dosar-radar

Find your case in the Romanian citizenship orders — in one search box.

A self-hosted search index over every art. 11 order published by the Romanian Ministry of Justice: 315,173 case numbers extracted from 2,214 PDF orders covering 2018–2026 — as of July 2026, and growing with every sync.

🇬🇧 English · 🇺🇦 Українська

---

The problem

If you have applied for Romanian citizenship under article 11, the decision on your file is published on cetatenie.just.ro — as a PDF attached to a ministerial order. There is no search. There is no index. There is no notification.

There are over two thousand such PDFs. To find out whether your case number appears in any of them, you are expected to open them one by one and read.

dosar-radar does that once, and turns it into a search box.

What it does

🔎 Search by case number	29293/2023, 48031/RD/2023, or just 29293 — all find the same file
📄 Full order index	Every order with date, number, case count and a link to the original PDF
🗂 Per-order case lists	Open any order and see every case number printed in it
🔄 Keeps itself current	Re-syncs every 3 hours on its own, parsing only what is new
🖥 CLI	Same search and sync without the web UI — useful in scripts and cron
🐳 Docker-first	make up and you have it; data lives in a named volume

Quick start

git clone https://github.com/paulfcdd/dosar-radar.git

cd dosar-radar && make up

The UI is on http://localhost:8000 — but the database starts empty. Fill it:

make sync

The first sync downloads and parses ~2,200 PDFs. Depending on your connection this takes roughly 20–40 minutes; every later run only touches orders that are new or previously failed, so it finishes in seconds. Watch progress in the web UI or in make logs.

Want to check the setup works before committing to a full sync?

make sync-fast

That parses only the 20 most recent orders.

How it works

cetatenie.just.ro
       │
       │  ① HTTP session that solves the site's proof-of-work challenge
       ▼
  index page ──② parse HTML──▶ 2,214 orders (date, number, PDF url)
       │
       │  ③ download PDFs, 4 in parallel
       ▼
   pdfplumber ──④ regex + noise filter──▶ 315,173 case numbers
       │
       ▼
    SQLite (WAL) ──▶ Flask UI  ·  CLI

Four parts are worth a closer look.

① The challenge. The ministry's site sits behind an anti-bot page that answers 503 with a JavaScript puzzle instead of content. It is a hashcash-style proof of work: the page carries a hex token c, and the client must find a counter i such that SHA1(c + str(i)) has the byte 0xb0 at position n and 0x0b at n+1, where n is the first hex digit of the token. The answer goes back as a cookie. session.py solves it in pure Python and caches the cookie for the rest of the run — no browser, no JS engine.

The obfuscated script rotates its string array by a different amount on every response, so positional parsing breaks within a day. Roles are resolved by content instead: the 40-char hex string is the token, the string ending in = is the cookie prefix.

② The index. Orders are grouped under plain <p> year headings rather than nested markup, so scraper.py walks <p>/<li> nodes in order and carries the current year forward as state.

③ Parsing. PDFs are parsed with pdfplumber. The tricky part is that a bare regex for number/year also matches legal references in the headers and footers — Regulament 2016/679, Legea 21/1991 — which are not case numbers. A noise filter drops those lines before extraction.

④ Search keys. Case numbers appear in two shapes: 16545/2024 and 48031/RD/2023. Each case is stored twice — as printed (case_number) and normalised to number/year (search_key) — so 48031/2023 finds the RD variant too, and a bare 29293 matches every year.

CLI

make search Q=29293/2023

make stats

Or directly, inside the container or a local virtualenv:

python cli.py sync --workers 8 --limit 50

python cli.py search 48031/2023

Make targets

Run make with no arguments for this list at any time.

Docker
make build	Build the image
make up	Start the web UI on :8000
make down	Stop (the volume and its data survive)
make logs	Follow logs
make sh	Shell inside the container
Data
make sync	Full sync — index + all unparsed PDFs
make sync-fast	Trial run on 20 orders
make sync-retry	Sync, ignoring retry backoff
make stats	Orders and cases in the database
make search Q=…	Look up a case number
make db-export	Snapshot the volume's database to ./data.sqlite3
make db-import	Load ./data.sqlite3 into the volume
make clean	Remove containers and the database volume
Without Docker
make venv	Create .venv and install dependencies
make dev	Run Flask locally

Variables: PORT (default 8000), WORKERS (default 4), Q for make search.

Moving the database between machines

A synced database is ~36 MB and is deliberately not in git. To avoid re-syncing from scratch elsewhere, move the file:

make db-export && scp data.sqlite3 other-host:/path/dosar-radar/

Then make db-import on the other side. Both targets go through VACUUM INTO and clean up stale WAL files, because copying a live SQLite database by hand silently loses — or reverts — recent writes.

Configuration

Variable	Default	Meaning
CETATENIE_DB	./data.sqlite3 (/data/data.sqlite3 in Docker)	Database path
PORT	8000	Host port published by compose
BIND	127.0.0.1	Interface compose binds to. Behind a reverse proxy leave it alone; set 0.0.0.0 only if you really want the app exposed directly
SYNC_INTERVAL_HOURS	3	How often the built-in scheduler re-syncs
SYNC_RETRY_MINUTES	5	First retry delay after an error; it doubles on subsequent failures
SYNC_RETRY_MAX_MINUTES	360	Maximum delay between failed attempts
CETATENIE_PROXY_URL	—	Approved HTTP(S) egress proxy for reaching the source
SYNC_ENABLED	1	Set to 0 to disable automatic syncing entirely

Syncing

A background thread in the web process re-syncs every SYNC_INTERVAL_HOURS. Both successful runs and failures are persisted, so restarting the container does not reset the schedule. A failed index refresh retries with exponential backoff (five minutes by default), instead of repeatedly hitting the ministry every minute.

If the ministry's WAF blocks the server IP, the service shows the precise error and pauses retries. It cannot fetch new data until the IP is allowlisted or CETATENIE_PROXY_URL points to an approved egress proxy.

There is no HTTP endpoint to trigger a sync. The service is meant to be public, and an unauthenticated trigger would let anyone use your server to hammer a government site. Manual runs go through the CLI:

make sync

Both paths take an exclusive flock on the data volume, so a manual run and a scheduled one can never overlap — whichever is second exits immediately.

Failed orders back off. Three orders have been dead on the ministry's own site for years; retrying them every three hours costs a 60-second timeout each and fills the logs with alarming errors from a perfectly healthy service. Retries are spaced out instead — 1h, 6h, 24h, 72h, then weekly. To ignore the backoff and retry everything now:

make sync-retry

The container runs gunicorn with one worker and eight threads on purpose: background sync progress is held in process memory, so with several workers /sync/status would answer sometimes from the process doing the downloading and sometimes from one that knows nothing about it.

Project layout

app.py                 Flask routes — index, search, order, sync
cli.py                 Command-line interface
cetatenie/
  session.py           HTTP session + proof-of-work challenge solver
  scraper.py           HTML index parser + PDF case extractor
  db.py                SQLite schema and queries
  sync.py              Orchestration, thread pool, progress state
templates/             Server-rendered UI, no build step, no JS framework

No frontend toolchain, no ORM, no migrations — about 600 lines of Python in total.

Limitations, honestly

• This is not an official source. The authoritative answer is always the PDF on cetatenie.just.ro, which this tool links to for every result. Numbers are extracted by regex from documents that were never meant to be machine-read.
• 3 of 2,214 orders do not parse. The ministry's own index links to three PDFs that return 404. They are marked error in the UI, with the reason shown.
• Scanned PDFs would be missed. Orders without a text layer are flagged rather than silently returning zero cases. None are currently in that state, but the check exists.
• Be a good citizen. Sync is manual and defaults to 4 parallel downloads. Please do not crank WORKERS up and hammer a public ministry's site.

License

MIT. Not affiliated with, endorsed by, or connected to the Ministry of Justice of Romania. All order data belongs to its publisher and is linked, not republished.