Wilson Freitas wilsonfreitas

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.

Lecture 1: Introduction to Research — [📝Lecture Notebooks] [▶️Video]
Lecture 2: Introduction to Python — [📝Lecture Notebooks] [▶️Video]
Lecture 3: Introduction to NumPy — [📝Lecture Notebooks] [▶️Video]
Lecture 4: Introduction to pandas — [📝Lecture Notebooks] [▶️Video]
Lecture 5: Plotting Data — [📝Lecture Notebooks] [[▶️Vide

A little experiment using restarts.

(And while we’re at it, let’s torture R’s syntax a little.)

![screenshot][]

In the following we will be using R’s “restarts” feature to implement the state machine that drives generators in languages such as Python. Generators allow lazily generating values on demand: a consumer invokes a generator, and consumes values as they are produced. A new value is only produced once the previous one has been consumed.

	f <- function(x) cos(x*20)
	expi <- function(x) cos(x) + sin(x)*1i
	par(mfrow = c(1,2))
	L <- 40
	medias_Re <- c()
	fs <- c()
	as <- seq(0.01, 1, l = 140)

	animation::saveGIF({
	for(a in as) {

	# =======================================
	# = Enhancements to data tidying =
	# = Hadley Wickham =
	# = https://rstd.io/tidyhancements-2019 =
	# =======================================

	# What is tidy data? ----------------------------------------------------------
	# 1. Each column is a variable.
	# 2. Each row is an observation.
	# 3. Each cell is a value.

	---
	title: "The Best Tacos in Austin according to Caitlin Hudon"
	output:
	flexdashboard::flex_dashboard:
	orientation: column
	vertical_layout: fill
	---

	```{r setup, include=FALSE}
	library(flexdashboard)

	upper <- seq(0, pi, length.out = 181)[-181]
	upper <-cbind(x = cos(upper), y = sin(upper))
	lower <- seq(pi, 2*pi, length.out = 181)[-181]
	lower <- cbind(x = cos(lower), y = sin(lower))
	right <- seq(1.5pi, 2.5pi, length.out = 181)[-181]
	right <- cbind(x = cos(right), y = sin(right))
	left <- seq(0.5pi, 1.5pi, length.out = 181)[-181]
	left <- cbind(x = cos(left), y = sin(left))
	full <- cbind(x = cos(seq(0, 2*pi, length.out = 361)[-361]),
	y = sin(seq(0, 2*pi, length.out = 361)[-361]))



	library(tidyverse)


	# Colliders

	x1 <- rnorm(1000)
	y <- 1.2 + 1.6 * x1 + rnorm(1000)
	z <- 0.6 + 1.3x1 + 1.5y + rnorm(1000)

	#!/usr/bin/env python
	# -- coding: UTF-8 --

	# Função para validar sequência de dígitos do Renavam.
	# Adaptado por Fabricio Biazzotto, a partir do original, feito em java:
	# http://blog.victorjabur.com/2010/05/28/renavam_veiculos_java


	def validar_renavam(renavam):
	# Retorna falso caso o valor passado não seja um número inteiro

	import asyncio
	import json

	import aiohttp


	SCHOOL_URL_FMT = 'http://educacao.dadosabertosbr.com/api/escola/{}'
	SEARCH_SCHOOL_URL = (
	'http://educacao.dadosabertosbr.com/api/escolas/buscaavancada?'
	'situacaoFuncionamento=1&energiaInexistente=on&aguaInexistente=on&'