Volodymyr wowkin2

LLM Wiki

A pattern for building personal knowledge bases using LLMs.

This is an idea file, it is designed to be copy pasted to your own LLM Agent (e.g. OpenAI Codex, Claude Code, OpenCode / Pi, or etc.). Its goal is to communicate the high level idea, but your agent will build out the specifics in collaboration with you.

The core idea

Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation. Ask a subtle question that requires synthesizing five documents, and the LLM has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM, ChatGPT file uploads, and most RAG systems work this way.

Setup

On every machine in the cluster install openmpi and mlx-lm:

conda install conda-forge::openmpi
pip install -U mlx-lm

Next download the pipeline parallel run script. Download it to the same path on every machine:

	#!/usr/bin/env python3
	"""
	Optimal randomized binary search strategy for Steve Ballmer's guessing game.

	Game: opponent picks `x` in `[1, 100]`, you guess with higher/lower feedback.
	Payoff = $6 - (number of guesses). Profit if you find it in 5 guesses or fewer.

	Solves the minimax LP over a pool of deterministic BST strategies to find
	the optimal randomized mixture. Verifies via Monte Carlo.


	import claude from "@anthropic-ai/tokenizer/claude.json";
	import { Tiktoken, TiktokenBPE } from "js-tiktoken";

	export function countTokens(text: string): number {
	const tokenizer = getTokenizer();
	const encoded = tokenizer.encode(text.normalize("NFKC"));
	return encoded.length;
	}

	#!/bin/bash

	# System information
	hostname=$(hostname)
	kernel_version=$(uname -r)
	uptime=$(uptime -p)

	# Memory information
	memory_total_kb=$(grep MemTotal /proc/meminfo \| awk '{print $2}')
	memory_total_mb=$((memory_total_kb / 1024))

	from django.core.management.base import BaseCommand

	from rq.registry import FailedJobRegistry
	from django_rq.queues import get_queue
	from django_rq.utils import get_jobs


	class Command(BaseCommand):
	help = 'Delete failed jobs from Django RQ.'

	import os
	import shlex
	import subprocess

	from django.core.management.base import BaseCommand
	from django.utils import autoreload


	class Command(BaseCommand):
	def add_arguments(self, parser):

	#!/usr/bin/python3.7

	import asyncio
	import ipaddress
	import re
	import sys


	MAX_NUMBER_WORKERS = 200

	import time
	import asyncio
	import concurrent.futures
	import requests


	URL_SESSION = 'http://localhost:8061/session'
	URL_REQUESTS = 'http://localhost:8061/requests'
	URL_REQUESTS_EXECUTOR = 'http://localhost:8061/requests_executor'

	// import_json_appsscript.js
	// https://gist.github.com/allenyllee/c764c86ed722417948fc256b7a5077c4
	//
	// Changelog:
	// (Oct. 16 2019) tag: allenyllee-20191016
	// 1. Fixed google script error: urlfetchapp - service invoked too many times https://stackoverflow.com/questions/10598179/google-apps-script-urlfetchapp-service-invoked-too-many-times
	// (Jul. 16 2018) tag: allenyllee-20180716
	// 1. Fixed the issue "If you try to query /arrayA[k]/arrayB[n]/arrayC[m]/.../member, you will always get /arrayA[k]/arrayB[k]/arrayC[k]/.../member."
	// (Nov. 30 2017) tag: allenyllee-20171130
	// 1. Add the ability to query array elements by using xpath like "/array[n]/member" where "n" is array index