APOC Data

Alaska campaign financial disclosure data from the Alaska Public Offices Commission.

This scrapes the CSV files from the APOC website once a day and uploads them to this repo's releases.

https://nickcrews.github.io/apoc-data

Who gave money to which candidates, and what they spent it on — click to filter, no SQL required. Pick a donor and every other chart, the totals, and the table of matching records narrow to them; drag across the timeline to pick a date range; flip between money in and money out.

It's a SQLRooms app running DuckDB-WASM: the data is downloaded into your browser and queried there, so nothing you do is sent anywhere. There's a SQL editor if you want one, and an "Ask an AI" panel that hands you a copy-pasteable prompt (schema and data URLs included) for questions the dashboard doesn't answer.

https://nickcrews.github.io/apoc-data/llms.txt

Hand that URL to any agent that can fetch a URL — Claude Code, Cursor, Codex, a ChatGPT with browsing on — and it has everything it needs: the parquet URLs, the full schema with row counts and date ranges, and the caveats that otherwise produce a confident wrong answer (names spelled six different ways, amended reports double-counting, report_year not being the year the money moved).

It doesn't repeat the schema — that's in manifest.json, which it links, and any reader that can run the query in it can equally run DESCRIBE. What it carries is the part that lives nowhere else.

For tools that can't fetch a URL — ChatGPT and Gemini, mostly — the "Ask an AI" button on the site copies the same guidance as a prompt, with the full schema spelled out inline and your question filled in, and points you at the CSV zip to attach.

There's also an AGENTS.md for agents working on this repo rather than on the data.

You can download the daily-scraped CSVs from the GitHub releases (this is what most users want).

Browse from this repo's releases. The latest release is also hosted on the site, as both CSV and parquet. Every file gets its own tab there, with its size and row count, and the rows themselves: search them, or filter from the distributions in the column headers — eg income.

Every CSV in one archive is at https://nickcrews.github.io/apoc-data/apoc-csvs.zip. That's the one to grab if you're feeding the data to ChatGPT or Gemini, which can't fetch a URL and only see the files you attach to the chat.

Using uv's uvx:

uvx apoc-data release download # downloads all files from the latest release to ./downloads/ folder
uvx apoc-data release download "20260702-125614" --destination mydownloads/ # specify explicitly
uvx apoc-data asset download debt.csv --destination apoc_debt.csv # download a single file
uvx apoc-data release list # see what releases are available
uvx apoc-data asset list --json # see what files are in the latest release, as JSON

Or, you can download these CSVs directly using the direct URLs from the releases page using curl, pandas, ibis, whatever!

curl -L https://github.com/NickCrews/apoc-data/releases/latest/download/candidate_registration.csv > candidate_registration.csv # get latest
curl -L https://github.com/NickCrews/apoc-data/releases/download/20240716-025636/candidate_registration.csv > candidate_registration.csv # or a different url pattern for specific releases

# query directly using duckdb
duckdb -c "SELECT count(*) FROM 'https://github.com/NickCrews/apoc-data/releases/latest/download/candidate_registration.csv'"
duckdb -c "SELECT count(*) FROM 'https://github.com/NickCrews/apoc-data/releases/download/20240716-025636/candidate_registration.csv'"

The same data is also published as parquet on the GitHub Pages site, with the strings parsed into real types: "$1,234.00" becomes a DECIMAL(18,2), "7/16/2024" becomes a DATE, and income.csv goes from 369MB to about 23MB. DuckDB reads them straight over HTTPS, and only fetches the columns and row groups your query touches:

duckdb -c "SELECT filer_name, sum(amount) AS raised
           FROM 'https://nickcrews.github.io/apoc-data/data/income.parquet'
           WHERE report_year = 2024
           GROUP BY 1 ORDER BY raised DESC LIMIT 10"

data/manifest.json lists every table with its row count, columns and types, and date range, plus the release the data came from, whose published_at is when it was scraped.

Note that the GitHub release URLs above don't send CORS headers, so they can't be fetched directly from a browser (eg from duckdb-wasm or a web app). That's why the latest release is also mirrored to GitHub Pages, which does allow cross-origin requests — it's what makes the in-browser explorer possible. eg at https://shell.duckdb.org/:

• this works: SELECT count(*) FROM 'https://nickcrews.github.io/apoc-data/campaign_form.csv'
• this does not: SELECT count(*) FROM 'https://github.com/NickCrews/apoc-data/releases/latest/download/campaign_form.csv'

We provide a python API too. uv add apoc-data and then

from apoc_data.releases import asset_download, release_download, release_list

release_list()  # all releases, newest first, as `Release` objects
release_download(destination="downloads/")  # all files from the latest release
asset_download("debt.csv", destination="apoc_debt.csv")  # a single file

You can also scrape fresh data directly from the APOC website (requires the scrape extra for playwright):

uvx "apoc-data[scrape]" scrape --directory scraped/

There is also a python API. Read the source code.

MIT, do as you wish with the data and code!

Create venv and install dev deps:

uv sync

scrape:

uv run apoc-data scrape --directory downloads --no-headless

convert the scraped CSVs to typed parquet (plus a manifest.json describing them):

uv run apoc-data convert --source scraped/ --destination _site/data/

The type of each column is detected rather than hardcoded, so new or renamed APOC columns keep working. See src/apoc_data/convert/.

The explorer lives in web/ — Vite + React + SQLRooms over DuckDB-WASM.

cd web
npm install
npm run data   # convert ../scraped into web/public/data (once)
npm run dev

Without local data, point it at the live site instead:

VITE_DATA_BASE_URL=https://nickcrews.github.io/apoc-data/data/ npm run dev

scripts/build_site.sh assembles everything that gets published, so the layout is reproducible outside CI:

scripts/build_site.sh scraped _site

_site/
  index.html, assets/   the explorer app, which lists every file below
  data/*.parquet        typed parquet + manifest.json
  *.csv, *.csv.zip      the raw scrape, at the URLs it has always had
  apoc-csvs.zip         every CSV in one file, to upload somewhere
  llms.txt              the schema and caveats, for AI agents

llms.txt and the "Ask an AI" prompt are two renderings of one source, web/src/aiGuide.ts — edit the caveats there, never in llms.txt, and both change together. A copy of llms.txt is committed at the repo root so its wording is reviewable in a diff; refresh it with

cd web && npx tsx ../scripts/gen-llms-txt.ts

It holds no row counts or timestamps, so it only changes when you change it.