Updated LLM config and added read_webpage

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.2.5",
+	"version": "0.4.0",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -30,7 +30,7 @@
 		"@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

@@ -1,11 +1,22 @@
 import * as os from 'node:os';
-import {LLM, LLMOptions} from './llm';
+import {LLM, AnthropicConfig, OllamaConfig, OpenAiConfig, LLMRequest} from './llm';
 import { Audio } from './audio.ts';
 import {Vision} from './vision.ts';
 
-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 */
@@ -13,11 +24,6 @@ export type AiOptions = LLMOptions & {
 		/** Model: `ggml-base.en.bin` */
 		model: string;
 	}
-	/** Tesseract OCR configuration */
-	tesseract?: {
-		/** Model: eng, eng_best, eng_fast */
-		model?: string;
-	}
 }
 
 export class Ai {

src/antrhopic.ts

@@ -1,8 +1,8 @@
 import {Anthropic as anthropic} from '@anthropic-ai/sdk';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, deepCopy} from '@ztimson/utils';
+import {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;
@@ -51,16 +51,16 @@ export class Anthropic 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, timestamp: Date.now()}]);
+			let history = [...options.history || [], {role: 'user', content: message, timestamp: Date.now()}];
-			const original = deepCopy(history);
 			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
+			history = this.fromStandard(<any>history);
 
-			const tools = options.tools || this.ai.options.tools || [];
+			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: tools.map(t => ({
 					name: t.name,
 					description: t.description,
@@ -117,9 +117,9 @@ export class Anthropic extends LLMProvider {
 				const toolCalls = resp.content.filter((c: any) => c.type === 'tool_use');
 				if(toolCalls.length && !controller.signal.aborted) {
 					history.push({role: 'assistant', content: resp.content});
-					original.push({role: 'assistant', content: resp.content});
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
 						const tool = tools.find(findByProp('name', toolCall.name));
+						if(options.stream) options.stream({tool: toolCall.name});
 						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
 						try {
 							const result = await tool.fn(toolCall.input, this.ai);

src/audio.ts

@@ -1,7 +1,7 @@
 import {spawn} from 'node:child_process';
 import fs from 'node:fs/promises';
 import Path from 'node:path';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 
 export class Audio {
 	private downloads: {[key: string]: Promise<string>} = {};
@@ -14,17 +14,11 @@ export class Audio {
 		}
 	}
 
-	/**
+	asr(path: string, model: string = this.whisperModel): AbortablePromise<string | null> {
-	 * Convert audio to text using Auditory Speech Recognition
-	 * @param {string} path Path to audio
-	 * @param model Whisper model
-	 * @returns {Promise<any>} Extracted text
-	 */
-	asr(path: string, model: string = this.whisperModel): {abort: () => void, response: Promise<string | null>} {
 		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
 		let abort: any = () => {};
-		const response = new Promise<string | null>((resolve, reject) => {
+		const p = new Promise<string | null>(async (resolve, reject) => {
-			this.downloadAsrModel(model).then(m => {
+			const m = await this.downloadAsrModel(model);
 			let output = '';
 			const proc = spawn(<string>this.ai.options.whisper?.binary, ['-nt', '-np', '-m', m, '-f', path], {stdio: ['ignore', 'pipe', 'ignore']});
 			abort = () => proc.kill('SIGTERM');
@@ -35,16 +29,9 @@ export class Audio {
 				else reject(new Error(`Exit code ${code}`));
 			});
 		});
-		});
+		return Object.assign(p, {abort});
-		return {response, abort};
 	}
 
-	/**
-	 * Downloads the specified Whisper model if it is not already present locally.
-	 *
-	 * @param {string} model Whisper model that will be downloaded
-	 * @return {Promise<string>} Absolute path to model file, resolves once downloaded
-	 */
 	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
 		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
 		if(!model.endsWith('.bin')) model += '.bin';

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

@@ -1,12 +1,16 @@
-import {pipeline} from '@xenova/transformers';
 import {JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
-import {Ollama} from './ollama.ts';
 import {OpenAi} from './open-ai.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
-import * as tf from '@tensorflow/tfjs';
+import {Worker} from 'worker_threads';
+import {fileURLToPath} from 'url';
+import {dirname, join} from 'path';
+
+export type 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 */
@@ -32,32 +36,6 @@ export type LLMMessage = {
 	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;
@@ -70,9 +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 */
@@ -83,14 +61,29 @@ export type LLMRequest = {
 }
 
 export class LLM {
-	private embedModel: any;
+	private embedWorker: Worker | null = null;
-	private providers: {[key: string]: LLMProvider} = {};
+	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) {
-		this.embedModel = pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2');
+		this.embedWorker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
-		if(ai.options.anthropic?.token) this.providers.anthropic = new Anthropic(this.ai, ai.options.anthropic.token, ai.options.anthropic.model);
+		this.embedWorker.on('message', ({ id, embedding }) => {
-		if(ai.options.ollama?.host) this.providers.ollama = new Ollama(this.ai, ai.options.ollama.host, ai.options.ollama.model);
+			const pending = this.embedQueue.get(id);
-		if(ai.options.openAi?.token) this.providers.openAi = new OpenAi(this.ai, ai.options.openAi.token, ai.options.openAi.model);
+			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);
+		});
 	}
 
 	/**
@@ -100,17 +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.ai.options)[options.model]?.model];
-		}
-		if(!options.model || model[1] == null) {
-			if(typeof this.ai.options.model == 'object') model = this.ai.options.model;
-			else model = [this.ai.options.model, (<any>this.ai.options)[this.ai.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]});
 	}
 
 	/**
@@ -148,49 +133,44 @@ export class LLM {
 		return denominator === 0 ? 0 : dotProduct / denominator;
 	}
 
-	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+	chunk(target: object | string, maxTokens = 500, overlapTokens = 50): string[] {
 		const objString = (obj: any, path = ''): string[] => {
-			if(obj === null || obj === undefined) return [];
+			if(!obj) return [];
 			return Object.entries(obj).flatMap(([key, value]) => {
 				const p = path ? `${path}${isNaN(+key) ? `.${key}` : `[${key}]`}` : key;
-				if(typeof value === 'object' && value !== null && !Array.isArray(value)) return objString(value, p);
+				if(typeof value === 'object' && !Array.isArray(value)) return objString(value, p);
-				const valueStr = Array.isArray(value) ? value.join(', ') : String(value);
+				return `${p}: ${Array.isArray(value) ? value.join(', ') : value}`;
-				return `${p}: ${valueStr}`;
 			});
 		};
 
-		const embed = async (text: string): Promise<number[]> => {
+		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
-			const model = await this.embedModel;
+		const tokens = lines.flatMap(l => [...l.split(/\s+/).filter(Boolean), '\n']);
-			const output = await model(text, {pooling: 'mean', normalize: true});
+		const chunks: string[] = [];
-			return Array.from(output.data);
+		for(let i = 0; i < tokens.length;) {
+			let text = '', j = i;
+			while(j < tokens.length) {
+				const next = text + (text ? ' ' : '') + tokens[j];
+				if(this.estimateTokens(next.replace(/\s*\n\s*/g, '\n')) > maxTokens && text) break;
+				text = next;
+				j++;
+			}
+			const clean = text.replace(/\s*\n\s*/g, '\n').trim();
+			if(clean) chunks.push(clean);
+			i = Math.max(j - overlapTokens, j === i ? i + 1 : j);
+		}
+		return chunks;
+	}
+
+	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+		const embed = (text: string): Promise<number[]> => {
+			return new Promise((resolve, reject) => {
+				const id = this.embedId++;
+				this.embedQueue.set(id, { resolve, reject });
+				this.embedWorker?.postMessage({ id, text });
+			});
 		};
 
-		// Tokenize
+		const chunks = this.chunk(target, maxTokens, overlapTokens);
-		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
-		const tokens = lines.flatMap(line => [...line.split(/\s+/).filter(w => w.trim()), '\n']);
-
-		// Chunk
-		const chunks: string[] = [];
-		let start = 0;
-		while (start < tokens.length) {
-			let end = start;
-			let text = '';
-			// Build chunk
-			while (end < tokens.length) {
-				const nextToken = tokens[end];
-				const testText = text + (text ? ' ' : '') + nextToken;
-				const testTokens = this.estimateTokens(testText.replace(/\s*\n\s*/g, '\n'));
-				if (testTokens > maxTokens && text) break;
-				text = testText;
-				end++;
-			}
-			// Save chunk
-			const cleanText = text.replace(/\s*\n\s*/g, '\n').trim();
-			if(cleanText) chunks.push(cleanText);
-			start = end - overlapTokens;
-			if (start <= end - tokens.length + end) start = end; // Safety: prevent infinite loop
-		}
-
 		return Promise.all(chunks.map(async (text, index) => ({
 			index,
 			embedding: await embed(text),

src/ollama.ts

@@ -1,122 +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, timestamp: history[i].timestamp};
-			}
-			if(!history[i]?.timestamp) history[i].timestamp = Date.now();
-		}
-		return history;
-	}
-
-	private fromStandard(history: LLMMessage[]): any[] {
-		return history.map((h: any) => {
-			const {timestamp, ...rest} = h;
-			if(h.role != 'tool') return rest;
-			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, timestamp: Date.now()}]);
-			if(history[0].roll == 'system') {
-				if(!system) system = history.shift();
-				else history.shift();
-			}
-			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min);
-			if(options.system) history.unshift({role: 'system', content: system})
-
-			const tools = options.tools || this.ai.options.tools || [];
-			const requestParams: any = {
-				model: options.model || this.model,
-				messages: history,
-				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: 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]) : []
-						}
-					}
-				}))
-			}
-
-			let resp: any, isFirstMessage = true;
-			do {
-				resp = await this.client.chat(requestParams).catch(err => {
-					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
-					throw err;
-				});
-
-				if(options.stream) {
-					if(!isFirstMessage) options.stream({text: '\n\n'});
-					else isFirstMessage = false;
-					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;
-					}
-				}
-
-				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 = 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[] {
@@ -64,16 +67,17 @@ 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, timestamp: Date.now()}]);
+			let history = [...options.history || [], {role: 'user', content: message, timestamp: Date.now()}];
 			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
+			history = this.fromStandard(<any>history);
 
-			const tools = options.tools || this.ai.options.tools || [];
+			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: tools.map(t => ({
 					type: 'function',
 					function: {
@@ -116,6 +120,7 @@ export class OpenAi extends LLMProvider {
 					history.push(resp.choices[0].message);
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
 						const tool = tools?.find(findByProp('name', toolCall.function.name));
+						if(options.stream) options.stream({tool: toolCall.function.name});
 						if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}'};
 						try {
 							const args = JSONAttemptParse(toolCall.function.arguments, {});

src/provider.ts

@@ -1,6 +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

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