Alex Liberzon alexlib

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.

Below are the generic details and resources of the machine I used:

Operating System: Linux Mint 22.3 - MATE 64-bit (safe-graphics / boot with 'nomodeset') - Linux Kernel: 6.14.0-37-generic
GPU: not relevant / not used
CPU: 2 cores – Intel® Core™2 Duo CPU E8400 @ 3.00GHz × 2
RAM: 8 GB (DDR2)

Given my recent personal good experience in running inference with CPU-only (no GPU) on local models up to 2B-4B parameters with more than acceptable speed and performance on really limited hardware and very little RAM, I decided to share the instructions and commands I used to compile llama.cpp on my machine and the arguments I currently use to launch llama.

📝 Scientific Manuscript Revision Checklist

Based on the Simons "Musings on Writing" Framework

🛠 Phase 1: The Physical Intervention

Physical Printout: Have you printed this draft and read it on paper? (Simons Rule: Repetition and flow issues are invisible on a screen).
The Reverse Outline: Have you written a single phrase summarizing the point of every paragraph?
- If you can't summarize it in one phrase, the paragraph needs revision.

🏛 Phase 2: Structure & Organization

The Opening: Does the paper establish a controversy, mystery, or conundrum? (Avoid "laundry list" introductions).

	#!/usr/bin/env bash
	set -euo pipefail

	# Config
	USER_LOGIN="${1:-alexlib}"
	REPORT_FILE="${2:-fork-sync-report.md}"
	SYNC="${SYNC:-1}" # SYNC=1 to sync clean forks, SYNC=0 to only report
	PER_PAGE="${PER_PAGE:-100}" # pagination size

	tmp_json="$(mktemp)"

	# /// script
	# requires-python = ">=3.14"
	# dependencies = [
	# "marimo>=0.19.9",
	# "matplotlib==3.10.8",
	# "numpy==2.4.2",
	# "pydantic-ai-slim==1.57.0",
	# "scipy==1.17.0",
	# ]
	# ///

	import marimo

	__generated_with = "0.19.6"
	app = marimo.App(width="medium", auto_download=["ipynb"])

	with app.setup:
	import cv2
	import numpy as np
	import matplotlib.pyplot as plt
	import marimo as mo

	import marimo

	__generated_with = "0.19.4"
	app = marimo.App(width="medium", auto_download=["ipynb"])


	@app.cell
	def _():
	import marimo as mo
	import cv2

	# /// script
	# requires-python = ">=3.13"
	# dependencies = [
	# "google-genai==1.56.0",
	# "matplotlib==3.10.8",
	# "nbformat==5.10.4",
	# "numpy==2.2.6",
	# "opencv-python==4.12.0.88",
	# "pandas==2.3.3",
	# ]