Bump 0.9.0

+ added wikitools
+ Improved webpage reading tool

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.8.1",
+	"version": "0.9.0",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -29,7 +29,7 @@
 		"@tensorflow/tfjs": "^4.22.0",
 		"@xenova/transformers": "^2.17.2",
 		"@ztimson/node-utils": "^1.0.7",
-		"@ztimson/utils": "^0.28.13",
+		"@ztimson/utils": "^0.28.16",
 		"cheerio": "^1.2.0",
 		"openai": "^6.22.0",
 		"tesseract.js": "^7.0.0"

src/antrhopic.ts

@@ -119,7 +119,7 @@ export class Anthropic extends LLMProvider {
 						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
 						try {
 							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'};
 						}

src/llm.ts

@@ -1,3 +1,4 @@
+import {sum} from '@tensorflow/tfjs';
 import {JSONAttemptParse} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
@@ -44,8 +45,6 @@ export type LLMMemory = {
 	fact: string;
 	/** Owner and fact embedding vector */
 	embeddings: [number[], number[]];
-	/** Creation time */
-	timestamp: Date;
 }
 
 export type LLMRequest = {
@@ -95,71 +94,86 @@ class LLM {
 	 * @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
+			// If memories were passed, find any relevant ones and add a tool for ADHOC lookups
 			if(options.memory) {
-				options.system = (options.system || '') + '\nYou have passive persistent memory never make any mention of your memory capabilities and what you can/cannot remember\n';
+				const search = async (query?: string | null, subject?: string | null, limit = 10) => {
-				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 || [])
+					return (options.memory || []).map(m => {
-						.map(m => ({...m, score: o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 1}))
+						const score = (o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 0)
-						.filter((m: any) => m.score >= 0.8)
+							+ (q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : 0);
-						.map((m: any) => ({...m, score: q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : m.score}))
+						return {...m, score};
-						.filter((m: any) => m.score >= 0.2)
+					}).toSorted((a: any, b: any) => a.score - b.score).slice(0, limit)
-						.toSorted((a: any, b: any) => a.score - b.score)
+						.map(m => `- ${m.owner}: ${m.fact}`).join('\n');
-						.slice(0, limit);
 				}
 
+				options.system += '\nYou have RAG memory and will be given the top_k closest memories regarding the users query. Save anything new you have learned worth remembering from the user message using the remember tool and feel free to recall memories manually.\n';
 				const relevant = await search(message);
-				if(relevant.length) options.history.push({role: 'assistant', content: 'Things I remembered:\n' + relevant.map(m => `${m.owner}: ${m.fact}`).join('\n')});
+				if(relevant.length) options.history.push({role: 'tool', name: 'recall', id: 'auto_recall_' + Math.random().toString(), args: {}, content: `Things I remembered:\n${relevant}`});
-				options.tools = [...options.tools || [], {
+				options.tools = [{
-					name: 'read_memory',
+					name: 'recall',
-					description: 'Check your long-term memory for more information',
+					description: 'Recall the closest memories you have regarding a query using RAG',
 					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'},
+						topK: {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);
+						return search(args.query, args.subject, args.topK);
 					}
-				}];
+				}, {
+					name: 'remember',
+					description: 'Store important facts user shares for future recall',
+					args: {
+						owner: {type: 'string', description: 'Subject/person this fact is about'},
+						fact: {type: 'string', description: 'The information to remember'}
+					},
+					fn: async (args) => {
+						if(!options.memory) return;
+						const e = await Promise.all([
+							this.embedding(args.owner),
+							this.embedding(`${args.owner}: ${args.fact}`)
+						]);
+						const newMem = {owner: args.owner, fact: args.fact, embeddings: <any>[e[0][0].embedding, e[1][0].embedding]};
+						options.memory.splice(0, options.memory.length, ...[
+							...options.memory.filter(m => {
+								return !(this.cosineSimilarity(newMem.embeddings[0], m.embeddings[0]) >= 0.9 && this.cosineSimilarity(newMem.embeddings[1], m.embeddings[1]) >= 0.8);
+							}),
+							newMem
+						]);
+						return 'Remembered!';
+					}
+				}, ...options.tools || []];
 			}
 
 			// Ask
 			const resp = await this.models[m].ask(message, options);
 
-			// Remove any memory calls
+			// Remove any memory calls from history
-			if(options.memory) {
+			if(options.memory) options.history.splice(0, options.history.length, ...options.history.filter(h => h.role != 'tool' || (h.name != 'recall' && h.name != 'remember')));
-				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);
+			// Compress message history
+			if(options.compress) {
+				const compressed = await this.ai.language.compressHistory(options.history, options.compress.max, options.compress.min, options);
+				options.history.splice(0, options.history.length, ...compressed);
 			}
 
-			// Handle compression and memory extraction
-			if(options.compress || options.memory) {
-				let compressed: any = null;
-				if(options.compress) {
-					compressed = await this.ai.language.compressHistory(options.history, options.compress.max, options.compress.min, options);
-					options.history.splice(0, options.history.length, ...compressed.history);
-				} else {
-					const i = options.history?.findLastIndex(m => m.role == 'user') ?? -1;
-					compressed = await this.ai.language.compressHistory(i != -1 ? options.history.slice(i) : options.history, 0, 0, options);
-				}
-				if(options.memory) {
-					const updated = options.memory
-						.filter(m => !compressed.memory.some(m2 => this.cosineSimilarity(m.embeddings[1], m2.embeddings[1]) > 0.8))
-						.concat(compressed.memory);
-					options.memory.splice(0, options.memory.length, ...updated);
-				}
-			}
 			return res(resp);
 		}), {abort});
 	}
@@ -181,32 +195,24 @@ 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[]}> {
+	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<LLMMessage[]> {
-		if(this.estimateTokens(history) < max) return {history, memory: []};
+		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(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), '{summary: string, facts: [[subject, fact]]}', {
+		const summary: any = await this.summarize(process.map(m => `[${m.role}]: ${m.content}`).join('\n\n'), 500, options);
-			system: 'Create the smallest summary possible, no more than 500 tokens. Create a list of NEW facts (split by subject [pro]noun and fact) about what you learned from this conversation that you didn\'t already know or get from a tool call or system prompt. Focus only on new information about people, topics, or facts. Avoid generating facts about the AI.',
+		const d = Date.now();
-			model: options?.model,
+		const h = [{role: <any>'tool', name: 'summary', id: `summary_` + d, args: {}, content: `Conversation Summary: ${summary?.summary}`, timestamp: d}, ...recent];
-			temperature: options?.temperature || 0.3
-		});
-		const timestamp = new Date();
-		const memory = await Promise.all((summary?.facts || [])?.map(async ([owner, fact]: [string, string]) => {
-			const e = await Promise.all([this.embedding(owner), this.embedding(`${owner}: ${fact}`)]);
-			return {owner, fact, embeddings: [e[0][0].embedding, e[1][0].embedding], timestamp};
-		}));
-		const h = [{role: 'assistant', content: `Conversation Summary: ${summary?.summary}`, timestamp: Date.now()}, ...recent];
 		if(system) h.splice(0, 0, system);
-		return {history: <any>h, memory};
+		return h;
 	}
 
 	/**
@@ -243,7 +249,7 @@ class LLM {
 				return `${p}: ${Array.isArray(value) ? value.join(', ') : value}`;
 			});
 		};
-		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
+		const lines = typeof target === 'object' ? objString(target) : target.toString().split('\n');
 		const tokens = lines.flatMap(l => [...l.split(/\s+/).filter(Boolean), '\n']);
 		const chunks: string[] = [];
 		for(let i = 0; i < tokens.length;) {
@@ -352,22 +358,64 @@ class LLM {
 	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
 	 */
 	async json(text: string, schema: string, options?: LLMRequest): Promise<any> {
-		const code = await this.code(text, {...options, system: [
+		let system =  `Your job is to convert input to JSON using tool calls. Call the \`submit\` tool at least once with JSON matching this schema:\n\`\`\`json\n${schema}\n\`\`\`\n\nResponses are ignored`;
-			options?.system,
+		if(options?.system) system += '\n\n' + options.system;
-			`Only respond using JSON matching this schema:\n\`\`\`json\n${schema}\n\`\`\``
+		return new Promise(async (resolve, reject) => {
-		].filter(t => !!t).join('\n')});
+			let done = false;
-		return code ? JSONAttemptParse(code, {}) : null;
+			const resp = await this.ask(text, {
+				temperature: 0.3,
+				...options,
+				system,
+				tools: [{
+					name: 'submit',
+					description: 'Submit JSON',
+					args: {json: {type: 'string', description: 'Javascript parsable JSON string', required: true}},
+					fn: (args) => {
+						try {
+							const json = JSON.parse(args.json);
+							resolve(json);
+							done = true;
+						} catch { return 'Invalid JSON'; }
+						return 'Saved';
+					}
+				}, ...(options?.tools || [])],
+			});
+			if(!done) reject(`AI failed to create JSON:\n${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> {
+	async summarize(text: string, length: number = 500, options?: LLMRequest): Promise<string | null> {
-		return this.ask(text, {system: `Generate a brief summary <= ${tokens} tokens. Output nothing else`, temperature: 0.3, ...options});
+		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}`);
+		});
 	}
 }
 

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
 		}));
 	}
 
@@ -67,7 +67,10 @@ 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()});
+			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 = {
@@ -100,15 +103,37 @@ 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 || ''
+										}
+									});
+								}
+							}
 						}
 					}
 				}
@@ -123,7 +148,7 @@ 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'})};
 						}

src/tools.ts

@@ -1,9 +1,17 @@
 import  * as cheerio from 'cheerio';
-import {$, $Sync} from '@ztimson/node-utils';
+import {$Sync} from '@ztimson/node-utils';
-import {ASet, consoleInterceptor, Http, fn as Fn} from '@ztimson/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) {
+			switch(args.language) {
-				case 'bash':
+				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 c = consoleInterceptor(null);
-		const resp = await Fn<any>({console}, args.code, true).catch((err: any) => console.output.error.push(err));
+		const resp = await Fn<any>({console: c}, args.code, true).catch((err: any) => c.output.error.push(err));
-		return {...console.output, return: resp, stdout: undefined, stderr: undefined};
+		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},
+		url: {type: 'string', description: 'URL to read', required: true},
-		focus: {type: 'string', description: 'Optional: What aspect to focus on (e.g., "pricing", "features", "contact info")'}
+		mimeRegex: {type: 'string', description: 'Optional regex to filter MIME types (e.g., "^image/", "text/")'}
 	},
-	fn: async (args: {url: string; focus?: string}) => {
+	fn: async (args: {url: string; mimeRegex?: string}) => {
-		const html = await fetch(args.url, {headers: {"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64)"}})
+		const ua = 'AiTools-Webpage/1.0';
-			.then(r => r.text()).catch(err => {throw new Error(`Failed to fetch: ${err.message}`)});
+		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();
+		$('script, style, nav, footer, header, aside, iframe, noscript, svg').remove();
-		const metadata = {
+		$('[role="navigation"], [role="banner"], [role="complementary"]').remove();
-			title: $('meta[property="og:title"]').attr('content') || $('title').text() || '',
+		$('[aria-hidden="true"], [hidden], .visually-hidden, .sr-only, .screen-reader-text').remove();
-			description: $('meta[name="description"]').attr('content') || $('meta[property="og:description"]').attr('content') || '',
+		$('.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'];
+		const selectors = ['article', 'main', '[role="main"]', '.content', '.post-content', '.entry-content', '.article-content'];
-		for (const selector of contentSelectors) {
+		for(const sel of selectors) {
-			const el = $(selector).first();
+			const el = $(sel).first();
 			if(el.length && el.text().trim().length > 200) {
-				content = el.text();
+				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);
+	}
+};