Making Montenegro's law searchable

Montenegro's Official Gazette publishes every law the country passes, and its register of 78,308 acts is complete and current. What it does not publish is text. Each act arrives as a set of page images, so the country's entire body of law can be looked at but not searched, quoted or handed to an AI.

The obvious workaround had already died. A private publisher, paragraf.me, carried clean consolidated text for 81 core statutes, then left the market, and its collection stopped on 31 March 2021. Building Pitaj, an assistant that answers legal questions and quotes back the articles it relied on, on top of text more than five years stale was not an option, so the law had to be read out of the Gazette's own pictures.

That turned out to be possible because of a detail the Gazette never mentions. The page images are renders of its own digital typesetting rather than scans of paper, clean enough that upscaled three times and run through Tesseract's Croatian model they come back with the diacritics intact and under two percent error on modern issues. The difficult part was the Gazette itself, which answers heavy traffic with HTTP 403, so the pipeline caches every page it has read, backs off instead of skipping, and publishes any act with a missing page flagged as incomplete rather than quietly broken.

Montenegro's 2026 VAT law raises the registration threshold to 30,000 euros, but it only applies from EU accession, so the 18,000 euro threshold in the older law is the one that applies today, and an assistant that trusted the most recent document would hand small businesses the wrong number. Every article in the vault therefore carries flags for whether it is actually in force, whether its text is the 2021 consolidated version or the current Gazette as published, and whether it has been repealed, and repealed articles never reach the assistant at all.

Each article is its own note, filed into one of thirteen categories and cross-linked to the laws it cites, so a person browsing the vault and the assistant answering a question are reading exactly the same words. As of 19 September 2026 about half the register, 38,696 acts, has full text, and the ingestion is still working through the long tail of municipal decisions.