ai-utils/src/llm.ts

import {JSONAttemptParse} from '@ztimson/utils';
import {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 {AiTool} from './tools.ts';

export type LLMMessage = {
	/** Message originator */
	role: 'assistant' | 'system' | 'user';
	/** Message content */
	content: string | any;
} | {
	/** 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;
}

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;
	/** 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 | [string, string];
	/** Stream response */
	stream?: (chunk: {text?: 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
	}
}

export class LLM {
	private providers: {[key: string]: LLMProvider} = {};

	constructor(public readonly ai: Ai, public readonly options: LLMOptions) {
		if(options.anthropic?.token) this.providers.anthropic = new Anthropic(this.ai, options.anthropic.token, options.anthropic.model);
		if(options.ollama?.host) this.providers.ollama = new Ollama(this.ai, options.ollama.host, options.ollama.model);
		if(options.openAi?.token) this.providers.openAi = new OpenAi(this.ai, options.openAi.token, options.openAi.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
	 */
	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.options)[options.model]?.model];
		}
		if(!options.model || model[1] == null) {
			if(typeof this.options.model == 'object') model = this.options.model;
			else model = [this.options.model, (<any>this.options)[this.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]});
	}

	/**
	 * 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 {LLMRequest} options LLM options
	 * @returns {Promise<LLMMessage[]>} New chat history will summary at index 0
	 */
	async compress(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<LLMMessage[]> {
		if(this.estimateTokens(history) < max) return history;
		let keep = 0, tokens = 0;
		for(let m of history.toReversed()) {
			tokens += this.estimateTokens(m.content);
			if(tokens < min) keep++;
			else break;
		}
		if(history.length <= keep) return history;
		const 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}`}, ...recent];
	}

	/**
	 * 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);
	}

	/**
	 * 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) {
		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), {});
	}

	/**
	 * 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})
			.then(history => <string>history.pop()?.content || null);
	}
}