ai-utils/src/llm.ts

import {JSONAttemptParse} from '@ztimson/utils';
import {AbortablePromise, Ai} from './ai.ts';
import {Anthropic} from './antrhopic.ts';
import {OpenAi} from './open-ai.ts';
import {LLMProvider} from './provider.ts';
import {AiTool} from './tools.ts';
import {Worker} from 'worker_threads';
import {fileURLToPath} from 'url';
import {dirname, join} from 'path';

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';
	/** Message content */
	content: string | any;
	/** Timestamp */
	timestamp?: number;
} | {
	/** Tool call */
	role: 'tool';
	/** Unique ID for call */
	id: string;
	/** Tool that was run */
	name: string;
	/** Tool arguments */
	args: any;
	/** Tool result */
	content: undefined | string;
	/** Tool error */
	error?: undefined | string;
	/** Timestamp */
	timestamp?: number;
}

/** Background information the AI will be fed */
export type LLMMemory = {
	/** What entity is this fact about */
	owner: string;
	/** The information that will be remembered */
	fact: string;
	/** Owner and fact embedding vector */
	embeddings: [number[], number[]];
	/** Creation time */
	timestamp: Date;
}

export type LLMRequest = {
	/** System prompt */
	system?: string;
	/** Message history */
	history?: LLMMessage[];
	/** Max tokens for request */
	max_tokens?: number;
	/** 0 = Rigid Logic, 1 = Balanced, 2 = Hyper Creative **/
	temperature?: number;
	/** Available tools */
	tools?: AiTool[];
	/** LLM  model */
	model?: string;
	/** Stream response */
	stream?: (chunk: {text?: string, tool?: string, done?: true}) => any;
	/** Compress old messages in the chat to free up context */
	compress?: {
		/** Trigger chat compression once context exceeds the token count */
		max: number;
		/** Compress chat until context size smaller than */
		min: number
	},
	/** Background information the AI will be fed */
	memory?: LLMMemory[],
}

class LLM {
	private embedWorker: Worker | null = null;
	private embedQueue = new Map<number, { resolve: (value: number[]) => void; reject: (error: any) => void }>();
	private embedId = 0;
	private models: {[model: string]: LLMProvider} = {};
	private defaultModel!: string;

	constructor(public readonly ai: Ai) {
		this.embedWorker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
		this.embedWorker.on('message', ({ id, embedding }) => {
			const pending = this.embedQueue.get(id);
			if (pending) {
				pending.resolve(embedding);
				this.embedQueue.delete(id);
			}
		});

		if(!ai.options.llm?.models) return;
		Object.entries(ai.options.llm.models).forEach(([model, config]) => {
			if(!this.defaultModel) this.defaultModel = model;
			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<string>}} Function to abort response and chat history
	 */
	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);
				}

				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);
					}
				}];
			}

			// 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 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<{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, 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: 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;
		for (let i = 0; i < v1.length; i++) {
			dotProduct += v1[i] * v2[i];
			normA += v1[i] * v1[i];
			normB += v2[i] * v2[i];
		}
		const denominator = Math.sqrt(normA) * Math.sqrt(normB);
		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 [];
			return Object.entries(obj).flatMap(([key, value]) => {
				const p = path ? `${path}${isNaN(+key) ? `.${key}` : `[${key}]`}` : key;
				if(typeof value === 'object' && !Array.isArray(value)) return objString(value, p);
				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[] = [];
		for(let i = 0; i < tokens.length;) {
			let text = '', j = i;
			while(j < tokens.length) {
				const next = text + (text ? ' ' : '') + tokens[j];
				if(this.estimateTokens(next.replace(/\s*\n\s*/g, '\n')) > maxTokens && text) break;
				text = next;
				j++;
			}
			const clean = text.replace(/\s*\n\s*/g, '\n').trim();
			if(clean) chunks.push(clean);
			i = Math.max(j - overlapTokens, j === i ? i + 1 : j);
		}
		return chunks;
	}

	/**
	 * 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) => {
				const id = this.embedId++;
				this.embedQueue.set(id, { resolve, reject });
				this.embedWorker?.postMessage({
					id,
					text,
					model: this.ai.options?.embedder || 'bge-small-en-v1.5',
					path: this.ai.options.path
				});
			});
		};
		const chunks = this.chunk(target, maxTokens, overlapTokens);
		return Promise.all(chunks.map(async (text, index) => ({
			index,
			embedding: await embed(text),
			text,
			tokens: this.estimateTokens(text),
		})));
	}

	/**
	 * Estimate variable as tokens
	 * @param history Object to size
	 * @returns {number} Rough token count
	 */
	estimateTokens(history: any): number {
		const text = JSON.stringify(history);
		return Math.ceil((text.length / 4) * 1.2);
	}

	/**
	 * Compare the difference between two strings using tensor math
	 * @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
	 */
	fuzzyMatch(target: string, ...searchTerms: string[]) {
		if(searchTerms.length < 2) throw new Error('Requires at least 2 strings to compare');
		const vector = (text: string, dimensions: number = 10): number[] => {
			return text.toLowerCase().split('').map((char, index) =>
				(char.charCodeAt(0) * (index + 1)) % dimensions / dimensions).slice(0, dimensions);
		}
		const v = vector(target);
		const similarities = searchTerms.map(t => vector(t)).map(refVector => this.cosineSimilarity(v, refVector))
		return {avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length, max: Math.max(...similarities), similarities}
	}

	/**
	 * Ask a question with JSON response
	 * @param {string} message Question
	 * @param {LLMRequest} options Configuration options and chat history
	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
	 */
	async json(message: string, options?: LLMRequest): Promise<any> {
		let resp = await this.ask(message, {system: 'Respond using a JSON blob matching any provided examples', ...options});
		if(!resp) return {};
		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
		const jsonStr = codeBlock ? codeBlock[1].trim() : resp;
		return JSONAttemptParse(jsonStr, {});
	}

	/**
	 * Create a summary of some text
	 * @param {string} text Text to summarize
	 * @param {number} tokens Max number of tokens
	 * @param options LLM request options
	 * @returns {Promise<string>} Summary
	 */
	summarize(text: string, tokens: number, options?: LLMRequest): Promise<string | null> {
		return this.ask(text, {system: `Generate a brief summary <= ${tokens} tokens. Output nothing else`, temperature: 0.3, ...options});
	}
}

export default LLM;