Files
ai-utils/src/llm.ts
ztimson 28904cddbe
All checks were successful
Publish Library / Build NPM Project (push) Successful in 49s
Publish Library / Tag Version (push) Successful in 16s
TTS
2026-01-30 15:39:29 -05:00

263 lines
9.1 KiB
TypeScript

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';
import {Worker} from 'worker_threads';
import {fileURLToPath} from 'url';
import {dirname, join} from 'path';
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;
}
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, 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
}
}
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} = {};
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.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);
}
/**
* 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.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]} / ${model[1]}`);
return this.providers[model[0]].ask(message, {...options, model: model[1]});
}
/**
* Compress chat history to reduce context size
* @param {LLMMessage[]} history Chatlog that will be compressed
* @param max Trigger compression once context is larger than max
* @param min Summarize until context size is less than min
* @param {LLMRequest} options LLM options
* @returns {Promise<LLMMessage[]>} New chat history will summary at index 0
*/
async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<LLMMessage[]> {
if(this.estimateTokens(history) < max) return history;
let keep = 0, tokens = 0;
for(let m of history.toReversed()) {
tokens += this.estimateTokens(m.content);
if(tokens < min) keep++;
else break;
}
if(history.length <= keep) return history;
const recent = keep == 0 ? [] : history.slice(-keep),
process = (keep == 0 ? history : history.slice(0, -keep)).filter(h => h.role === 'assistant' || h.role === 'user');
const summary = await this.summarize(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), 250, options);
return [{role: 'assistant', content: `Conversation Summary: ${summary}`, timestamp: Date.now()}, ...recent];
}
cosineSimilarity(v1: number[], v2: number[]): number {
if (v1.length !== v2.length) throw new Error('Vectors must be same length');
let dotProduct = 0, normA = 0, normB = 0;
for (let i = 0; i < v1.length; i++) {
dotProduct += v1[i] * v2[i];
normA += v1[i] * v1[i];
normB += v2[i] * v2[i];
}
const denominator = Math.sqrt(normA) * Math.sqrt(normB);
return denominator === 0 ? 0 : dotProduct / denominator;
}
chunk(target: object | string, maxTokens = 500, overlapTokens = 50): string[] {
const objString = (obj: any, path = ''): string[] => {
if(!obj) return [];
return Object.entries(obj).flatMap(([key, value]) => {
const p = path ? `${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;
}
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 });
});
};
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 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) {
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);
}
}