Proper error handling for OCR

+ added wikitools
+ Improved webpage reading tool

README.md

@@ -3,7 +3,7 @@
   <br />
 
   <!-- Logo -->
-  <img src="https://git.zakscode.com/repo-avatars/a90851ca730480ec37a5c0c2c4f1b4609eee5eadf806eaf16c83ac4cb7493aa9" alt="Logo" width="200" height="200">
+  <img alt="Logo" width="200" height="200" src="https://git.zakscode.com/repo-avatars/a82d423674763e7a0c1c945bdbb07e249b2bb786d3c9beae76d5b196a10f5c0f">
 
   <!-- Title -->
 ### @ztimson/ai-utils
@@ -53,13 +53,15 @@ A TypeScript library that provides a unified interface for working with multiple
 - **Provider Abstraction**: Switch between AI providers without changing your code
 
 ### Built With
-[![Anthropic](https://img.shields.io/badge/Anthropic-191919?style=for-the-badge&logo=anthropic&logoColor=white)](https://anthropic.com/)
-[![OpenAI](https://img.shields.io/badge/OpenAI-412991?style=for-the-badge&logo=openai&logoColor=white)](https://openai.com/)
-[![Ollama](https://img.shields.io/badge/Ollama-000000?style=for-the-badge&logo=ollama&logoColor=white)](https://ollama.com/)
-[![TensorFlow](https://img.shields.io/badge/TensorFlow-FF6F00?style=for-the-badge&logo=tensorflow&logoColor=white)](https://tensorflow.org/)
-[![Tesseract](https://img.shields.io/badge/Tesseract-3C8FC7?style=for-the-badge&logo=tesseract&logoColor=white)](https://tesseract-ocr.github.io/)
+[![Anthropic](https://img.shields.io/badge/Anthropic-de7356?style=for-the-badge&logo=anthropic&logoColor=white)](https://anthropic.com/)
+[![llama](https://img.shields.io/badge/llama.cpp-fff?style=for-the-badge&logo=ollama&logoColor=black)](https://github.com/ggml-org/llama.cpp)
+[![OpenAI](https://img.shields.io/badge/OpenAI-000?style=for-the-badge&logo=openai-gym&logoColor=white)](https://openai.com/)
+[![Pyannote](https://img.shields.io/badge/Pyannote-458864?style=for-the-badge&logo=python&logoColor=white)](https://github.com/pyannote)
+[![TensorFlow](https://img.shields.io/badge/TensorFlow-fff?style=for-the-badge&logo=tensorflow&logoColor=ff6f00)](https://tensorflow.org/)
+[![Tesseract](https://img.shields.io/badge/Tesseract-B874B2?style=for-the-badge&logo=hack-the-box&logoColor=white)](https://tesseract-ocr.github.io/)
+[![Transformers.js](https://img.shields.io/badge/Transformers.js-000?style=for-the-badge&logo=hugging-face&logoColor=yellow)](https://huggingface.co/docs/transformers.js/en/index)
 [![TypeScript](https://img.shields.io/badge/TypeScript-3178C6?style=for-the-badge&logo=typescript&logoColor=white)](https://typescriptlang.org/)
-[![Whisper](https://img.shields.io/badge/Whisper-412991?style=for-the-badge&logo=openai&logoColor=white)](https://github.com/ggerganov/whisper.cpp)
+[![Whisper](https://img.shields.io/badge/Whisper.cpp-000?style=for-the-badge&logo=openai-gym&logoColor=white)](https://github.com/ggerganov/whisper.cpp)
 
 ## Setup
 
@@ -75,6 +77,7 @@ 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>
 
@@ -87,17 +90,138 @@ A TypeScript library that provides a unified interface for working with multiple
 
 #### Prerequisites
 - [Node.js](https://nodejs.org/en/download)
+- _[Whisper.cpp](https://github.com/ggml-org/whisper.cpp/releases/tag) (ASR)_
+- _[Pyannote](https://github.com/pyannote) (ASR Diarization):_ `pip install pyannote.audio`
 
 #### Instructions
 1. Install the dependencies: `npm i`
-2. Build library: `npm build`
-3. Run unit tests: `npm test`
+2. For speaker diarization: `pip install pyannote.audio`
+3. Build library: `npm build`
+4. Run unit tests: `npm test`
 
 </details>
 
 ## Documentation
 
-[Available Here](https://ai-utils.docs.zakscode.com/)
+### Setup
+```javascript
+const ai = new Ai({
+    path: '/ai-models',
+    
+    // Setup audio
+    whisper: '/path/to/binary', // Required for ASR
+    hfToken: '...', // Required for diarization
+    asr: 'ggml-base.en.bin', // Override default ASR model
+    
+    // Setup LLM
+    embedder: 'bge-small-en-v1.5', // Override default embedder model
+    llm: {
+        system: 'You are a helpful assistant.',
+        compress: {max: 90_000, min: 50_000}, // Compress chat history to min tokens when max is reached
+        temperature: 0.8,
+        max_tokens: 100_000,
+        memoryModel: 'gpt-4o', // Cheap model for managing memories in background, defaults to current model
+        models: {
+            'claude-3-5-sonnet': {proto: 'anthropic', token: process.env.ANTHROPIC_TOKEN},
+            'gpt-4o':            {proto: 'openai',    token: process.env.OPENAI_TOKEN},
+            'llama3':            {proto: 'ollama',    host: 'http://localhost:11434'},
+        },
+        mcp: [
+            {name: 'files', url: 'https://mcp.example.com', token: process.env.MCP_TOKEN}
+        ],
+        skills: [
+            {name: 'Tone of voice', description: 'Brand writing guidelines', content: '# Tone of Voice\n\nAlways be concise and friendly...'}
+        ],
+        tools: [{
+            name: 'Marco?',
+            description: 'Where is marco polo?',
+            args: {
+                shout: {type: 'boolean', default: 'Shout into the void?', description: false, required: false}
+            },
+            fn: (args: any, stream: LLMRequest['stream'], ai: Ai) => {
+                const {shout} = args;
+                return shout ? 'Polo!' : 'Polo';
+            }
+        }],
+    },
+	
+    // Setup Vision
+    ocr: 'eng' // Override default OCR model
+});
+
+```
+
+### Audio
+
+```javascript
+// Crate audio transcript
+const text = await ai.audio.asr('./path/to/audio.mp3');
+console.log(text);
+
+// Break transcript into speakers
+const text = await ai.audio.asr('./path/to/audio.mp3', {diarization: true});
+console.log(text);
+
+// Break transcript into named speakers
+const text = await ai.audio.asr('./path/to/audio.mp3', {diarization: 'llm'});
+console.log(text);
+```
+
+### Language
+
+```javascript
+const history = [], memory = [];
+
+// Wait for entire response
+const text = await ai.language.ask('My favorite color is blue, whats yours?', {history, memory});
+console.log(text);
+
+// Stream response
+const chunks = '';
+await ai.language.ask('Write me a poem', {
+	history, memory,
+	stream: chunk => chunks += chunk,
+});
+console.log(chunks);
+
+// Manually compile history into memories at end of conversation
+// Happens automatically when coverstaions are compressed
+await ai.language.updateMemory(history, memory);
+
+// Summarize text
+const summary = await ai.language.summarize(longText, 200);
+
+// Code response (no conversation or extra BS)
+const code = await ai.language.code('Write a fibonacci function');
+
+// Structured JSON response
+const data = await ai.language.json('Extract the name and age', `{
+    "name": "string",
+    "age": "number"
+}`, {system: 'Extract from user input'});
+```
+
+#### Premade LLM Tools:
+- `cli`: Run a shell command, returns its output
+- `get_datetime`: Returns local date/time
+- `get_datetime_utc`: Returns current UTC date/time
+- `exec`: Execute code in cli, node, or python
+- `fetch`: Make HTTP requests (GET/POST/PUT/DELETE)
+- `exec_javascript`: Execute CommonJS JavaScript
+- `exec_python`: Execute Python via python -c
+- `read_webpage`: Scrape & clean content from a URL, handles HTML, JSON, CSV, media, PDFs etc.
+- `web_search`: Anonymous DuckDuckGo search, returns a list of URLs
+- `wikipedia_lookup`: Fetch a Wikipedia article (intro or full)
+- `wikipedia_search`: Search Wikipedia and return matching articles
+- `get_weather`: Fetch current weather + forecast for a location (just built!)
+
+### Vision
+
+```javascript
+// Extract text from image
+const text = await ai.vision.ocr('./path/to/image.png');
+console.log(text);
+```
 
 ## License
 

main.mjs

@@ -0,0 +1,25 @@
+import {Ai} from './dist/index.mjs';
+
+const ai = new Ai({
+	path: './',
+	llm: {
+		system: 'You are a testbed for developing an AI library',
+		models: {
+			'qwen/qwen3.5-9b': {proto: 'openai', host: 'http://127.0.0.1:1234/v1'}
+		}
+	}
+});
+
+const skills = [{
+	name: 'Momentum',
+	description: 'Learn how to use the Momentum API',
+	content: 'You can initialize it with: new Momentum(url);'
+}];
+
+const history = [], memory = [];
+await ai.language.ask('My favorite color is red', {history, memory});
+await ai.language.updateMemory(history, memory);
+
+history.splice(0, history.length);
+console.log(await ai.language.ask('Whats my favorite color?', {history, memory}));
+console.log(history);

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.5.4",
+	"version": "1.0.5",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -25,21 +25,21 @@
 		"watch": "npx vite build --watch"
 	},
 	"dependencies": {
-		"@anthropic-ai/sdk": "^0.67.0",
+		"@anthropic-ai/sdk": "^0.102.0",
 		"@tensorflow/tfjs": "^4.22.0",
-		"@xenova/transformers": "^2.17.2",
-		"@ztimson/node-utils": "^1.0.4",
-		"@ztimson/utils": "^0.27.9",
+		"@huggingface/transformers": "^4.2.0",
+		"@ztimson/node-utils": "^1.0.7",
+		"@ztimson/utils": "^0.29.4",
 		"cheerio": "^1.2.0",
-		"openai": "^6.6.0",
-		"tesseract.js": "^6.0.1"
+		"openai": "^6.42.0",
+		"tesseract.js": "^7.0.0"
 	},
 	"devDependencies": {
-		"@types/node": "^24.8.1",
+		"@types/node": "^24.13.1",
 		"typedoc": "^0.26.7",
-		"typescript": "^5.3.3",
-		"vite": "^7.2.7",
-		"vite-plugin-dts": "^4.5.3"
+		"typescript": "^5.6.3",
+		"vite": "^8.0.16",
+		"vite-plugin-dts": "^5.0.2"
 	},
 	"files": [
 		"dist"

src/ai.ts

@@ -8,26 +8,22 @@ export type AbortablePromise<T> = Promise<T> & {
 };
 
 export type AiOptions = {
+	/** Token to pull diarization models from hugging face */
+	hfToken?: string;
 	/** Path to models */
 	path?: string;
-	/** Embedding model */
-	embedder?: string; // all-MiniLM-L6-v2, bge-small-en-v1.5, bge-large-en-v1.5
+	/** Whisper ASR model: ggml-tiny.en.bin, ggml-base.en.bin */
+	asr?: string;
+	/** Embedding model: all-MiniLM-L6-v2, bge-small-en-v1.5, bge-large-en-v1.5 */
+	embedder?: string;
 	/** Large language models, first is default */
 	llm?: Omit<LLMRequest, 'model'> & {
 		models: {[model: string]: AnthropicConfig | OllamaConfig | OpenAiConfig};
 	}
-	/** Tesseract OCR configuration */
-	tesseract?: {
-		/** Model: eng, eng_best, eng_fast */
-		model?: string;
-	}
-	/** Whisper ASR configuration */
-	whisper?: {
-		/** Whisper binary location */
-		binary: string;
-		/** Model: `ggml-base.en.bin` */
-		model: string;
-	}
+	/** OCR model: eng, eng_best, eng_fast */
+	ocr?: string;
+	/** Whisper binary */
+	whisper?: string;
 }
 
 export class Ai {

src/antrhopic.ts

@@ -13,25 +13,25 @@ export class Anthropic extends LLMProvider {
 	}
 
 	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
-			const orgI = i;
-			if(typeof history[orgI].content != 'string') {
-				if(history[orgI].role == 'assistant') {
-					history[orgI].content.filter((c: any) => c.type =='tool_use').forEach((c: any) => {
-						history.splice(i + 1, 0, {role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: Date.now()});
-					});
-				} else if(history[orgI].role == 'user') {
-					history[orgI].content.filter((c: any) => c.type =='tool_result').forEach((c: any) => {
-						const h = history.find((h: any) => h.id == c.tool_use_id);
-						h[c.is_error ? 'error' : 'content'] = c.content;
+		const timestamp = Date.now();
+		const messages: LLMMessage[] = [];
+		for(let h of history) {
+			if(typeof h.content == 'string') {
+				messages.push(<any>{timestamp, ...h});
+			} else {
+				const textContent = h.content?.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
+				if(textContent) messages.push({timestamp, role: h.role, content: textContent});
+				h.content.forEach((c: any) => {
+					if(c.type == 'tool_use') {
+						messages.push({timestamp, role: 'tool', id: c.id, name: c.name, args: c.input, content: undefined});
+					} else if(c.type == 'tool_result') {
+						const m: any = messages.findLast(m => (<any>m).id == c.tool_use_id);
+						if(m) m[c.is_error ? 'error' : 'content'] = c.content;
+					}
 				});
 			}
-				history[orgI].content = history[orgI].content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
-				if(!history[orgI].content) history.splice(orgI, 1);
 		}
-			if(!history[orgI].timestamp) history[orgI].timestamp = Date.now();
-		}
-		return history.filter(h => !!h.content);
+		return messages;
 	}
 
 	private fromStandard(history: LLMMessage[]): any[] {
@@ -50,8 +50,8 @@ export class Anthropic extends LLMProvider {
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		const controller = new AbortController();
-		return Object.assign(new Promise<any>(async (res, rej) => {
-			const history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
+		return Object.assign(new Promise<any>(async (res) => {
+			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
@@ -73,7 +73,6 @@ export class Anthropic extends LLMProvider {
 			};
 
 			let resp: any, isFirstMessage = true;
-			const assistantMessages: string[] = [];
 			do {
 				resp = await this.client.messages.create(requestParams).catch(err => {
 					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
@@ -119,9 +118,8 @@ export class Anthropic extends LLMProvider {
 						if(options.stream) options.stream({tool: toolCall.name});
 						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
 						try {
-							console.log(typeof tool.fn);
 							const result = await tool.fn(toolCall.input, options?.stream, this.ai);
-							return {type: 'tool_result', tool_use_id: toolCall.id, content: JSONSanitize(result)};
+							return {type: 'tool_result', tool_use_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
 						} catch (err: any) {
 							return {type: 'tool_result', tool_use_id: toolCall.id, is_error: true, content: err?.message || err?.toString() || 'Unknown'};
 						}
@@ -131,7 +129,7 @@ export class Anthropic extends LLMProvider {
 				}
 			} 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')});
-			this.toStandard(history);
+			history = this.toStandard(history);
 
 			if(options.stream) options.stream({done: true});
 			if(options.history) options.history.splice(0, options.history.length, ...history);

src/audio.ts

@@ -1,39 +1,258 @@
-import {spawn} from 'node:child_process';
+import {execSync, spawn} from 'node:child_process';
+import {mkdtempSync} from 'node:fs';
 import fs from 'node:fs/promises';
-import Path from 'node:path';
+import {tmpdir} from 'node:os';
+import Path, {join} from 'node:path';
 import {AbortablePromise, Ai} from './ai.ts';
 
 export class Audio {
 	private downloads: {[key: string]: Promise<string>} = {};
+	private pyannote!: 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';
+		if(ai.options.whisper) {
+			this.whisperModel = ai.options.asr || 'ggml-base.en.bin';
 			this.downloadAsrModel();
 		}
+
+		this.pyannote = `
+import sys
+import json
+import os
+from pyannote.audio import Pipeline
+
+os.environ['TORCH_HOME'] = r"${ai.options.path}"
+pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token="${ai.options.hfToken}")
+output = pipeline(sys.argv[1])
+
+segments = []
+for turn, speaker in output.speaker_diarization:
+    segments.append({"start": turn.start, "end": turn.end, "speaker": speaker})
+
+print(json.dumps(segments))
+`;
 	}
 
-	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);
+	private async addPunctuation(timestampData: any, llm?: boolean, cadence = 150): Promise<string> {
+		const countSyllables = (word: string): number => {
+			word = word.toLowerCase().replace(/[^a-z]/g, '');
+			if(word.length <= 3) return 1;
+			const matches = word.match(/[aeiouy]+/g);
+			let count = matches ? matches.length : 1;
+			if(word.endsWith('e')) count--;
+			return Math.max(1, count);
+		};
+
+		let result = '';
+		timestampData.transcription.filter((word, i) => {
+			let skip = false;
+			const prevWord = timestampData.transcription[i - 1];
+			const nextWord = timestampData.transcription[i + 1];
+			if(!word.text && nextWord) {
+				nextWord.offsets.from = word.offsets.from;
+				nextWord.timestamps.from = word.offsets.from;
+			} else if(word.text && word.text[0] != ' ' && prevWord) {
+				prevWord.offsets.to = word.offsets.to;
+				prevWord.timestamps.to = word.timestamps.to;
+				prevWord.text += word.text;
+				skip = true;
+			}
+			return !!word.text && !skip;
+		}).forEach((word: any) => {
+			const capital = /^[A-Z]/.test(word.text.trim());
+			const length = word.offsets.to - word.offsets.from;
+			const syllables = countSyllables(word.text.trim());
+			const expected = syllables * cadence;
+			if(capital && length > expected * 2 && word.text[0] == ' ') result += '.';
+			result += word.text;
+		});
+		if(!llm) return result.trim();
+		return this.ai.language.ask(result, {
+			system: 'Remove any misplaced punctuation from the following ASR transcript using the replace tool. Avoid modifying words unless there is an obvious typo',
+			temperature: 0.1,
+			tools: [{
+				name: 'replace',
+				description: 'Use find and replace to fix errors',
+				args: {
+					find: {type: 'string', description: 'Text to find', required: true},
+					replace: {type: 'string', description: 'Text to replace', required: true}
+				},
+				fn: (args) => result = result.replace(args.find, args.replace)
+			}]
+		}).then(() => result);
+	}
+
+	private async diarizeTranscript(timestampData: any, speakers: any[], llm: boolean): Promise<string> {
+		const speakerMap = new Map();
+		let speakerCount = 0;
+		speakers.forEach((seg: any) => {
+			if(!speakerMap.has(seg.speaker)) speakerMap.set(seg.speaker, ++speakerCount);
+		});
+
+		const punctuatedText = await this.addPunctuation(timestampData, llm);
+		const sentences = punctuatedText.match(/[^.!?]+[.!?]+/g) || [punctuatedText];
+		const words = timestampData.transcription.filter((w: any) => w.text.trim());
+
+		// Assign speaker to each sentence
+		const sentencesWithSpeakers = sentences.map(sentence => {
+			sentence = sentence.trim();
+			if(!sentence) return null;
+
+			const sentenceWords = sentence.toLowerCase().replace(/[^\w\s]/g, '').split(/\s+/);
+			const speakerWordCount = new Map<number, number>();
+
+			sentenceWords.forEach(sw => {
+				const word = words.find((w: any) => sw === w.text.trim().toLowerCase().replace(/[^\w]/g, ''));
+				if(!word) return;
+
+				const wordTime = word.offsets.from / 1000;
+				const speaker = speakers.find((seg: any) => wordTime >= seg.start && wordTime <= seg.end);
+				if(speaker) {
+					const spkNum = speakerMap.get(speaker.speaker);
+					speakerWordCount.set(spkNum, (speakerWordCount.get(spkNum) || 0) + 1);
+				}
+			});
+
+			let bestSpeaker = 1;
+			let maxWords = 0;
+			speakerWordCount.forEach((count, speaker) => {
+				if(count > maxWords) {
+					maxWords = count;
+					bestSpeaker = speaker;
+				}
+			});
+
+			return {speaker: bestSpeaker, text: sentence};
+		}).filter(s => s !== null);
+
+		// Merge adjacent sentences from same speaker
+		const merged: Array<{speaker: number, text: string}> = [];
+		sentencesWithSpeakers.forEach(item => {
+			const last = merged[merged.length - 1];
+			if(last && last.speaker === item.speaker) {
+				last.text += ' ' + item.text;
+			} else {
+				merged.push({...item});
+			}
+		});
+
+		let transcript = merged.map(item => `[Speaker ${item.speaker}]: ${item.text}`).join('\n').trim();
+		if(!llm) return transcript;
+		let chunks = this.ai.language.chunk(transcript, 500, 0);
+		if(chunks.length > 4) chunks = [...chunks.slice(0, 3), <string>chunks.at(-1)];
+		const names = await this.ai.language.json(chunks.join('\n'), '{1: "Detected Name", 2: "Second Name"}', {
+			system: 'Use the following transcript to identify speakers. Only identify speakers you are positive about, dont mention speakers you are unsure about in your response',
+			temperature: 0.1,
+		});
+		Object.entries(names).forEach(([speaker, name]) => transcript = transcript.replaceAll(`[Speaker ${speaker}]`, `[${name}]`));
+		return transcript;
+	}
+
+	private runAsr(file: string, opts: {model?: string, diarization?: boolean} = {}): AbortablePromise<any> {
+		let proc: any;
+		const p = new Promise<any>((resolve, reject) => {
+			this.downloadAsrModel(opts.model).then(m => {
+				if(opts.diarization) {
+					let output = join(Path.dirname(file), 'transcript');
+					proc = spawn(<string>this.ai.options.whisper,
+						['-m', m, '-f', file, '-np', '-ml', '1', '-oj', '-of', output],
+						{stdio: ['ignore', 'ignore', 'pipe']}
+					);
+					proc.on('error', (err: Error) => reject(err));
+					proc.on('close', async (code: number) => {
+						if(code === 0) {
+							output = await fs.readFile(output + '.json', 'utf-8');
+							fs.rm(output + '.json').catch(() => { });
+							try { resolve(JSON.parse(output)); }
+							catch(e) { reject(new Error('Failed to parse whisper JSON')); }
+						} else {
+							reject(new Error(`Exit code ${code}`));
+						}
+					});
+				} else {
 					let output = '';
-			const proc = spawn(<string>this.ai.options.whisper?.binary, ['-nt', '-np', '-m', m, '-f', path], {stdio: ['ignore', 'pipe', 'ignore']});
-			abort = () => proc.kill('SIGTERM');
+					proc = spawn(<string>this.ai.options.whisper, ['-m', m, '-f', file, '-np', '-nt']);
 					proc.on('error', (err: Error) => reject(err));
 					proc.stdout.on('data', (data: Buffer) => output += data.toString());
+					proc.on('close', async (code: number) => {
+						if(code === 0) {
+							resolve(output.trim() || null);
+						} else {
+							reject(new Error(`Exit code ${code}`));
+						}
+					});
+				}
+			});
+		});
+		return <any>Object.assign(p, {abort: () => proc?.kill('SIGTERM')});
+	}
+
+	private runDiarization(file: string): AbortablePromise<any> {
+		let aborted = false, abort = () => { aborted = true; };
+		const checkPython = (cmd: string) => {
+			return new Promise<boolean>((resolve) => {
+				const proc = spawn(cmd, ['-W', 'ignore', '-c', 'import pyannote.audio']);
+				proc.on('close', (code: number) => resolve(code === 0));
+				proc.on('error', () => resolve(false));
+			});
+		};
+		const p = Promise.all<any>([
+			checkPython('python'),
+			checkPython('python3'),
+		]).then(<any>(async ([p, p3]: [boolean, boolean]) => {
+			if(aborted) return;
+			if(!p && !p3) throw new Error('Pyannote is not installed: pip install pyannote.audio');
+			const binary = p3 ? 'python3' : 'python';
+			return new Promise((resolve, reject) => {
+				if(aborted) return;
+				let output = '';
+				const proc = spawn(binary, ['-W', 'ignore', '-c', this.pyannote, file]);
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.stderr.on('data', (data: Buffer) => console.error(data.toString()));
 				proc.on('close', (code: number) => {
-				if(code === 0) resolve(output.trim() || null);
-				else reject(new Error(`Exit code ${code}`));
+					if(code === 0) {
+						try { resolve(JSON.parse(output)); }
+						catch (err) { reject(new Error('Failed to parse diarization output')); }
+					} else {
+						reject(new Error(`Python process exited with code ${code}`));
+					}
 				});
+				proc.on('error', reject);
+				abort = () => proc.kill('SIGTERM');
 			});
-		return Object.assign(p, {abort});
+		}));
+		return <any>Object.assign(p, {abort});
+	}
+
+	asr(path: string, options: { model?: string; diarization?: boolean | 'llm' } = {}): AbortablePromise<string | null> {
+		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
+
+		const tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
+		execSync(`ffmpeg -i "${path}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
+		const clean = () => fs.rm(Path.dirname(tmp), {recursive: true, force: true}).catch(() => {});
+
+		if(!options.diarization) return this.runAsr(tmp, {model: options.model});
+		const timestamps = this.runAsr(tmp, {model: options.model, diarization: true});
+		const diarization = this.runDiarization(tmp);
+		let aborted = false, abort = () => {
+			aborted = true;
+			timestamps.abort();
+			diarization.abort();
+			clean();
+		};
+
+		const response = Promise.allSettled([timestamps, diarization]).then(async ([ts, d]) => {
+			if(ts.status == 'rejected') throw new Error('Whisper.cpp timestamps:\n' + ts.reason);
+			if(d.status == 'rejected') throw new Error('Pyannote:\n' + d.reason);
+			if(aborted || !options.diarization) return ts.value;
+			return this.diarizeTranscript(ts.value, d.value, options.diarization == 'llm');
+		}).finally(() => clean());
+		return <any>Object.assign(response, {abort});
 	}
 
 	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
-		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
+		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
 		if(!model.endsWith('.bin')) model += '.bin';
 		const p = Path.join(<string>this.ai.options.path, model);
 		if(await fs.stat(p).then(() => true).catch(() => false)) return p;

src/embedder.ts

@@ -1,14 +1,13 @@
-import { pipeline } from '@xenova/transformers';
-import { parentPort } from 'worker_threads';
+import { pipeline } from '@huggingface/transformers';
 
-let embedder: any;
+const [modelDir, model] = process.argv.slice(2);
 
-parentPort?.on('message', async ({ id, text, model, path }) => {
-	if(!embedder) embedder = await pipeline('feature-extraction', 'Xenova/' + model, {
-		quantized: true,
-		cache_dir: path,
-	});
+let text = '';
+process.stdin.on('data', chunk => text += chunk);
+process.stdin.on('end', async () => {
+	const embedder = await pipeline('feature-extraction', 'Xenova/' + model, {cache_dir: modelDir});
 	const output = await embedder(text, { pooling: 'mean', normalize: true });
 	const embedding = Array.from(output.data);
-	parentPort?.postMessage({ id, embedding });
+	process.stdout.write(JSON.stringify({embedding}));
+	process.exit();
 });

src/index.ts

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

src/llm.ts

@@ -1,12 +1,12 @@
-import {JSONAttemptParse} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
 import {OpenAi} from './open-ai.ts';
 import {LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
-import {Worker} from 'worker_threads';
 import {fileURLToPath} from 'url';
 import {dirname, join} from 'path';
+import {spawn} from 'node:child_process';
+import {Memory, MemoryManager} from './memory.ts';
 
 export type AnthropicConfig = {proto: 'anthropic', token: string};
 export type OllamaConfig = {proto: 'ollama', host: string};
@@ -36,18 +36,6 @@ export type LLMMessage = {
 	timestamp?: number;
 }
 
-/** Background information the AI will be fed */
-export type LLMMemory = {
-	/** What entity is this fact about */
-	owner: string;
-	/** The information that will be remembered */
-	fact: string;
-	/** Owner and fact embedding vector */
-	embeddings: [number[], number[]];
-	/** Creation time */
-	timestamp: Date;
-}
-
 export type LLMRequest = {
 	/** System prompt */
 	system?: string;
@@ -64,33 +52,43 @@ export type LLMRequest = {
 	/** Stream response */
 	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[],
+	compress?: {max: number; min: number};
+	/** User's memory documents - RAG injected automatically each turn */
+	memory?: Memory[];
+	/** Model to use for memory operations */
+	memoryModel?: string;
+	/** Skill documents the AI can browse and read on demand */
+	skills?: Skill[];
+	/** MCP servers to connect and expose as tools */
+	mcp?: McpServer[];
 }
 
+export type McpServer = {
+	/** MCP server name for humans */
+	name: string;
+	/** Host URL */
+	host: string;
+	/** Server access token */
+	token?: string;
+}
+
+export type Skill = {
+	/** Name of skill for humans */
+	name: string;
+	/** Description LLM will use to decide to learn a skill */
+	description: string;
+	/** Skill instructions */
+	content: string;
+}
+
+
 class LLM {
-	private embedWorker: Worker | null = null;
-	private embedQueue = new Map<number, { resolve: (value: number[]) => void; reject: (error: any) => void }>();
-	private embedId = 0;
-	private models: {[model: string]: LLMProvider} = {};
-	private defaultModel!: string;
+	private memoryManager!: MemoryManager;
+
+	defaultModel!: string;
+	models: {[model: string]: LLMProvider} = {};
 
 	constructor(public readonly ai: Ai) {
-		this.embedWorker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
-		this.embedWorker.on('message', ({ id, embedding }) => {
-			const pending = this.embedQueue.get(id);
-			if (pending) {
-				pending.resolve(embedding);
-				this.embedQueue.delete(id);
-			}
-		});
-
 		if(!ai.options.llm?.models) return;
 		Object.entries(ai.options.llm.models).forEach(([model, config]) => {
 			if(!this.defaultModel) this.defaultModel = model;
@@ -98,84 +96,143 @@ class LLM {
 			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);
 		});
+		this.memoryManager = new MemoryManager(this);
+	}
+
+	private async setupMcp(servers: McpServer[] = []): Promise<{prompt: string, tools: AiTool[]}> {
+		if(!servers?.length) return {prompt: '', tools: []};
+		const allTools: AiTool[] = [];
+		await Promise.all(servers.map(async server => {
+			const res = await fetch(`${server.host}/tools`, {headers: server.token ? {Authorization: `Bearer ${server.token}`} : {}});
+			const mcp: any = await res.json();
+			if(!mcp?.tools) return;
+			for(const t of mcp.tools) {
+				const args: Record<string, any> = {};
+				if(t.inputSchema?.properties) {
+					for(const [key, val] of Object.entries<any>(t.inputSchema.properties)) {
+						args[key] = {type: val.type || 'string', description: val.description || '', required: t.inputSchema.required?.includes(key)};
+					}
+				}
+				allTools.push({
+					name: `${server.name}_${t.name}`,
+					description: t.description || '',
+					args,
+					fn: async (a: any) => {
+						const r = await fetch(`${server.host}/tools/call`, {
+							method: 'POST',
+							headers: {'Content-Type': 'application/json', ...(server.token ? {Authorization: `Bearer ${server.token}`} : {})},
+							body: JSON.stringify({name: t.name, arguments: a})
+						});
+						const data: any = await r.json();
+						return data?.content?.[0]?.text ?? JSON.stringify(data);
+					}
+				});
+			}
+		}));
+
+		const list = allTools.map(t => `- ${t.name}: ${t.description}`).join('\n');
+		return {
+			prompt: `You have access to the following MCP tools:\n${list}`,
+			tools: allTools
+		};
+	}
+
+	private setupSkills(skills: Skill[] = []): {prompt: string, tools: AiTool[]} {
+		if(!skills?.length) return {prompt: '', tools: []};
+		const list = skills.map(s => `- ${s.name}: ${s.description}`).join('\n');
+		return {
+			prompt: `You have access to the following skill documents, use \`read_skill\` to access them:\n${list}`,
+			tools: [{
+				name: 'read_skill',
+				description: 'Read the full content of a skill/knowledge document',
+				args: {
+					name: {type: 'string', description: 'Exact skill name', required: true}
+				},
+				fn: (args: any) => {
+					const skill = skills.find(s => s.name === args.name);
+					if(!skill) return `Skill not found. Available:\n${list}`;
+					return `# ${skill.name}\n${skill.content}`;
+				}
+			}]
+		}
 	}
 
-	/**
-	 * Chat with LLM
-	 * @param {string} message Question
-	 * @param {LLMRequest} options Configuration options and chat history
-	 * @returns {{abort: () => void, response: Promise<string>}} Function to abort response and chat history
-	 */
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
+		options = <any>{
+			system: '',
+			temperature: 0.8,
+			...this.ai.options.llm,
+			models: undefined,
+			history: [],
+			...options,
+		}
 		const m = options.model || this.defaultModel;
 		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
 		let abort = () => {};
 		return Object.assign(new Promise<string>(async res => {
-			if(!options.history) options.history = [];
-			// If memories were passed, find any relivant ones and add a tool for ADHOC lookups
+			let tools: AiTool[] = options.tools || this.ai.options.llm?.tools || [];
+			const prompts: string[] = [];
+			let history = options.history || [];
+
+			// MCP
+			const mcp = options.mcp || this.ai.options?.llm?.mcp;
+			if(mcp?.length) {
+				const m = await this.setupMcp(mcp);
+				prompts.unshift(m.prompt);
+				tools.push(...m.tools);
+			}
+
+			// Skills
+			const skills = options.skills || this.ai.options?.llm?.skills;
+			if(skills?.length) {
+				const s = this.setupSkills(skills);
+				prompts.unshift(s.prompt);
+				tools.push(...s.tools);
+			}
+
+			// Memory
 			if(options.memory) {
-				options.system = (options.system || '') + '\nYou have passive persistent memory never make any mention of your memory capabilities and what you can/cannot remember\n';
-				const search = async (query?: string | null, subject?: string | null, limit = 50) => {
-					const [o, q] = await Promise.all([
-						subject ? this.embedding(subject) : Promise.resolve(null),
-						query ? this.embedding(query) : Promise.resolve(null),
-					]);
-					return (options.memory || [])
-						.map(m => ({...m, score: o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 1}))
-						.filter((m: any) => m.score >= 0.8)
-						.map((m: any) => ({...m, score: q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : m.score}))
-						.filter((m: any) => m.score >= 0.2)
-						.toSorted((a: any, b: any) => a.score - b.score)
-						.slice(0, limit);
+				const relevant = await this.memoryManager.recollect(message, options.memory, 1);
+				prompts.unshift(`You have access to the following memory files:
+${options.memory.map(m => `- ${m.name}: ${m.description}`).join('\n')}
+${relevant.length ? `
+The closest memory has been added primitively:
+\`\`\`
+Name: ${relevant[0].name}
+Description: ${relevant[0].description}
+${relevant[0].content}
+\`\`\`
+`: ''}`.trim());
+				tools.push(this.memoryManager.tools.read(<Memory[]>options.memory));
 			}
 
-				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);
-					}
-				}];
-			}
+			prompts.unshift(options.system || this.ai.options.llm?.system || '');
+			const resp = await this.models[m].ask(message, {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
 
-			// Ask
-			const resp = await this.models[m].ask(message, options);
-
-			// Remove any memory calls
+			// Trim memory injections from history
 			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);
+				history.splice(0, history.length, ...history.filter(h => h.role !== 'tool' || h.name !== 'recall'));
 			}
 
-			// 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);
-				}
+			// Auto-memorize before compressing
+			if(options.compress && this.estimateTokens(history) >= options.compress.max) {
+				if(options.memory) await this.memoryManager.memorize(history, options.memory, options);
+				const compressed = await this.compressHistory(history, options.compress.max, options.compress.min, options);
+				if(options.history) options.history.splice(0, options.history.length, ...compressed);
 			}
+
 			return res(resp);
 		}), {abort});
 	}
 
+	/**
+	 * Digest full conversation history into memory documents.
+	 * Call on session end to persist the conversation.
+	 */
+	async updateMemory(history: LLMMessage[], memories: Memory[], options: LLMRequest = {}): Promise<void> {
+		await this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
+	}
+
 	/**
 	 * Compress chat history to reduce context size
 	 * @param {LLMMessage[]} history Chatlog that will be compressed
@@ -184,27 +241,24 @@ class LLM {
 	 * @param {LLMRequest} options LLM options
 	 * @returns {Promise<LLMMessage[]>} New chat history will summary at index 0
 	 */
-	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<{history: LLMMessage[], memory: LLMMemory[]}> {
-		if(this.estimateTokens(history) < max) return {history, memory: []};
+	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()) {
 			tokens += this.estimateTokens(m.content);
 			if(tokens < min) keep++;
 			else break;
 		}
-		if(history.length <= keep) return {history, memory: []};
+		if(history.length <= keep) return history;
 		const system = history[0].role == 'system' ? history[0] : null,
 			recent = keep == 0 ? [] : history.slice(-keep),
 			process = (keep == 0 ? history : history.slice(0, -keep)).filter(h => h.role === 'assistant' || h.role === 'user');
-		const summary: any = await this.json(`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. Match this format: {summary: string, facts: [[subject, fact]]}\n\n${process.map(m => `${m.role}: ${m.content}`).join('\n\n')}`, {model: options?.model, temperature: options?.temperature || 0.3});
-		const timestamp = new Date();
-		const memory = await Promise.all((summary?.facts || [])?.map(async ([owner, fact]: [string, string]) => {
-			const e = await Promise.all([this.embedding(owner), this.embedding(`${owner}: ${fact}`)]);
-			return {owner, fact, embeddings: [e[0][0].embedding, e[1][0].embedding], timestamp};
-		}));
-		const h = [{role: 'assistant', content: `Conversation Summary: ${summary?.summary}`, timestamp: Date.now()}, ...recent];
+
+		const summary: any = await this.summarize(process.map(m => `[${m.role}]: ${m.content}`).join('\n\n'), 500, options);
+		const d = Date.now();
+		const h = [{role: <any>'tool', name: 'summary', id: `summary_` + d, args: {}, content: `Conversation Summary: ${summary?.summary}`, timestamp: d}, ...recent];
 		if(system) h.splice(0, 0, system);
-		return {history: <any>h, memory};
+		return h;
 	}
 
 	/**
@@ -241,7 +295,7 @@ class LLM {
 				return `${p}: ${Array.isArray(value) ? value.join(', ') : value}`;
 			});
 		};
-		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
+		const lines = typeof target === 'object' ? objString(target) : target.toString().split('\n');
 		const tokens = lines.flatMap(l => [...l.split(/\s+/).filter(Boolean), '\n']);
 		const chunks: string[] = [];
 		for(let i = 0; i < tokens.length;) {
@@ -262,30 +316,55 @@ class LLM {
 	/**
 	 * Create a vector representation of a string
 	 * @param {object | string} target Item that will be embedded (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)
+	 * @param {maxTokens?: number, overlapTokens?: number} opts Options for embedding such as chunk sizes
 	 * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
 	 */
-	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+	embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}): AbortablePromise<{index: number, embedding: number[], text: string, tokens: number}[]> {
+		let {maxTokens = 500, overlapTokens = 50} = opts;
+		let aborted = false;
+		const abort = () => { aborted = true; };
+
 		const embed = (text: string): Promise<number[]> => {
 			return new Promise((resolve, reject) => {
-				const id = this.embedId++;
-				this.embedQueue.set(id, { resolve, reject });
-				this.embedWorker?.postMessage({
-					id,
-					text,
-					model: this.ai.options?.embedder || 'bge-small-en-v1.5',
-					path: this.ai.options.path
+				if(aborted) return reject(new Error('Aborted'));
+				const args: string[] = [
+					join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'),
+					<string>this.ai.options.path,
+					this.ai.options?.embedder || 'bge-small-en-v1.5'
+				];
+				const proc = spawn('node', args, {stdio: ['pipe', 'pipe', 'ignore']});
+				proc.stdin.write(text);
+				proc.stdin.end();
+				let output = '';
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.on('close', (code: number) => {
+					if(aborted) return reject(new Error('Aborted'));
+					if(code === 0) {
+						try {
+							const result = JSON.parse(output);
+							resolve(result.embedding);
+						} catch(err) {
+							reject(err);
+						}
+					} else {
+						reject(new Error(`Embedder process exited with code ${code}`));
+					}
 				});
+				proc.on('error', reject);
 			});
 		};
-		const chunks = this.chunk(target, maxTokens, overlapTokens);
-		return Promise.all(chunks.map(async (text, index) => ({
-			index,
-			embedding: await embed(text),
-			text,
-			tokens: this.estimateTokens(text),
-		})));
+
+		const p = (async () => {
+			const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
+			for(let i = 0; i < chunks.length; i++) {
+				if(aborted) break;
+				const text = chunks[i];
+				const embedding = await embed(text);
+				results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
+			}
+			return results;
+		})();
+		return <any>Object.assign(p, {abort});
 	}
 
 	/**
@@ -311,33 +390,101 @@ class LLM {
 				(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}
+		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): Promise<any> {
-		let resp = await this.ask(message, {system: 'Respond using a JSON blob matching any provided examples', ...options});
-		if(!resp) return {};
-		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
-		const jsonStr = codeBlock ? codeBlock[1].trim() : resp;
-		return JSONAttemptParse(jsonStr, {});
+	async json(text: string, schema: string, options?: LLMRequest): Promise<any> {
+		let system = `Your job is to convert input to JSON using tool calls. Call the \`submit\` tool at least once with JSON matching this schema:\n\`\`\`json\n${schema}\n\`\`\`\n\nResponses are ignored`;
+		if(options?.system) system += '\n\n' + options.system;
+		return new Promise(async (resolve, reject) => {
+			let done = false;
+			const resp = await this.ask(text, {
+				temperature: 0.3,
+				...options,
+				system,
+				tools: [{
+					name: 'submit',
+					description: 'Submit JSON',
+					args: {json: {type: 'string', description: 'Javascript parsable JSON string', required: true}},
+					fn: (args) => {
+						try {
+							const json = JSON.parse(args.json);
+							resolve(json);
+							done = true;
+						} catch { return 'Invalid JSON'; }
+						return 'Saved';
+					}
+				}, ...(options?.tools || [])],
+			});
+			if(!done) reject(`AI failed to create JSON:\n${resp}`);
+		});
 	}
 
 	/**
 	 * Create a summary of some text
 	 * @param {string} text Text to summarize
-	 * @param {number} tokens Max number of tokens
+	 * @param {number} length Max number of words
 	 * @param options LLM request options
 	 * @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});
+	async summarize(text: string, length: number = 500, options?: LLMRequest): Promise<string | null> {
+		let system = `Your job is to summarize the users message using tool calls. Call the \`submit\` tool at least once with the shortest summary possible that's <= ${length} words. The tool call will respond with the token count. Responses are ignored`;
+		if(options?.system) system += '\n\n' + options.system;
+		return new Promise(async (resolve, reject) => {
+			let done = false;
+			const resp = await this.ask(text, {
+				temperature: 0.3,
+				...options,
+				system,
+				tools: [{
+					name: 'submit',
+					description: 'Submit summary',
+					args: {summary: {type: 'string', description: 'Text summarization', required: true}},
+					fn: (args) => {
+						if(!args.summary) return 'No summary provided';
+						const count = args.summary.split(' ').length;
+						if(count > length) return `Too long: ${length} words`;
+						done = true;
+						resolve(args.summary || null);
+						return `Saved: ${length} words`;
+					}
+				}, ...(options?.tools || [])],
+			});
+			if(!done) reject(`AI failed to create summary:\n${resp}`);
+		});
+	}
+
+	addModel(name: string, config: AnthropicConfig | OllamaConfig | OpenAiConfig, setDefault = false) {
+		if(config.proto == 'anthropic') this.models[name] = new Anthropic(this.ai, config.token, name);
+		else if(config.proto == 'ollama') this.models[name] = new OpenAi(this.ai, config.host, 'not-needed', name);
+		else if(config.proto == 'openai') this.models[name] = new OpenAi(this.ai, config.host || null, config.token, name);
+		if(setDefault || !this.defaultModel) this.defaultModel = name;
+	}
+
+	removeModel(name: string) {
+		delete this.models[name];
+		if(this.defaultModel === name) {
+			this.defaultModel = Object.keys(this.models)[0] ?? '';
+		}
+	}
+
+	setModels(models: {[model: string]: AnthropicConfig | OllamaConfig | OpenAiConfig}, replace = true) {
+		if(replace) this.models = {};
+		Object.entries(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);
+		});
+		this.defaultModel = Object.keys(this.models)[0] ?? '';
 	}
 }
 

src/memory.ts

@@ -0,0 +1,177 @@
+// memory.ts
+import {LLMRequest, LLMMessage} from './llm.ts';
+import {AiTool} from './tools.ts';
+
+/** Background information the AI will be fed as a knowledge document */
+export type Memory = {
+	/** Memory subject */
+	name: string;
+	/** Short description of what this document contains - used for RAG retrieval */
+	description: string;
+	/** Full markdown content of the document */
+	content: string;
+	/** Embedding vector of the description - used for similarity search */
+	embedding: number[];
+}
+
+export type MemoryCollection = {
+	/** Memory subject */
+	name: string;
+	/** Short description - required if isNew */
+	description?: string;
+	/** Extracted facts to merge */
+	facts: string[];
+}
+
+export class MemoryManager {
+
+	tools = {
+		edit: (memory: Memory): AiTool => ({
+			name: 'edit_memory',
+			description: 'Edit a memory. Omit start/end to append. Pass start only to replace from that line on (Note line 0 = first line of content / line AFTER description). Pass start+end to replace a specific range. start=0 replaces the whole document. Returns updated document',
+			args: {
+				content: {type: 'string', description: 'New content', required: true},
+				start: {type: 'number', description: 'First line to replace (0-indexed, inclusive). Omit to append.'},
+				end: {type: 'number', description: 'Last line to replace (0-indexed, inclusive). Omit to replace from start to end of doc.'},
+			},
+			fn: (args: any) => {
+				const lines = memory.content ? memory.content.split('\n') : [];
+				const newLines = args.content.split('\n');
+				if(args.start === undefined) lines.push(...newLines);
+				else if(args.end === undefined) lines.splice(args.start, lines.length - args.start, ...newLines);
+				else lines.splice(args.start, args.end - args.start + 1, ...newLines);
+				memory.content = lines.join('\n');
+				return memory.content;
+			}
+		}),
+		extract: (pools: MemoryCollection[]): AiTool => ({
+			name: 'extract_facts',
+			description: 'Extract a list of facts to group into a single memory',
+			args: {
+				name: {type: 'string', description: 'Exact name of an existing memory, or a new name if none fits ([pro]nouns only)', required: true},
+				description: {type: 'string', description: 'One sentence description of the memory subject', required: true},
+				facts: {type: 'string', description: 'Comma separated list of extracted facts', required: true},
+			},
+			fn: (args: any) => {
+				pools.push({
+					name: args.name,
+					description: args.description,
+					facts: args.facts.split(',').map((f: string) => f.trim()).filter(Boolean),
+				});
+				return 'Success';
+			}}),
+		read: (memories: Memory[]): AiTool => ({
+			name: 'read_memory',
+			description: 'Read entire memory',
+			args: {
+				name: {type: 'string', description: 'Exact memory name', required: true},
+			},
+			fn: (args: any) => {
+				const mem = memories.find(m => m.name === args.name);
+				if(!mem) return 'Document not found';
+				return `Name: ${mem.name}\nDescription: ${mem.description}\n\n${mem.content}`;
+			}
+		}),
+	}
+
+	constructor(private llm: any, private model?: string) {}
+
+	/**
+	 * Extracts facts from conversation and groups them into individual memories
+	 * @param {string} conversation Full conversation formatted as [role]: content
+	 * @param {Memory[]} memories The user's memory documents
+	 * @param {LLMRequest} options LLM options
+	 * @returns {Promise<MemoryCollection[]>} Fact pools grouped by target document
+	 */
+	private async extract(conversation: string, memories: Memory[], options: LLMRequest): Promise<MemoryCollection[]> {
+		const existingDocs = memories.map(m => `Name: ${m.name}\nDescription: ${m.description}`).join('\n\n');
+		const pools: MemoryCollection[] = [];
+		await this.llm.ask(conversation, {
+			model: this.model || options.model,
+			temperature: 0.2,
+			system: `You are a fact extractor. Analyze this conversation and extract facts worth remembering long term.
+Rules:
+- ONLY extract facts the USER explicitly stated about themselves or their business
+- ONLY extract decisions that were MADE during this conversation
+- DO NOT extract anything the AI said, its name, capabilities, or how it introduced itself
+- DO NOT extract greetings, pleasantries or generic exchanges
+- If nothing worth remembering was said, dont do anything, skip calling tools
+
+For each fact decide whether it belongs in an existing document or needs a new one, then call the \`extract_facts\` tool.
+
+Existing documents:\n${existingDocs || 'None yet.'}`,
+			tools: [this.tools.extract(pools)]
+		});
+		return pools;
+	}
+
+	/**
+	 * Bot 2 - Editor: merges a pool of facts into a specific document using surgical line-based edits.
+	 * Receives full document content and uses read + amend tools to make precise edits.
+	 * @param {MemoryCollection} newMem The fact pool to merge
+	 * @param {Memory[]} memories The user's memory documents
+	 * @param {LLMRequest} options LLM options
+	 */
+	private async edit(newMem: MemoryCollection, memories: Memory[], options: LLMRequest): Promise<void> {
+		const existing = memories.find(m => m.name === newMem.name);
+		const mem: Memory = existing || {name: newMem.name, description: newMem.description || '', content: '', embedding: []};
+		const isNew = !existing;
+
+		await this.llm.ask(newMem.facts.map(f => `- ${f}`).join('\n'),
+			{
+				model: this.model || options.model,
+				temperature: 0.2,
+				system: `You are a document editor. Merge the users list of facts into the following document using the \`edit_memory\` tool; call it as many times as necessary:
+\`\`\`
+${mem.content}
+\`\`\``,
+				tools: [this.tools.edit(mem)]
+			}
+		);
+
+		if(isNew || mem.description !== existing?.description) {
+			const e = await this.llm.embedding(mem.description);
+			mem.embedding = e?.[0]?.embedding;
+		}
+
+		if(isNew) memories.push(mem);
+		else {
+			const idx = memories.findIndex(m => m.name === newMem.name);
+			if(idx >= 0) memories[idx] = mem;
+		}
+	}
+
+	/**
+	 * Find relevant memory documents for a query using description embeddings
+	 * @param {string} query The query to search against
+	 * @param {Memory[]} memories The user's memory documents
+	 * @param {number} limit Max number of results to return
+	 * @returns {Promise<Memory[]>} The most relevant memory documents
+	 */
+	async recollect(query: string, memories: Memory[], limit = 5): Promise<Memory[]> {
+		const [e] = await this.llm.embedding(query);
+		return memories
+			.filter(m => m.embedding?.length)
+			.map(m => ({...m, score: this.llm.cosineSimilarity(m.embedding, e.embedding)}))
+			.toSorted((a: any, b: any) => b.score - a.score)
+			.slice(0, limit);
+	}
+
+	/**
+	 * Two-stage memory pipeline: classify facts from conversation history then surgically merge them into documents.
+	 * Bot 1 (classify) extracts and groups facts cheaply. Bot 2 (edit) runs per-document in parallel with full content access.
+	 * @param {LLMMessage[]} history Full conversation history to digest
+	 * @param {Memory[]} memories The user's memory documents — mutated in place
+	 * @param {LLMRequest} options LLM options
+	 */
+	async memorize(history: LLMMessage[], memories: Memory[], options: LLMRequest): Promise<void> {
+		const conversation = history
+			.filter(h => h.role === 'user' || h.role === 'assistant')
+			.map(h => `[${h.role}]: ${h.content}`)
+			.join('\n\n');
+		if(!conversation.trim()) return;
+		const pools = await this.extract(conversation, memories, options);
+		if(!pools.length) return;
+		await Promise.all(pools.map(pool => this.edit(pool, memories, options)));
+	}
+}

src/open-ai.ts

@@ -11,7 +11,7 @@ export class OpenAi extends LLMProvider {
 		super();
 		this.client = new openAI(clean({
 			baseURL: host,
-			apiKey: token
+			apiKey: token || host ? 'ignored' : undefined
 		}));
 	}
 
@@ -67,8 +67,11 @@ export class OpenAi extends LLMProvider {
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		const controller = new AbortController();
 		return Object.assign(new Promise<any>(async (res, rej) => {
-			if(options.system && options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
-			const history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
+			if(options.system) {
+				if(options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
+				else options.history[0].content = options.system;
+			}
+			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
@@ -100,19 +103,42 @@ export class OpenAi extends LLMProvider {
 				if(options.stream) {
 					if(!isFirstMessage) options.stream({text: '\n\n'});
 					else isFirstMessage = false;
-					resp.choices = [{message: {content: '', tool_calls: []}}];
+					resp.choices = [{message: {role: 'assistant', content: '', tool_calls: []}}];
 					for await (const chunk of resp) {
 						if(controller.signal.aborted) break;
 						if(chunk.choices[0].delta.content) {
 							resp.choices[0].message.content += chunk.choices[0].delta.content;
 							options.stream({text: chunk.choices[0].delta.content});
 						}
+
 						if(chunk.choices[0].delta.tool_calls) {
-							resp.choices[0].message.tool_calls = chunk.choices[0].delta.tool_calls;
+							for(const deltaTC of chunk.choices[0].delta.tool_calls) {
+								const existing = resp.choices[0].message.tool_calls.find(tc => tc.index === deltaTC.index);
+								if(existing) {
+									if(deltaTC.id) existing.id = deltaTC.id;
+									if(deltaTC.type) existing.type = deltaTC.type;
+									if(deltaTC.function) {
+										if(!existing.function) existing.function = {};
+										if(deltaTC.function.name) existing.function.name = deltaTC.function.name;
+										if(deltaTC.function.arguments) existing.function.arguments = (existing.function.arguments || '') + deltaTC.function.arguments;
+									}
+								} else {
+									resp.choices[0].message.tool_calls.push({
+										index: deltaTC.index,
+										id: deltaTC.id || '',
+										type: deltaTC.type || 'function',
+										function: {
+											name: deltaTC.function?.name || '',
+											arguments: deltaTC.function?.arguments || ''
+										}
+									});
+								}
+							}
 						}
 					}
 				}
 
+				if(resp.error) throw new Error(resp.error);
 				const toolCalls = resp.choices[0].message.tool_calls || [];
 				if(toolCalls.length && !controller.signal.aborted) {
 					history.push(resp.choices[0].message);
@@ -123,7 +149,7 @@ export class OpenAi extends LLMProvider {
 						try {
 							const args = JSONAttemptParse(toolCall.function.arguments, {});
 							const result = await tool.fn(args, options.stream, this.ai);
-							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize(result)};
+							return {role: 'tool', tool_call_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
 						} catch (err: any) {
 							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'})};
 						}
@@ -132,8 +158,8 @@ 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 || ''});
-			this.toStandard(history);
+			history.push({role: 'assistant', content: resp.choices[0].message.content.trim() || ''});
+			history = this.toStandard(history);
 
 			if(options.stream) options.stream({done: true});
 			if(options.history) options.history.splice(0, options.history.length, ...history);

src/tools.ts

@@ -1,9 +1,17 @@
 import  * as cheerio from 'cheerio';
-import {$, $Sync} from '@ztimson/node-utils';
-import {ASet, consoleInterceptor, Http, fn as Fn} from '@ztimson/utils';
+import {$Sync} from '@ztimson/node-utils';
+import {ASet, consoleInterceptor, Http, fn as Fn, decodeHtml} from '@ztimson/utils';
+import * as os from 'node:os';
 import {Ai} from './ai.ts';
 import {LLMRequest} from './llm.ts';
 
+const UA = 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)';
+
+const getShell = () => {
+	if(os.platform() == 'win32') return 'cmd';
+	return $Sync`echo $SHELL`?.split('/').pop() || 'bash';
+}
+
 export type AiToolArg = {[key: string]: {
 	/** Argument type */
 	type: 'array' | 'boolean' | 'number' | 'object' | 'string',
@@ -40,11 +48,18 @@ export const CliTool: AiTool = {
 	name: 'cli',
 	description: 'Use the command line interface, returns any output',
 	args: {command: {type: 'string', description: 'Command to run', required: true}},
-	fn: (args: {command: string}) => $`${args.command}`
+	fn: (args: {command: string}) => $Sync`${args.command}`
 }
 
 export const DateTimeTool: AiTool = {
 	name: 'get_datetime',
+	description: 'Get local date / time',
+	args: {},
+	fn: async () => new Date().toString()
+}
+
+export const DateTimeUTCTool: AiTool = {
+	name: 'get_datetime_utc',
 	description: 'Get current UTC date / time',
 	args: {},
 	fn: async () => new Date().toUTCString()
@@ -54,19 +69,20 @@ export const ExecTool: AiTool = {
 	name: 'exec',
 	description: 'Run code/scripts',
 	args: {
-		language: {type: 'string', description: 'Execution language', enum: ['cli', 'node', 'python'], required: true},
+		language: {type: 'string', description: `Execution language (CLI: ${getShell()})`, enum: ['cli', 'node', 'python'], required: true},
 		code: {type: 'string', description: 'Code to execute', required: true}
 	},
 	fn: async (args, stream, ai) => {
 		try {
-			switch(args.type) {
-				case 'bash':
+			switch(args.language) {
+				case 'cli':
 					return await CliTool.fn({command: args.code}, stream, ai);
 				case 'node':
 					return await JSTool.fn({code: args.code}, stream, ai);
-				case 'python': {
+				case 'python':
 					return await PythonTool.fn({code: args.code}, stream, ai);
-				}
+				default:
+					throw new Error(`Unsupported language: ${args.language}`);
 			}
 		} catch(err: any) {
 			return {error: err?.message || err.toString()};
@@ -98,9 +114,9 @@ export const JSTool: AiTool = {
 		code: {type: 'string', description: 'CommonJS javascript', required: true}
 	},
 	fn: async (args: {code: string}) => {
-		const console = consoleInterceptor(null);
-		const resp = await Fn<any>({console}, args.code, true).catch((err: any) => console.output.error.push(err));
-		return {...console.output, return: resp, stdout: undefined, stderr: undefined};
+		const c = consoleInterceptor(null);
+		const resp = await Fn<any>({console: c}, args.code, true).catch((err: any) => c.output.error.push(err));
+		return {...c.output, return: resp, stdout: undefined, stderr: undefined};
 	}
 }
 
@@ -115,37 +131,107 @@ export const PythonTool: AiTool = {
 
 export const ReadWebpageTool: AiTool = {
 	name: 'read_webpage',
-	description: 'Extract clean, structured content from a webpage. Use after web_search to read specific URLs',
+	description: 'Extract clean content from webpages, or convert media/documents to accessible formats',
 	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")'}
+		url: {type: 'string', description: 'URL to read', required: true},
+		mimeRegex: {type: 'string', description: 'Optional regex to filter MIME types (e.g., "^image/", "text/")'}
 	},
-	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}`)});
+	fn: async (args: {url: string; mimeRegex?: string}) => {
+		const ua = 'AiTools-Webpage/1.0';
+		const maxSize = 10 * 1024 * 1024;
 
+		const response = await fetch(args.url, {
+			headers: {
+				'User-Agent': ua,
+				'Accept': 'text/html,application/xhtml+xml,application/xml;q=0.9,image/webp,*/*;q=0.8',
+				'Accept-Language': 'en-US,en;q=0.5'
+			},
+			redirect: 'follow'
+		}).catch(err => {throw new Error(`Failed to fetch: ${err.message}`)});
+
+		const contentType = response.headers.get('content-type') || '';
+		const mimeType = contentType.split(';')[0].trim().toLowerCase();
+
+		if(args.mimeRegex && !new RegExp(args.mimeRegex, 'i').test(mimeType)) {
+			return `❌ MIME type rejected: ${mimeType} (filter: ${args.mimeRegex})`;
+		}
+
+		if(mimeType.match(/^(image|audio|video)\//)) {
+			const buffer = await response.arrayBuffer();
+			if(buffer.byteLength > maxSize) {
+				return `❌ File too large: ${(buffer.byteLength / 1024 / 1024).toFixed(1)}MB (max 10MB)\nType: ${mimeType}`;
+			}
+			const base64 = Buffer.from(buffer).toString('base64');
+			return `## Media File\n**Type:** ${mimeType}\n**Size:** ${(buffer.byteLength / 1024).toFixed(1)}KB\n**Data URL:** \`data:${mimeType};base64,${base64.slice(0, 100)}...\``;
+		}
+
+		if(mimeType.match(/^text\/(plain|csv|xml)/) || args.url.match(/\.(txt|csv|xml|md|yaml|yml)$/i)) {
+			const text = await response.text();
+			const truncated = text.length > 50000 ? text.slice(0, 50000) : text;
+			return `## Text File\n**Type:** ${mimeType}\n**URL:** ${args.url}\n\n${truncated}`;
+		}
+
+		if(mimeType.match(/application\/(json|xml|csv)/)) {
+			const text = await response.text();
+			const truncated = text.length > 50000 ? text.slice(0, 50000) : text;
+			return `## Structured Data\n**Type:** ${mimeType}\n**URL:** ${args.url}\n\n\`\`\`\n${truncated}\n\`\`\``;
+		}
+
+		if(mimeType === 'application/pdf' || (mimeType.startsWith('application/') && !mimeType.includes('html'))) {
+			const buffer = await response.arrayBuffer();
+			if(buffer.byteLength > maxSize) {
+				return `❌ File too large: ${(buffer.byteLength / 1024 / 1024).toFixed(1)}MB (max 10MB)\nType: ${mimeType}`;
+			}
+			const base64 = Buffer.from(buffer).toString('base64');
+			return `## Binary File\n**Type:** ${mimeType}\n**Size:** ${(buffer.byteLength / 1024).toFixed(1)}KB\n**Data URL:** \`data:${mimeType};base64,${base64.slice(0, 100)}...\``;
+		}
+
+		// HTML
+		const html = await response.text();
 		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') || '',
-		};
-
+		$('script, style, nav, footer, header, aside, iframe, noscript, svg').remove();
+		$('[role="navigation"], [role="banner"], [role="complementary"]').remove();
+		$('[aria-hidden="true"], [hidden], .visually-hidden, .sr-only, .screen-reader-text').remove();
+		$('.ad, .ads, .advertisement, .cookie, .popup, .modal, .sidebar, .related, .comments, .social-share').remove();
+		$('button, [class*="share"], [class*="follow"], [class*="social"]').remove();
+		const title = $('meta[property="og:title"]').attr('content') || $('title').text().trim() || '';
+		const description = $('meta[name="description"]').attr('content') || $('meta[property="og:description"]').attr('content') || '';
+		const author = $('meta[name="author"]').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();
+		const selectors = ['article', 'main', '[role="main"]', '.content', '.post-content', '.entry-content', '.article-content'];
+		for(const sel of selectors) {
+			const el = $(sel).first();
+			if(el.length && el.text().trim().length > 200) {
+				const paragraphs: string[] = [];
+				el.find('p').each((_, p) => {
+					const text = $(p).text().trim();
+					if(text.length > 80) paragraphs.push(text);
+				});
+				if(paragraphs.length > 2) {
+					content = paragraphs.join('\n\n');
 					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};
 		}
-}
+
+		if(!content) {
+			const paragraphs: string[] = [];
+			$('body p').each((_, p) => {
+				const text = $(p).text().trim();
+				if(text.length > 80) paragraphs.push(text);
+			});
+			content = paragraphs.slice(0, 30).join('\n\n');
+		}
+
+		// Decode escaped newlines and clean
+		const parts = [`## ${title || 'Webpage'}`];
+		if(description) parts.push(`_${description}_`);
+		if(author) parts.push(`👤 ${author}`);
+		parts.push(`🔗 ${args.url}\n`);
+		parts.push(content);
+		return decodeHtml(parts.join('\n\n').replaceAll(/\n{3,}/g, '\n\n'));
+	}
+};
 
 export const WebSearchTool: AiTool = {
 	name: 'web_search',
@@ -159,7 +245,7 @@ export const WebSearchTool: AiTool = {
 		length: number;
 	}) => {
 		const html = await fetch(`https://html.duckduckgo.com/html/?q=${encodeURIComponent(args.query)}`, {
-			headers: {"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64)", "Accept-Language": "en-US,en;q=0.9"}
+			headers: {"User-Agent": UA, "Accept-Language": "en-US,en;q=0.9"}
 		}).then(resp => resp.text());
 		let match, regex = /<a .*?href="(.+?)".+?<\/a>/g;
 		const results = new ASet<string>();
@@ -172,3 +258,94 @@ export const WebSearchTool: AiTool = {
 		return results;
 	}
 }
+
+class WikipediaClient {
+	private async get(url: string): Promise<any> {
+		const resp = await fetch(url, {headers: {'User-Agent': UA}});
+		return resp.json();
+	}
+
+	private api(params: Record<string, any>): Promise<any> {
+		const qs = new URLSearchParams({...params, format: 'json', utf8: '1'}).toString();
+		return this.get(`https://en.wikipedia.org/w/api.php?${qs}`);
+	}
+
+	private clean(text: string): string {
+		return text.replace(/\n{3,}/g, '\n\n').replace(/ {2,}/g, ' ').replace(/\[\d+\]/g, '').trim();
+	}
+
+	private truncate(text: string, max: number): string {
+		if(text.length <= max) return text;
+		const cut = text.slice(0, max);
+		const lastPara = cut.lastIndexOf('\n\n');
+		return lastPara > max * 0.7 ? cut.slice(0, lastPara) : cut;
+	}
+
+	private async searchTitles(query: string, limit = 6): Promise<any[]> {
+		const data = await this.api({action: 'query', list: 'search', srsearch: query, srlimit: limit, srprop: 'snippet'});
+		return data.query?.search || [];
+	}
+
+	private async fetchExtract(title: string, intro = false): Promise<string> {
+		const params: any = {action: 'query', prop: 'extracts', titles: title, explaintext: 1, redirects: 1};
+		if(intro) params.exintro = 1;
+		const data = await this.api(params);
+		const page = Object.values(data.query?.pages || {})[0] as any;
+		return this.clean(page?.extract || '');
+	}
+
+	private pageUrl(title: string): string {
+		return `https://en.wikipedia.org/wiki/${encodeURIComponent(title.replace(/ /g, '_'))}`;
+	}
+
+	private stripHtml(text: string): string {
+		return text.replace(/<[^>]+>/g, '');
+	}
+
+	async lookup(query: string, detail: 'intro' | 'full' = 'intro'): Promise<string> {
+		const results = await this.searchTitles(query, 6);
+		if(!results.length) return `❌ No Wikipedia articles found for "${query}"`;
+		const title = results[0].title;
+		const url = this.pageUrl(title);
+		const content = await this.fetchExtract(title, detail === 'intro');
+		const text = this.truncate(content, detail === 'intro' ? 2000 : 8000);
+		return `## ${title}\n🔗 ${url}\n\n${text}`;
+	}
+
+	async search(query: string): Promise<string> {
+		const results = await this.searchTitles(query, 8);
+		if(!results.length) return `❌ No results for "${query}"`;
+		const lines = [`### Search results for "${query}"\n`];
+		for(let i = 0; i < results.length; i++) {
+			const r = results[i];
+			const snippet = this.truncate(this.stripHtml(r.snippet || ''), 150);
+			lines.push(`**${i + 1}. ${r.title}**\n${snippet}\n${this.pageUrl(r.title)}`);
+		}
+		return lines.join('\n\n');
+	}
+}
+
+export const WikipediaLookupTool: AiTool = {
+	name: 'wikipedia_lookup',
+	description: 'Get Wikipedia article content',
+	args: {
+		query: {type: 'string', description: 'Topic or article title', required: true},
+		detail: {type: 'string', description: 'Content level: "intro" (summary, default) or "full" (complete article)', enum: ['intro', 'full'], default: 'intro'}
+	},
+	fn: async (args: {query: string; detail?: 'intro' | 'full'}) => {
+		const wiki = new WikipediaClient();
+		return wiki.lookup(args.query, args.detail || 'intro');
+	}
+};
+
+export const WikipediaSearchTool: AiTool = {
+	name: 'wikipedia_search',
+	description: 'Search Wikipedia for matching articles',
+	args: {
+		query: {type: 'string', description: 'Search terms', required: true}
+	},
+	fn: async (args: {query: string}) => {
+		const wiki = new WikipediaClient();
+		return wiki.search(args.query);
+	}
+};

src/vision.ts

@@ -3,7 +3,7 @@ import {AbortablePromise, Ai} from './ai.ts';
 
 export class Vision {
 
-	constructor(private ai: Ai) { }
+	constructor(private ai: Ai) {}
 
 	/**
 	 * Convert image to text using Optical Character Recognition
@@ -12,12 +12,16 @@ export class Vision {
 	 */
 	ocr(path: string): AbortablePromise<string | null> {
 		let worker: any;
-		const p = new Promise<string | null>(async res => {
-			worker = await createWorker(this.ai.options.tesseract?.model || 'eng', 2, {cachePath: this.ai.options.path});
-			const {data} = await worker.recognize(path);
-			await worker.terminate();
-			res(data.text.trim() || null);
+		const p = (async () => {
+			worker = await createWorker(this.ai.options.ocr || 'eng', 2, {cachePath: this.ai.options.path});
+			worker.setParameters({}).catch(() => {}); // force error handler attachment
+			return await new Promise<string | null>((res, rej) => {
+				worker.on?.('error', rej); // catch worker-level throws
+				worker.recognize(path)
+					.then(({data}: any) => res(data.text.trim() || null))
+					.catch(rej);
 			});
+		})().finally(() => worker?.terminate());
 		return Object.assign(p, {abort: () => worker?.terminate()});
 	}
 }

tsconfig.json

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

vite.config.ts

@@ -1,6 +1,5 @@
 import {defineConfig} from 'vite';
 import dts from 'vite-plugin-dts';
-import {resolve} from 'path';
 
 export default defineConfig({
 	build: {