Added memory system

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.2.7",
+	"version": "0.5.1",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -30,7 +30,7 @@
 		"@xenova/transformers": "^2.17.2",
 		"@ztimson/node-utils": "^1.0.4",
 		"@ztimson/utils": "^0.27.9",
-		"ollama": "^0.6.0",
+		"cheerio": "^1.2.0",
 		"openai": "^6.6.0",
 		"tesseract.js": "^6.0.1"
 	},

src/ai.ts

@@ -1,11 +1,24 @@
 import * as os from 'node:os';
-import {LLM, LLMOptions} from './llm';
+import LLM, {AnthropicConfig, OllamaConfig, OpenAiConfig, LLMRequest} from './llm';
 import { Audio } from './audio.ts';
 import {Vision} from './vision.ts';
 
-export type AiOptions = LLMOptions & {
+export type AbortablePromise<T> = Promise<T> & {
+	abort: () => any
+};
+
+export type AiOptions = {
 	/** Path to models */
 	path?: string;
+	/** Large language models, first is default */
+	llm?: Omit<LLMRequest, 'model'> & {
+		models: {[model: string]: AnthropicConfig | OllamaConfig | OpenAiConfig};
+	}
+	/** Tesseract OCR configuration */
+	tesseract?: {
+		/** Model: eng, eng_best, eng_fast */
+		model?: string;
+	}
 	/** Whisper ASR configuration */
 	whisper?: {
 		/** Whisper binary location */
@@ -13,11 +26,6 @@ export type AiOptions = LLMOptions & {
 		/** Model: `ggml-base.en.bin` */
 		model: string;
 	}
-	/** Tesseract OCR configuration */
-	tesseract?: {
-		/** Model: eng, eng_best, eng_fast */
-		model?: string;
-	}
 }
 
 export class Ai {

src/antrhopic.ts

@@ -1,8 +1,8 @@
 import {Anthropic as anthropic} from '@anthropic-ai/sdk';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, deepCopy} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
+import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 
 export class Anthropic extends LLMProvider {
 	client!: anthropic;
@@ -18,8 +18,7 @@ export class Anthropic extends LLMProvider {
 			if(typeof history[orgI].content != 'string') {
 				if(history[orgI].role == 'assistant') {
 					history[orgI].content.filter((c: any) => c.type =='tool_use').forEach((c: any) => {
-						i++;
-						history.splice(i, 0, {role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: Date.now()});
+						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) => {
@@ -28,6 +27,7 @@ export class Anthropic extends LLMProvider {
 					});
 				}
 				history[orgI].content = history[orgI].content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
+				if(!history[orgI].content) history.splice(orgI, 1);
 			}
 			if(!history[orgI].timestamp) history[orgI].timestamp = Date.now();
 		}
@@ -48,19 +48,16 @@ export class Anthropic extends LLMProvider {
 		return history.map(({timestamp, ...h}) => h);
 	}
 
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
+	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		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);
-			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
-
-			const tools = options.tools || this.ai.options.tools || [];
+		return Object.assign(new Promise<any>(async (res, rej) => {
+			const 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,
-				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,
@@ -117,12 +114,13 @@ export class Anthropic extends LLMProvider {
 				const toolCalls = resp.content.filter((c: any) => c.type === 'tool_use');
 				if(toolCalls.length && !controller.signal.aborted) {
 					history.push({role: 'assistant', content: resp.content});
-					original.push({role: 'assistant', content: resp.content});
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
 						const tool = tools.find(findByProp('name', toolCall.name));
+						if(options.stream) options.stream({tool: toolCall.name});
 						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
 						try {
-							const result = await tool.fn(toolCall.input, this.ai);
+							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)};
 						} catch (err: any) {
 							return {type: 'tool_result', tool_use_id: toolCall.id, is_error: true, content: err?.message || err?.toString() || 'Unknown'};
@@ -132,11 +130,12 @@ export class Anthropic extends LLMProvider {
 					requestParams.messages = history;
 				}
 			} 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);
 
 			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {role: 'assistant', content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')}]));
-		});
-
-		return Object.assign(response, {abort: () => controller.abort()});
+			if(options.history) options.history.splice(0, options.history.length, ...history);
+			res(history.at(-1)?.content);
+		}), {abort: () => controller.abort()});
 	}
 }

src/audio.ts

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

src/index.ts

@@ -1,4 +1,9 @@
 export * from './ai';
 export * from './antrhopic';
+export * from './audio';
+export * from './embedder'
 export * from './llm';
+export * from './open-ai';
+export * from './provider';
 export * from './tools';
+export * from './vision';

src/llm.ts

@@ -1,14 +1,17 @@
 import {JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
-import {Ollama} from './ollama.ts';
 import {OpenAi} from './open-ai.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
 import {Worker} from 'worker_threads';
 import {fileURLToPath} from 'url';
 import {dirname, join} from 'path';
 
+export type AnthropicConfig = {proto: 'anthropic', token: string};
+export type OllamaConfig = {proto: 'ollama', host: string};
+export type OpenAiConfig = {proto: 'openai', host?: string, token: string};
+
 export type LLMMessage = {
 	/** Message originator */
 	role: 'assistant' | 'system' | 'user';
@@ -28,36 +31,22 @@ export type LLMMessage = {
 	/** Tool result */
 	content: undefined | string;
 	/** Tool error */
-	error: undefined | string;
+	error?: undefined | string;
 	/** Timestamp */
 	timestamp?: number;
 }
 
-export type LLMOptions = {
-	/** Anthropic settings */
-	anthropic?: {
-		/** API Token */
-		token: string;
-		/** Default model */
-		model: string;
-	},
-	/** Ollama settings */
-	ollama?: {
-		/** connection URL */
-		host: string;
-		/** Default model */
-		model: string;
-	},
-	/** Open AI settings */
-	openAi?: {
-		/** API Token */
-		token: string;
-		/** Default model */
-		model: string;
-	},
-	/** Default provider & model */
-	model: string | [string, string];
-} & Omit<LLMRequest, 'model'>;
+/** 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 */
@@ -71,24 +60,26 @@ export type LLMRequest = {
 	/** Available tools */
 	tools?: AiTool[];
 	/** LLM  model */
-	model?: string | [string, string];
+	model?: string;
 	/** Stream response */
-	stream?: (chunk: {text?: string, done?: true}) => any;
+	stream?: (chunk: {text?: string, tool?: string, done?: true}) => any;
 	/** Compress old messages in the chat to free up context */
 	compress?: {
 		/** Trigger chat compression once context exceeds the token count */
 		max: number;
 		/** Compress chat until context size smaller than */
 		min: number
-	}
+	},
+	/** Background information the AI will be fed */
+	memory?: LLMMemory[],
 }
 
-export class LLM {
+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'));
@@ -100,54 +91,128 @@ 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);
+		});
 	}
 
 	/**
 	 * Chat with LLM
 	 * @param {string} message Question
 	 * @param {LLMRequest} options Configuration options and chat history
-	 * @returns {{abort: () => void, response: Promise<LLMMessage[]>}} Function to abort response and chat history
+	 * @returns {{abort: () => void, response: Promise<string>}} 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];
+	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
+		const m = options.model || this.defaultModel;
+		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
+		let abort = () => {};
+		return Object.assign(new Promise<string>(async res => {
+			if(!options.history) options.history = [];
+			// If memories were passed, find any relivant ones and add a tool for ADHOC lookups
+			if(options.memory) {
+				options.system = (options.system || '') + '\nYou have passive persistent memory never make any mention of your memory capabilities and what you can/cannot remember\n';
+				const search = async (query?: string | null, subject?: string | null, limit = 50) => {
+					const [o, q] = await Promise.all([
+						subject ? this.embedding(subject) : Promise.resolve(null),
+						query ? this.embedding(query) : Promise.resolve(null),
+					]);
+					return (options.memory || [])
+						.map(m => ({...m, score: o ? this.cosineSimilarity(m.embeddings[0], o[0].embedding) : 1}))
+						.filter((m: any) => m.score >= 0.8)
+						.map((m: any) => ({...m, score: q ? this.cosineSimilarity(m.embeddings[1], q[0].embedding) : m.score}))
+						.filter((m: any) => m.score >= 0.2)
+						.toSorted((a: any, b: any) => a.score - b.score)
+						.slice(0, limit);
 				}
-		if(!options.model || model[1] == null) {
-			if(typeof this.ai.options.model == 'object') model = this.ai.options.model;
-			else model = [this.ai.options.model, (<any>this.ai.options)[this.ai.options.model]?.model];
+
+				const relevant = await search(message);
+				if(relevant.length) options.history.push({role: 'assistant', content: 'Things I remembered:\n' + relevant.map(m => `${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);
 					}
-		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]});
+				}];
+			}
+
+			// Ask
+			const resp = await this.models[m].ask(message, options);
+
+			// Remove any memory calls
+			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);
+			}
+
+			// 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);
+				}
+			}
+			return res(resp);
+		}), {abort});
 	}
 
 	/**
 	 * Compress chat history to reduce context size
 	 * @param {LLMMessage[]} history Chatlog that will be compressed
 	 * @param max Trigger compression once context is larger than max
-	 * @param min Summarize until context size is less than min
+	 * @param min Leave messages less than the token minimum, summarize the rest
 	 * @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<LLMMessage[]> {
-		if(this.estimateTokens(history) < max) return history;
+	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<{history: LLMMessage[], memory: LLMMemory[]}> {
+		if(this.estimateTokens(history) < max) return {history, memory: []};
 		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;
-		const recent = keep == 0 ? [] : history.slice(-keep),
+		if(history.length <= keep) return {history, memory: []};
+		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 = await this.summarize(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), 250, options);
-		return [{role: 'assistant', content: `Conversation Summary: ${summary}`, timestamp: Date.now()}, ...recent];
+		const summary: any = await this.json(`Create the smallest summary possible, no more than 500 tokens. Create a list of NEW facts (split by subject [pro]noun and fact) about what you learned from this conversation that you didn't already know or get from a tool call or system prompt. Focus only on new information about people, topics, or facts. Avoid generating facts about the AI. Match this format: {summary: string, facts: [[subject, fact]]}\n\n${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];
+		if(system) h.splice(0, 0, system);
+		return {history: <any>h, memory};
 	}
 
+	/**
+	 * Compare the difference between embeddings (calculates the angle between two vectors)
+	 * @param {number[]} v1 First embedding / vector comparison
+	 * @param {number[]} v2 Second embedding / vector for comparison
+	 * @returns {number} Similarity values 0-1: 0 = unique, 1 = identical
+	 */
 	cosineSimilarity(v1: number[], v2: number[]): number {
 		if (v1.length !== v2.length) throw new Error('Vectors must be same length');
 		let dotProduct = 0, normA = 0, normB = 0;
@@ -160,6 +225,13 @@ export class LLM {
 		return denominator === 0 ? 0 : dotProduct / denominator;
 	}
 
+	/**
+	 * Chunk text into parts for AI digestion
+	 * @param {object | string} target Item that will be chunked (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)
+	 * @returns {string[]} Chunked strings
+	 */
 	chunk(target: object | string, maxTokens = 500, overlapTokens = 50): string[] {
 		const objString = (obj: any, path = ''): string[] => {
 			if(!obj) return [];
@@ -169,7 +241,6 @@ export class LLM {
 				return `${p}: ${Array.isArray(value) ? value.join(', ') : value}`;
 			});
 		};
-
 		const lines = typeof target === 'object' ? objString(target) : target.split('\n');
 		const tokens = lines.flatMap(l => [...l.split(/\s+/).filter(Boolean), '\n']);
 		const chunks: string[] = [];
@@ -188,6 +259,13 @@ export class LLM {
 		return chunks;
 	}
 
+	/**
+	 * 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)
+	 * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
+	 */
 	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
 		const embed = (text: string): Promise<number[]> => {
 			return new Promise((resolve, reject) => {
@@ -196,7 +274,6 @@ export class LLM {
 				this.embedWorker?.postMessage({ id, text });
 			});
 		};
-
 		const chunks = this.chunk(target, maxTokens, overlapTokens);
 		return Promise.all(chunks.map(async (text, index) => ({
 			index,
@@ -218,7 +295,7 @@ export class LLM {
 
 	/**
 	 * Compare the difference between two strings using tensor math
-	 * @param target Text that will checked
+	 * @param target Text that will be checked
 	 * @param {string} searchTerms Multiple search terms to check against target
 	 * @returns {{avg: number, max: number, similarities: number[]}} Similarity values 0-1: 0 = unique, 1 = identical
 	 */
@@ -239,13 +316,12 @@ export class LLM {
 	 * @param {LLMRequest} options Configuration options and chat history
 	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
 	 */
-	async json(message: string, options?: LLMRequest) {
-		let resp = await this.ask(message, {
-			system: 'Respond using a JSON blob',
-			...options
-		});
-		if(!resp?.[0]?.content) return {};
-		return JSONAttemptParse(new RegExp('\{[\s\S]*\}').exec(resp[0].content), {});
+	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, {});
 	}
 
 	/**
@@ -256,7 +332,8 @@ export class LLM {
 	 * @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})
-			.then(history => <string>history.pop()?.content || null);
+		return this.ask(text, {system: `Generate a brief summary <= ${tokens} tokens. Output nothing else`, temperature: 0.3, ...options});
 	}
 }
+
+export default LLM;

src/ollama.ts

@@ -1,122 +0,0 @@
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
-import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
-import {Ollama as ollama} from 'ollama';
-
-export class Ollama extends LLMProvider {
-	client!: ollama;
-
-	constructor(public readonly ai: Ai, public host: string, public model: string) {
-		super();
-		this.client = new ollama({host});
-	}
-
-	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
-			if(history[i].role == 'assistant' && history[i].tool_calls) {
-				if(history[i].content) delete history[i].tool_calls;
-				else {
-					history.splice(i, 1);
-					i--;
-				}
-			} else if(history[i].role == 'tool') {
-				const error = history[i].content.startsWith('{"error":');
-				history[i] = {role: 'tool', name: history[i].tool_name, args: history[i].args, [error ? 'error' : 'content']: history[i].content, timestamp: history[i].timestamp};
-			}
-			if(!history[i]?.timestamp) history[i].timestamp = Date.now();
-		}
-		return history;
-	}
-
-	private fromStandard(history: LLMMessage[]): any[] {
-		return history.map((h: any) => {
-			const {timestamp, ...rest} = h;
-			if(h.role != 'tool') return rest;
-			return {role: 'tool', tool_name: h.name, content: h.error || h.content}
-		});
-	}
-
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
-		const controller = new AbortController();
-		const response = new Promise<any>(async (res, rej) => {
-			let system = options.system || this.ai.options.system;
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
-			if(history[0].roll == 'system') {
-				if(!system) system = history.shift();
-				else history.shift();
-			}
-			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min);
-			if(options.system) history.unshift({role: 'system', content: system})
-
-			const tools = options.tools || this.ai.options.tools || [];
-			const requestParams: any = {
-				model: options.model || this.model,
-				messages: history,
-				stream: !!options.stream,
-				signal: controller.signal,
-				options: {
-					temperature: options.temperature || this.ai.options.temperature || 0.7,
-					num_predict: options.max_tokens || this.ai.options.max_tokens || 4096,
-				},
-				tools: tools.map(t => ({
-					type: 'function',
-					function: {
-						name: t.name,
-						description: t.description,
-						parameters: {
-							type: 'object',
-							properties: t.args ? objectMap(t.args, (key, value) => ({...value, required: undefined})) : {},
-							required: t.args ? Object.entries(t.args).filter(t => t[1].required).map(t => t[0]) : []
-						}
-					}
-				}))
-			}
-
-			let resp: any, isFirstMessage = true;
-			do {
-				resp = await this.client.chat(requestParams).catch(err => {
-					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
-					throw err;
-				});
-
-				if(options.stream) {
-					if(!isFirstMessage) options.stream({text: '\n\n'});
-					else isFirstMessage = false;
-					resp.message = {role: 'assistant', content: '', tool_calls: []};
-					for await (const chunk of resp) {
-						if(controller.signal.aborted) break;
-						if(chunk.message?.content) {
-							resp.message.content += chunk.message.content;
-							options.stream({text: chunk.message.content});
-						}
-						if(chunk.message?.tool_calls) resp.message.tool_calls = chunk.message.tool_calls;
-						if(chunk.done) break;
-					}
-				}
-
-				if(resp.message?.tool_calls?.length && !controller.signal.aborted) {
-					history.push(resp.message);
-					const results = await Promise.all(resp.message.tool_calls.map(async (toolCall: any) => {
-						const tool = tools.find(findByProp('name', toolCall.function.name));
-						if(!tool) return {role: 'tool', tool_name: toolCall.function.name, content: '{"error": "Tool not found"}'};
-						const args = typeof toolCall.function.arguments === 'string' ? JSONAttemptParse(toolCall.function.arguments, {}) : toolCall.function.arguments;
-						try {
-							const result = await tool.fn(args, this.ai);
-							return {role: 'tool', tool_name: toolCall.function.name, args, content: JSONSanitize(result)};
-						} catch (err: any) {
-							return {role: 'tool', tool_name: toolCall.function.name, args, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'})};
-						}
-					}));
-					history.push(...results);
-					requestParams.messages = history;
-				}
-			} while (!controller.signal.aborted && resp.message?.tool_calls?.length);
-
-			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {role: 'assistant', content: resp.message?.content}]));
-		});
-
-		return Object.assign(response, {abort: () => controller.abort()});
-	}
-}

src/open-ai.ts

@@ -1,15 +1,18 @@
 import {OpenAI as openAI} from 'openai';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean} from '@ztimson/utils';
+import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 
 export class OpenAi extends LLMProvider {
 	client!: openAI;
 
-	constructor(public readonly ai: Ai, public readonly apiToken: string, public model: string) {
+	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string, public model: string) {
 		super();
-		this.client = new openAI({apiKey: apiToken});
+		this.client = new openAI(clean({
+			baseURL: host,
+			apiKey: token
+		}));
 	}
 
 	private toStandard(history: any[]): LLMMessage[] {
@@ -61,19 +64,18 @@ export class OpenAi extends LLMProvider {
 		}, [] as any[]);
 	}
 
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
+	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		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()}]);
-			if(options.compress) history = await this.ai.language.compressHistory(<any>history, options.compress.max, options.compress.min, options);
-
-			const tools = options.tools || this.ai.options.tools || [];
+		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()}]);
+			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: {
@@ -116,10 +118,11 @@ export class OpenAi extends LLMProvider {
 					history.push(resp.choices[0].message);
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
 						const tool = tools?.find(findByProp('name', toolCall.function.name));
+						if(options.stream) options.stream({tool: toolCall.function.name});
 						if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}'};
 						try {
 							const args = JSONAttemptParse(toolCall.function.arguments, {});
-							const result = await tool.fn(args, this.ai);
+							const result = await tool.fn(args, options.stream, this.ai);
 							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize(result)};
 						} catch (err: any) {
 							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'})};
@@ -129,10 +132,12 @@ 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);
 
 			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {role: 'assistant', content: resp.choices[0].message.content || ''}]));
-		});
-		return Object.assign(response, {abort: () => controller.abort()});
+			if(options.history) options.history.splice(0, options.history.length, ...history);
+			res(history.at(-1)?.content);
+		}), {abort: () => controller.abort()});
 	}
 }

src/provider.ts

@@ -1,7 +1,6 @@
-import {LLMMessage, LLMOptions, LLMRequest} from './llm.ts';
-
-export type AbortablePromise<T> = Promise<T> & {abort: () => void};
+import {AbortablePromise} from './ai.ts';
+import {LLMMessage, LLMRequest} from './llm.ts';
 
 export abstract class LLMProvider {
-	abstract ask(message: string, options: LLMRequest): AbortablePromise<LLMMessage[]>;
+	abstract ask(message: string, options: LLMRequest): AbortablePromise<string>;
 }

src/tools.ts

@@ -1,6 +1,8 @@
+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';
+import {LLMRequest} from './llm.ts';
 
 export type AiToolArg = {[key: string]: {
 	/** Argument type */
@@ -31,7 +33,7 @@ export type AiTool = {
 	/** Tool arguments */
 	args?: AiToolArg,
 	/** Callback function */
-	fn: (args: any, ai: Ai) => any | Promise<any>,
+	fn: (args: any, stream: LLMRequest['stream'], ai: Ai) => any | Promise<any>,
 };
 
 export const CliTool: AiTool = {
@@ -43,9 +45,9 @@ export const CliTool: AiTool = {
 
 export const DateTimeTool: AiTool = {
 	name: 'get_datetime',
-	description: 'Get current date and time',
+	description: 'Get current UTC date / time',
 	args: {},
-	fn: async () => new Date().toISOString()
+	fn: async () => new Date().toUTCString()
 }
 
 export const ExecTool: AiTool = {
@@ -55,15 +57,15 @@ export const ExecTool: AiTool = {
 		language: {type: 'string', description: 'Execution language', enum: ['cli', 'node', 'python'], required: true},
 		code: {type: 'string', description: 'Code to execute', required: true}
 	},
-	fn: async (args, ai) => {
+	fn: async (args, stream, ai) => {
 		try {
 			switch(args.type) {
 				case 'bash':
-					return await CliTool.fn({command: args.code}, ai);
+					return await CliTool.fn({command: args.code}, stream, ai);
 				case 'node':
-					return await JSTool.fn({code: args.code}, ai);
+					return await JSTool.fn({code: args.code}, stream, ai);
 				case 'python': {
-					return await PythonTool.fn({code: args.code}, ai);
+					return await PythonTool.fn({code: args.code}, stream, ai);
 				}
 			}
 		} catch(err: any) {
@@ -111,9 +113,43 @@ 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},

src/vision.ts

@@ -1,5 +1,5 @@
 import {createWorker} from 'tesseract.js';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 
 export class Vision {
 
@@ -8,18 +8,16 @@ export class Vision {
 	/**
 	 * Convert image to text using Optical Character Recognition
 	 * @param {string} path Path to image
-	 * @returns {{abort: Function, response: Promise<string | null>}} Abort function & Promise of extracted text
+	 * @returns {AbortablePromise<string | null>} Promise of extracted text with abort method
 	 */
-	ocr(path: string): {abort: () => void, response: Promise<string | null>} {
+	ocr(path: string): AbortablePromise<string | null> {
 		let worker: any;
-		return {
-			abort: () => { worker?.terminate(); },
-			response: new Promise(async res => {
+		const p = new Promise<string | null>(async res => {
 			worker = await createWorker(this.ai.options.tesseract?.model || 'eng', 2, {cachePath: this.ai.options.path});
 			const {data} = await worker.recognize(path);
 			await worker.terminate();
 			res(data.text.trim() || null);
-			})
-		}
+		});
+		return Object.assign(p, {abort: () => worker?.terminate()});
 	}
 }