.whisper?.model + '.bin';
+			this.downloadAsrModel();
+		}
+	}
+
+	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 ${code}`));
+			});
+		});
+		return Object.assign(p, {abort});
+	}
+
+	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
+		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
+		if(!model.endsWith('.bin')) model += '.bin';
+		const p = Path.join(<string>this.ai.options.path, model);
+		if(await fs.stat(p).then(() => true).catch(() => false)) return p;
+		if(!!this.downloads[model]) return this.downloads[model];
+		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${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];
+	}
+}

src/embedder.ts

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

src/index.ts

@@ -1,4 +1,9 @@
 export * from './ai';
 export * from './antrhopic';
+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';
@@ -24,34 +32,10 @@ export type LLMMessage = {
 	content: undefined | string;
 	/** Tool error */
 	error: undefined | string;
+	/** Timestamp */
+	timestamp?: number;
 }
 
-export type LLMOptions = {
-	/** Anthropic settings */
-	anthropic?: {
-		/** API Token */
-		token: string;
-		/** Default model */
-		model: string;
-	},
-	/** Ollama settings */
-	ollama?: {
-		/** 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 */
 	system?: string;
@@ -64,9 +48,9 @@ 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 */
@@ -77,12 +61,29 @@ export type LLMRequest = {
 }
 
 export class LLM {
-	private providers: {[key: string]: LLMProvider} = {};
+	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;
 
-	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);
+		this.embedWorker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
-		if(options.ollama?.host) this.providers.ollama = new Ollama(this.ai, options.ollama.host, options.ollama.model);
+		this.embedWorker.on('message', ({ id, embedding }) => {
-		if(options.openAi?.token) this.providers.openAi = new OpenAi(this.ai, options.openAi.token, options.openAi.model);
+			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;
+			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);
+		});
 	}
 
 	/**
@@ -92,17 +93,9 @@ export class LLM {
 	 * @returns {{abort: () => void, response: Promise<LLMMessage[]>}} Function to abort response and chat history
 	 */
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
-		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;
+		return this.models[m].ask(message, options);
-			else model = [options.model, (<any>this.options)[options.model]?.model];
-		}
-		if(!options.model || model[1] == null) {
-			if(typeof this.options.model == 'object') model = this.options.model;
-			else model = [this.options.model, (<any>this.options)[this.options.model]?.model];
-		}
-		if(!model[0] || !model[1]) throw new Error(`Unknown LLM provider or model: ${model[0]} / ${model[1]}`);
-		return this.providers[model[0]].ask(message, {...options, model: model[1]});
 	}
 
 	/**
@@ -113,7 +106,7 @@ export class LLM {
 	 * @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<LLMMessage[]> {
 		if(this.estimateTokens(history) < max) return history;
 		let keep = 0, tokens = 0;
 		for(let m of history.toReversed()) {
@@ -125,7 +118,65 @@ export class LLM {
 		const 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];
+		return [{role: 'assistant', content: `Conversation Summary: ${summary}`, timestamp: Date.now()}, ...recent];
+	}
+
+	cosineSimilarity(v1: number[], v2: number[]): number {
+		if (v1.length !== v2.length) throw new Error('Vectors must be same length');
+		let dotProduct = 0, normA = 0, normB = 0;
+		for (let i = 0; i < v1.length; i++) {
+			dotProduct += v1[i] * v2[i];
+			normA += v1[i] * v1[i];
+			normB += v2[i] * v2[i];
+		}
+		const denominator = Math.sqrt(normA) * Math.sqrt(normB);
+		return denominator === 0 ? 0 : dotProduct / denominator;
+	}
+
+	chunk(target: object | string, maxTokens = 500, overlapTokens = 50): string[] {
+		const objString = (obj: any, path = ''): string[] => {
+			if(!obj) return [];
+			return Object.entries(obj).flatMap(([key, value]) => {
+				const p = path ? `${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;
+	}
+
+	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+		const embed = (text: string): Promise<number[]> => {
+			return new Promise((resolve, reject) => {
+				const id = this.embedId++;
+				this.embedQueue.set(id, { resolve, reject });
+				this.embedWorker?.postMessage({ id, text });
+			});
+		};
+
+		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),
+		})));
 	}
 
 	/**
@@ -138,6 +189,23 @@ 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 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

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;
 	}
@@ -46,14 +50,15 @@ 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[]);
@@ -62,16 +67,18 @@ export class OpenAi extends LLMProvider {
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
 		const controller = new AbortController();
 		const response = new Promise<any>(async (res, rej) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
+			let history = [...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);
+			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
+			history = this.fromStandard(<any>history);
 
+			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,
@@ -85,28 +92,35 @@ 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, {});
@@ -124,7 +138,6 @@ export class OpenAi extends LLMProvider {
 			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()});
 	}
 }

src/provider.ts

@@ -1,6 +1,5 @@
-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[]>;

src/tools.ts

@@ -1,3 +1,4 @@
+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';
@@ -111,9 +112,43 @@ 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.tesseract?.model || '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 +1,19 @@
 import {defineConfig} from 'vite';
 import dts from 'vite-plugin-dts';
+import {resolve} from 'path';
 
 export default defineConfig({
 	build: {
 		lib: {
-			entry: './src/index.ts',
+			entry: {
+				index: './src/index.ts',
+				embedder: './src/embedder.ts',
+			},
 			name: 'utils',
-			fileName: (format) => (format === 'es' ? 'index.mjs' : 'index.js'),
+			fileName: (format, entryName) => {
+				if (entryName === 'embedder') return 'embedder.js';
+				return format === 'es' ? 'index.mjs' : 'index.js';
+			},
 		},
 		ssr: true,
 		emptyOutDir: true,