Fabrício Raphael fabriciorsf

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.

Add user `ssh-agent` as daemon to Ubuntu 18.04LTS server

Create a new systemd user unit, which starts ssh-agent upon login to server. Will remain resident until the final session for the user has logged out.

Steps

Create /etc/systemd/user/ssh-agent.service.
Run the following commands (under your user account, not root) to install the systemd unit and start:

https://linuxconfig.org/overclock-your-radeon-gpu-with-amdgpu

Install sensors:

sudo apt install lm-sensors

Edit device files

sudo nano /sys/class/drm/card2/device/pp_sclk_od

Download latest AMD APP SDK from this page: http://developer.amd.com/tools-and-sdks/opencl-zone/amd-accelerated-parallel-processing-app-sdk/
Open terminal and write:

cd
cd Downloads
tar xvf AMD-APP-SDKInstaller-v3.0.130.136-GA-linux64.tar.bz2
sudo sh ./AMD-APP-SDK-v3.0.130.136-GA-linux64.sh

It should look like this:

Some notes on how to setup various open-source XMR Miners on Ubuntu 16

Using Systemd

setup

Running AMD Radeon GPUS on Ubuntu 16.04

Cards tested (so far):

Model	Released
Radeon RX 550	April 2017
Radeon RX 560	April 2017
Radeon HD 7800 Series	March 2012

Install 17.40.2712 Beta Linux Driver

Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank

Introduction

The paper introduces:
- Sentiment Sentiment Treebank - A dataset containing 215,154 phrases with fine-grained sentiment labels (5 classes).
- Recursive Neural Tensor Network - Model to learn these fine-grained sentiment labels.
Link to the paper

	import tensorflow as tf
	def conv1d(input_, output_size, width, stride):
	'''
	:param input_: A tensor of embedded tokens with shape [batch_size,max_length,embedding_size]
	:param output_size: The number of feature maps we'd like to calculate
	:param width: The filter width
	:param stride: The stride
	:return: A tensor of the concolved input with shape [batch_size,max_length,output_size]
	'''
	inputSize = input_.get_shape()[-1] # How many channels on the input (The size of our embedding for instance)

	import keras
	from keras.models import Model
	import keras.backend as K
	from keras.engine.topology import Layer, InputLayer, Input, Node
	from keras.layers import merge, Dense, TimeDistributed, LSTM
	import theano.tensor as T
	import theano

	import numpy as np

	# Reuters-21578 dataset downloader and parser
	#
	# Author: Eustache Diemert <eustache@diemert.fr>
	# http://scikit-learn.org/stable/auto_examples/applications/plot_out_of_core_classification.html
	#
	# Modified by @herrfz, get pandas DataFrame from the orig SGML
	# License: BSD 3 clause

	from __future__ import print_function