Behnam Moh ibehnam

Extracting financial disclosure reports and police blotter narratives using OpenAI's Structured Output

tl;dr this demo shows how to call OpenAI's gpt-4o-mini model, provide it with URL of a screenshot of a document, and extract data that follows a schema you define. The results are pretty solid even with little effort in defining the data — and no effort doing data prep. OpenAI's API could be a cost-efficient tool for large scale data gathering projects involving public documents.

OpenAI announced Structured Outputs for its API, a feature that allows users to specify the fields and schema of extracted data, and guarantees that the JSON output will follow that specification.

For example, given a Congressional financial disclosure report, with assets defined in a table like this:

Purpose

Bootstrap knowledge of LLMs ASAP. With a bias/focus to GPT.

Avoid being a link dump. Try to provide only valuable well tuned information.

Prelude

Neural network links before starting with transformers.

Karabiner layouts for symbols and navigation

Gavin Sinclair, January 2022

Introduction

I use Karabiner (configured with Gosu) to make advanced key mappings on my Apple computer. Karabiner allows you to create “layers”, perhaps simulating those on a programmable mechanical keyboard. I make good use of these layers to give me easy access (home-row or nearby) to all symbols and navigational controls, and even a numpad.

The motivation is to keep hand movement to a minimum. Decades of coding on standard keyboards has unfortunately left me with hand and wrist pain. I will soon enough own a small split keyboard which will force me to use layers to access symbols etc., so this Karabiner solution, which has evolved over months, is a training run for that.

	#cloud-config

	# Enable automatic package updates and upgrades during cloud-init execution
	package_update: true
	package_upgrade: true

	packages:
	# Security and Hardening
	- ufw
	- fail2ban


	Jeff has two brothers and each of his brothers has three sisters and each of
	the sisters has four step brothers. How many step brothers does each brother
	have?

	<think> Okay, so I've got this problem: Jeff has two brothers, and each of his
	brothers has three sisters. Then it says that each of the sisters has four step
	brothers. The question is asking how many step brothers each brother has.

	Hmm, let me start by breaking down the information piece by piece. First, Jeff

	Question: Jeff has two brothers and each of his brothers has three sisters
	and each of the sisters has four step brothers. How many step brothers does
	each brother have?


	Model's Response:


	<think> okay, let's try to figure out this problem step by step. so jeff has
	two brothers. that means there are three boys in total: jeff and his two

	font-family = BerkeleyMono Nerd Font
	#font-family = Iosevka Nerd Font
	# font-family = SFMono Nerd Font
	font-size = 20
	theme = GruvboxDarkHard
	shell-integration-features = no-cursor,sudo,no-title
	cursor-style = block
	adjust-cell-height = 35%
	# background-opacity = 0.96

	{
	;; :!Ca is keycode :a and prefix a with !C

	;; C \| left_command
	;; T \| left_control
	;; O \| left_option
	;; S \| left_shift
	;; F \| fn
	;; Q \| right_command
	;; W \| right_control

	window:
	# Window dimensions (changes require restart)
	#
	# Number of lines/columns (not pixels) in the terminal. The number of columns
	# must be at least `2`, while using a value of `0` for columns and lines will
	# fall back to the window manager's recommended size.
	dimensions:
	columns: 500
	lines: 30

	"""
	Lisp languages support an image-based development cycle at the REPL.

	R - having its roots in Lisp - it was first implemented in Scheme - has overtaken this.
	R's base command `save.image(filepath)` can save the state of an entire REPL session in a file,
	and next time you start R, you can start where you stopped, by hitting `load(filepath)`.

	I was asking myself, since Python's REPL is ipython, what supports
	similar functionailities for pythonistas.