Field manual

How to run your own investigation

A step-by-step for journalists, researchers, and anyone sitting on a fresh FOIA/RTI release — from raw documents to a graph you can question.

01

Get your release

File (or already have) a FOIA/RTI request and receive the government's response — a PDF bundle, a folder of scanned letters, or a plain-text dump. FOIAtlas works on the released files as-is, redactions and all.

02

Split it into documents

One file per record works best — a memo, a single email, one set of meeting minutes. If your release is one giant merged PDF, split it into logical documents first (any PDF tool) so each becomes its own node with its own date and source.

03

Drop them on /upload

Drag in as many PDFs or .txt files as you have — or paste text straight from an OCR tool or a copy-paste. Watch the live pipeline: parse → extract → resolve → embed → write, per file, with a running node/edge count.

04

Let Gemini read for you

Each document is read against the graph schema: people, organizations, meetings, topics — and every black bar becomes a RedactedSpan carrying its exemption code and surrounding context. Aliases resolve automatically across every file you've uploaded.

05

Ask, don't search

On /chat, ask in plain English — “who was at the meeting before the vote?”, “what’s hidden behind the b(5) redactions?”. Answers are grounded in your specific documents and cite them by title.

06

Walk the graph

On /explorer, see the whole network — filter by type, hover any node to spotlight its connections, and follow a redaction back to the meeting or memo it sits in.

07

Clean up as you go

Extraction isn't perfect — if it merges two people who share a name, or misses an obvious link, just say so in chat: “rename this node”, “these are different people, split them”, “connect X to Y”. The agent edits the graph on your instruction.

Accepted formats

What you can drop on /upload

PDF

Native or scanned — pdf-parse extracts the text layer. Scanned images with no text layer need OCR first (see below).

.txt

Plain text — the fastest path if you've already transcribed or OCR'd a document.

Pasted text

No file at all: paste straight into the /upload text box for a one-off email or a quote you're tracking down.

Scanned, no text layer?

Run it through an OCR tool first (Adobe Acrobat, Google Drive's built-in OCR, or a free tool like OCRmyPDF) so the text is selectable, then upload the result — FOIAtlas needs machine- readable text, not just an image of a page.

Where to find real releases

Practice on the real thing

Government FOIA portals

foia.gov (federal agencies), agency-specific reading rooms, and state/local portals.

RTI replies

The scanned PDF an Indian public authority sends back under Section 6/7 of the RTI Act.

MuckRock / DocumentCloud

Crowdsourced FOIA request archives — thousands of already-released sets to practice on.

FBI Vault / National Archives

Bulk historical releases, often merged into single large PDFs — split before uploading.

Three things that make extraction better

  • Keep document titles descriptive ("Memo — Atwater Mill site assessment, 2024-03-12") — they become the node label and the chat citation.
  • Upload chronologically when you can. Dates seed the Event/Document nodes and make “what happened before X?” answerable.
  • One redaction per exemption code, if the source lets you tell — it keeps investigative leads distinct instead of merged.