Updated LLM config and added read_webpage

2026-02-01 13:16:08 -05:00
parent 28904cddbe
commit 7b57a0ded1
11 changed files with 840 additions and 384 deletions
--- a/package-lock.json
+++ b/package-lock.json
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.3.0",
+	"version": "0.4.0",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -30,7 +30,7 @@
 		"@xenova/transformers": "^2.17.2",
 		"@ztimson/node-utils": "^1.0.4",
 		"@ztimson/utils": "^0.27.9",
-		"ollama": "^0.6.0",
+		"cheerio": "^1.2.0",
 		"openai": "^6.6.0",
 		"tesseract.js": "^6.0.1"
 	},
@@ -42,7 +42,6 @@
 		"vite-plugin-dts": "^4.5.3"
 	},
 	"files": [
-		"bin",
 		"dist"
 	]
 }
--- a/src/ai.ts
+++ b/src/ai.ts
@@ -1,18 +1,17 @@
 import * as os from 'node:os';
-import {LLM, LLMOptions} from './llm';
+import {LLM, AnthropicConfig, OllamaConfig, OpenAiConfig, LLMRequest} from './llm';
 import { Audio } from './audio.ts';
 import {Vision} from './vision.ts';

 export type AbortablePromise<T> = Promise<T> & {abort: () => any};

-export type AiOptions = LLMOptions & {
+export type AiOptions = {
 	/** Path to models */
 	path?: string;
-	/** Piper TTS configuratoin */
-	piper?: {
-		/** Model URL: `https://huggingface.co/rhasspy/piper-voices/tree/main/.../model.onnx` */
-		model: 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 */
--- a/src/antrhopic.ts
+++ b/src/antrhopic.ts
@@ -1,5 +1,5 @@
 import {Anthropic as anthropic} from '@anthropic-ai/sdk';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, deepCopy} from '@ztimson/utils';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
 import {LLMProvider} from './provider.ts';
@@ -51,16 +51,16 @@ export class Anthropic extends LLMProvider {
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
 		const controller = new AbortController();
 		const response = new Promise<any>(async (res, rej) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
-			const original = deepCopy(history);
+			let history = [...options.history || [], {role: 'user', content: message, timestamp: Date.now()}];
 			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
+			history = this.fromStandard(<any>history);

-			const tools = options.tools || this.ai.options.tools || [];
+			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
-				max_tokens: options.max_tokens || this.ai.options.max_tokens || 4096,
-				system: options.system || this.ai.options.system || '',
-				temperature: options.temperature || this.ai.options.temperature || 0.7,
+				max_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || 4096,
+				system: options.system || this.ai.options.llm?.system || '',
+				temperature: options.temperature || this.ai.options.llm?.temperature || 0.7,
 				tools: tools.map(t => ({
 					name: t.name,
 					description: t.description,
@@ -117,7 +117,6 @@ export class Anthropic extends LLMProvider {
 				const toolCalls = resp.content.filter((c: any) => c.type === 'tool_use');
 				if(toolCalls.length && !controller.signal.aborted) {
 					history.push({role: 'assistant', content: resp.content});
-					original.push({role: 'assistant', content: resp.content});
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
 						const tool = tools.find(findByProp('name', toolCall.name));
 						if(options.stream) options.stream({tool: toolCall.name});
--- a/src/audio.ts
+++ b/src/audio.ts
@@ -1,6 +1,4 @@
 import {spawn} from 'node:child_process';
-import * as os from 'node:os';
-import {platform, arch} from 'node:os';
 import fs from 'node:fs/promises';
 import Path from 'node:path';
 import {AbortablePromise, Ai} from './ai.ts';
@@ -8,21 +6,12 @@ import {AbortablePromise, Ai} from './ai.ts';
 export class Audio {
 	private downloads: {[key: string]: Promise<string>} = {};
 	private whisperModel!: string;
-	private piperBinary?: string;

 	constructor(private ai: Ai) {
 		if(ai.options.whisper?.binary) {
 			this.whisperModel = ai.options.whisper?.model.endsWith('.bin') ? ai.options.whisper?.model : ai.options.whisper?.model + '.bin';
 			this.downloadAsrModel();
 		}
-		if(ai.options.piper?.model) {
-			if(!ai.options.piper.model.startsWith('http') || !ai.options.piper.model.endsWith('.onnx'))
-				throw new Error('Piper model should be a URL to an onnx file to download');
-			if(platform() != 'linux' || (arch() != 'x64' && arch() != 'arm64'))
-				throw new Error('Piper TTS only supported on Linux x64/arm64');
-			this.piperBinary = Path.join(import.meta.dirname, '../bin/piper');
-			this.downloadTtsModel();
-		}
 	}

 	asr(path: string, model: string = this.whisperModel): AbortablePromise<string | null> {
@@ -43,38 +32,6 @@ export class Audio {
 		return Object.assign(p, {abort});
 	}

-	tts(text: string, outputPath?: string, model: string = <string>this.ai.options.piper?.model): AbortablePromise<Buffer | string> {
-		if(!this.piperBinary) throw new Error('Piper not configured');
-		if(!model) throw new Error('Invalid Piper model');
-		let abort: any = () => {};
-		const p = new Promise<Buffer | string>(async (resolve, reject) => {
-			const modelPath = await this.downloadTtsModel(model);
-			const tmpFile = outputPath || Path.join(os.tmpdir(), `piper_${Date.now()}.wav`);
-			const proc = spawn(<string>this.piperBinary, ['--model', modelPath, '--output_file', tmpFile], {
-				stdio: ['pipe', 'ignore', 'ignore'],
-				env: {...process.env, LD_LIBRARY_PATH: Path.dirname(<string>this.piperBinary)}
-			});
-			abort = () => proc.kill('SIGTERM');
-			proc.stdin.write(text);
-			proc.stdin.end();
-			proc.on('error', (err: Error) => reject(err));
-			proc.on('close', async (code: number) => {
-				if(code === 0) {
-					if(outputPath) {
-						resolve(outputPath);
-					} else {
-						const buffer = await fs.readFile(tmpFile);
-						await fs.unlink(tmpFile).catch(() => {});
-						resolve(buffer);
-					}
-				} else {
-					reject(new Error(`Exit code ${code}`));
-				}
-			});
-		});
-		return Object.assign(p, {abort});
-	}
-
 	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
 		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
 		if(!model.endsWith('.bin')) model += '.bin';
@@ -90,24 +47,4 @@ export class Audio {
 			});
 		return this.downloads[model];
 	}
-
-	async downloadTtsModel(model: string = <string>this.ai.options.piper?.model): Promise<string> {
-		if(!model) throw new Error('Invalid Piper model');
-		const m = <string>model.split('/').pop();
-		const p = Path.join(<string>this.ai.options.path, m);
-		const [onnxExists, jsonExists] = await Promise.all([
-			fs.stat(p).then(() => true).catch(() => false),
-			fs.stat(p + '.json').then(() => true).catch(() => false)
-		]);
-		if(onnxExists && jsonExists) return p;
-		if(!!this.downloads[m]) return this.downloads[m];
-		this.downloads[m] = Promise.all([
-			onnxExists ? Promise.resolve() : fetch(model).then(r => r.arrayBuffer()).then(b => fs.writeFile(p, Buffer.from(b))),
-			jsonExists ? Promise.resolve() : fetch(model + '.json').then(r => r.arrayBuffer()).then(b => fs.writeFile(p + '.json', Buffer.from(b)))
-		]).then(() => {
-			delete this.downloads[m];
-			return p;
-		});
-		return this.downloads[m];
-	}
 }
--- a/src/index.ts
+++ b/src/index.ts
@@ -1,4 +1,9 @@
 export * from './ai';
 export * from './antrhopic';
+export * from './audio';
+export * from './embedder'
 export * from './llm';
+export * from './open-ai';
+export * from './provider';
 export * from './tools';
+export * from './vision';
--- a/src/llm.ts
+++ b/src/llm.ts
@@ -1,7 +1,6 @@
 import {JSONAttemptParse} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
-import {Ollama} from './ollama.ts';
 import {OpenAi} from './open-ai.ts';
 import {LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
@@ -9,6 +8,10 @@ import {Worker} from 'worker_threads';
 import {fileURLToPath} from 'url';
 import {dirname, join} from 'path';

+export type AnthropicConfig = {proto: 'anthropic', token: string};
+export type OllamaConfig = {proto: 'ollama', host: string};
+export type OpenAiConfig = {proto: 'openai', host?: string, token: string};
+
 export type LLMMessage = {
 	/** Message originator */
 	role: 'assistant' | 'system' | 'user';
@@ -33,32 +36,6 @@ export type LLMMessage = {
 	timestamp?: number;
 }

-export type LLMOptions = {
-	/** Anthropic settings */
-	anthropic?: {
-		/** API Token */
-		token: string;
-		/** Default model */
-		model: string;
-	},
-	/** Ollama settings */
-	ollama?: {
-		/** connection URL */
-		host: string;
-		/** Default model */
-		model: string;
-	},
-	/** Open AI settings */
-	openAi?: {
-		/** API Token */
-		token: string;
-		/** Default model */
-		model: string;
-	},
-	/** Default provider & model */
-	model: string | [string, string];
-} & Omit<LLMRequest, 'model'>;
-
 export type LLMRequest = {
 	/** System prompt */
 	system?: string;
@@ -71,7 +48,7 @@ export type LLMRequest = {
 	/** Available tools */
 	tools?: AiTool[];
 	/** LLM  model */
-	model?: string | [string, string];
+	model?: string;
 	/** Stream response */
 	stream?: (chunk: {text?: string, tool?: string, done?: true}) => any;
 	/** Compress old messages in the chat to free up context */
@@ -87,8 +64,8 @@ export class LLM {
 	private embedWorker: Worker | null = null;
 	private embedQueue = new Map<number, { resolve: (value: number[]) => void; reject: (error: any) => void }>();
 	private embedId = 0;
-	private providers: {[key: string]: LLMProvider} = {};
-
+	private models: {[model: string]: LLMProvider} = {};
+	private defaultModel!: string;

 	constructor(public readonly ai: Ai) {
 		this.embedWorker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
@@ -100,9 +77,13 @@ export class LLM {
 			}
 		});

-		if(ai.options.anthropic?.token) this.providers.anthropic = new Anthropic(this.ai, ai.options.anthropic.token, ai.options.anthropic.model);
-		if(ai.options.ollama?.host) this.providers.ollama = new Ollama(this.ai, ai.options.ollama.host, ai.options.ollama.model);
-		if(ai.options.openAi?.token) this.providers.openAi = new OpenAi(this.ai, ai.options.openAi.token, ai.options.openAi.model);
+		if(!ai.options.llm?.models) return;
+		Object.entries(ai.options.llm.models).forEach(([model, config]) => {
+			if(!this.defaultModel) this.defaultModel = model;
+			if(config.proto == 'anthropic') this.models[model] = new Anthropic(this.ai, config.token, model);
+			else if(config.proto == 'ollama') this.models[model] = new OpenAi(this.ai, config.host, 'not-needed', model);
+			else if(config.proto == 'openai') this.models[model] = new OpenAi(this.ai, config.host || null, config.token, model);
+		});
 	}

 	/**
@@ -112,17 +93,9 @@ export class LLM {
 	 * @returns {{abort: () => void, response: Promise<LLMMessage[]>}} Function to abort response and chat history
 	 */
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
-		let model: any = [null, null];
-		if(options.model) {
-			if(typeof options.model == 'object') model = options.model;
-			else model = [options.model, (<any>this.ai.options)[options.model]?.model];
-		}
-		if(!options.model || model[1] == null) {
-			if(typeof this.ai.options.model == 'object') model = this.ai.options.model;
-			else model = [this.ai.options.model, (<any>this.ai.options)[this.ai.options.model]?.model];
-		}
-		if(!model[0] || !model[1]) throw new Error(`Unknown LLM provider or model: ${model[0]} / ${model[1]}`);
-		return this.providers[model[0]].ask(message, {...options, model: model[1]});
+		const m = options.model || this.defaultModel;
+		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
+		return this.models[m].ask(message, options);
 	}

 	/**
--- a/src/ollama.ts
+++ b/src/ollama.ts
@@ -1,123 +0,0 @@
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
-import {AbortablePromise, Ai} from './ai.ts';
-import {LLMMessage, LLMRequest} from './llm.ts';
-import {LLMProvider} from './provider.ts';
-import {Ollama as ollama} from 'ollama';
-
-export class Ollama extends LLMProvider {
-	client!: ollama;
-
-	constructor(public readonly ai: Ai, public host: string, public model: string) {
-		super();
-		this.client = new ollama({host});
-	}
-
-	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
-			if(history[i].role == 'assistant' && history[i].tool_calls) {
-				if(history[i].content) delete history[i].tool_calls;
-				else {
-					history.splice(i, 1);
-					i--;
-				}
-			} else if(history[i].role == 'tool') {
-				const error = history[i].content.startsWith('{"error":');
-				history[i] = {role: 'tool', name: history[i].tool_name, args: history[i].args, [error ? 'error' : 'content']: history[i].content, timestamp: history[i].timestamp};
-			}
-			if(!history[i]?.timestamp) history[i].timestamp = Date.now();
-		}
-		return history;
-	}
-
-	private fromStandard(history: LLMMessage[]): any[] {
-		return history.map((h: any) => {
-			const {timestamp, ...rest} = h;
-			if(h.role != 'tool') return rest;
-			return {role: 'tool', tool_name: h.name, content: h.error || h.content}
-		});
-	}
-
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
-		const controller = new AbortController();
-		const response = new Promise<any>(async (res, rej) => {
-			let system = options.system || this.ai.options.system;
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
-			if(history[0].roll == 'system') {
-				if(!system) system = history.shift();
-				else history.shift();
-			}
-			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min);
-			if(options.system) history.unshift({role: 'system', content: system})
-
-			const tools = options.tools || this.ai.options.tools || [];
-			const requestParams: any = {
-				model: options.model || this.model,
-				messages: history,
-				stream: !!options.stream,
-				signal: controller.signal,
-				options: {
-					temperature: options.temperature || this.ai.options.temperature || 0.7,
-					num_predict: options.max_tokens || this.ai.options.max_tokens || 4096,
-				},
-				tools: tools.map(t => ({
-					type: 'function',
-					function: {
-						name: t.name,
-						description: t.description,
-						parameters: {
-							type: 'object',
-							properties: t.args ? objectMap(t.args, (key, value) => ({...value, required: undefined})) : {},
-							required: t.args ? Object.entries(t.args).filter(t => t[1].required).map(t => t[0]) : []
-						}
-					}
-				}))
-			}
-
-			let resp: any, isFirstMessage = true;
-			do {
-				resp = await this.client.chat(requestParams).catch(err => {
-					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
-					throw err;
-				});
-
-				if(options.stream) {
-					if(!isFirstMessage) options.stream({text: '\n\n'});
-					else isFirstMessage = false;
-					resp.message = {role: 'assistant', content: '', tool_calls: []};
-					for await (const chunk of resp) {
-						if(controller.signal.aborted) break;
-						if(chunk.message?.content) {
-							resp.message.content += chunk.message.content;
-							options.stream({text: chunk.message.content});
-						}
-						if(chunk.message?.tool_calls) resp.message.tool_calls = chunk.message.tool_calls;
-						if(chunk.done) break;
-					}
-				}
-
-				if(resp.message?.tool_calls?.length && !controller.signal.aborted) {
-					history.push(resp.message);
-					const results = await Promise.all(resp.message.tool_calls.map(async (toolCall: any) => {
-						const tool = tools.find(findByProp('name', toolCall.function.name));
-						if(options.stream) options.stream({tool: toolCall.function.name});
-						if(!tool) return {role: 'tool', tool_name: toolCall.function.name, content: '{"error": "Tool not found"}'};
-						const args = typeof toolCall.function.arguments === 'string' ? JSONAttemptParse(toolCall.function.arguments, {}) : toolCall.function.arguments;
-						try {
-							const result = await tool.fn(args, this.ai);
-							return {role: 'tool', tool_name: toolCall.function.name, args, content: JSONSanitize(result)};
-						} catch (err: any) {
-							return {role: 'tool', tool_name: toolCall.function.name, args, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'})};
-						}
-					}));
-					history.push(...results);
-					requestParams.messages = history;
-				}
-			} while (!controller.signal.aborted && resp.message?.tool_calls?.length);
-
-			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {role: 'assistant', content: resp.message?.content}]));
-		});
-
-		return Object.assign(response, {abort: () => controller.abort()});
-	}
-}
--- a/src/open-ai.ts
+++ b/src/open-ai.ts
@@ -1,5 +1,5 @@
 import {OpenAI as openAI} from 'openai';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean} from '@ztimson/utils';
 import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
 import {LLMProvider} from './provider.ts';
@@ -7,9 +7,12 @@ import {LLMProvider} from './provider.ts';
 export class OpenAi extends LLMProvider {
 	client!: openAI;

-	constructor(public readonly ai: Ai, public readonly apiToken: string, public model: string) {
+	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string, public model: string) {
 		super();
-		this.client = new openAI({apiKey: apiToken});
+		this.client = new openAI(clean({
+			baseURL: host,
+			apiKey: token
+		}));
 	}

 	private toStandard(history: any[]): LLMMessage[] {
@@ -64,16 +67,17 @@ export class OpenAi extends LLMProvider {
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
 		const controller = new AbortController();
 		const response = new Promise<any>(async (res, rej) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
+			let history = [...options.history || [], {role: 'user', content: message, timestamp: Date.now()}];
 			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
+			history = this.fromStandard(<any>history);

-			const tools = options.tools || this.ai.options.tools || [];
+			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
 				messages: history,
 				stream: !!options.stream,
-				max_tokens: options.max_tokens || this.ai.options.max_tokens || 4096,
-				temperature: options.temperature || this.ai.options.temperature || 0.7,
+				max_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || 4096,
+				temperature: options.temperature || this.ai.options.llm?.temperature || 0.7,
 				tools: tools.map(t => ({
 					type: 'function',
 					function: {
--- a/src/provider.ts
+++ b/src/provider.ts
@@ -1,5 +1,5 @@
 import {AbortablePromise} from './ai.ts';
-import {LLMMessage, LLMOptions, LLMRequest} from './llm.ts';
+import {LLMMessage, LLMRequest} from './llm.ts';

 export abstract class LLMProvider {
 	abstract ask(message: string, options: LLMRequest): AbortablePromise<LLMMessage[]>;
--- a/src/tools.ts
+++ b/src/tools.ts
@@ -1,3 +1,4 @@
+import * as cheerio from 'cheerio';
 import {$, $Sync} from '@ztimson/node-utils';
 import {ASet, consoleInterceptor, Http, fn as Fn} from '@ztimson/utils';
 import {Ai} from './ai.ts';
@@ -111,9 +112,43 @@ export const PythonTool: AiTool = {
 	fn: async (args: {code: string}) => ({result: $Sync`python -c "${args.code}"`})
 }

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