Saku Pajak

What it is

Saku Pajak does one thing: search tax regulations. Type a question, and what comes back are the relevant articles, along with whether each one is in force and where that status comes from.

It is a static site. Search runs on the user’s own device — no server, no account, no telemetry, and no network requests while searching. Since it only needs a browser, it can be used on office computers that don’t allow installing any app. On a phone, the same site can be installed and used without internet.

This site is not an official government site, and that is written on every page. Its name does not use the word “official”, the national emblem, or any agency’s logo.

Built-in data: public regulations only

The built-in data is only public regulations from the regulation catalog of the Directorate General of Taxes (DJP) and the legal documentation network (JDIH) of the Ministry of Finance, split by article. Every document always carries its source URL and the date it was fetched, both in the data and on screen.

Coverage in v1.0.0: income tax (PPh), general provisions and procedures (KUP), and VAT (PPN). The other categories in the DJP catalog — stamp duty, land and building acquisition duty, property tax, and “Other” — are being fetched for v1.1, together with an incremental update command that checks for new or changed regulations without fetching every detail again. Those results are not on the site yet.

Rules that must not be broken, whatever the reason:

• Status is never guessed. Each source’s claim is shown as it is, with the date it was fetched. Contradictions are marked uncertain. There is no single status label made up by the app itself.
• Documents without text still appear in lists and search results, with a note that the text is not available at the source and a link to the original.
• Scanned attachments are shown as the original PDF, never as OCR text.
• Links between regulations are hints, not truth, so they always come with a quote of the source sentence.

Personal collection

Many everyday references — circulars, internal memos, confirmation letters, rulings — are not available from public sources that can be fetched. For that there is a personal collection: users can import their own PDFs, text files, or pasted text, then search them with the same engine as the built-in data.

Those documents are stored only on the user’s own device (IndexedDB), and every collection document is labeled “Not verified”. References inside them to public regulations are linked automatically. This was proven with network recordings in a real Chrome: zero requests while searching, opening, exporting, deleting, and restoring the collection, and no document content or id in any URL, request body, tab title, or console.

Documents that prospective users once sent as test material never went into the repo, the built-in data, or any commit. All collection tests use mock documents written from scratch.

Data pipeline trial: the income tax category

Before the specification was written, data fetching was tested first for one category, income tax. Its report became the basis for nearly every decision above.

• More than half the list is not substantive regulation. Of the 2,468 income tax rows in the DJP catalog, 1,430 (58%) are weekly exchange rate or monthly interest rate announcements. Those were left out of scope. After merging with JDIH, 1,122 substantive documents remain.
• Status often disagrees between sources. Of the 337 documents whose status could be compared across two sources, 121 (36%) differ. Adding conflicts with revocation evidence in their own text, 190 documents (17%) have an uncertain status. That is why status is never inferred.
• Even source texts contain OCR errors. 23 of 570 DJP text files contain typical uncorrected OCR patterns — clan for dan, rn for m. Search is made robust to those errors without changing the original text. On scanned attachments, the OCR engine reported 0.93–0.99 confidence for clearly broken text, so OCR confidence scores are never used to decide whether something is usable.
• Many older regulations have no text. In the DJP catalog, regulations from the 2010s and 2020s have full text, but documents without text reach 51% for the 2000s, 90% for the 1990s, and 96% for the 1980s. Their pages only hold metadata. Such documents still appear, with a link to the source.

Offline, and proof that the content is the same

For offline use, document and article pages are not stored one by one; they are built on the device from the data — by the same function used at build time. The on-device storage is about 81 MB, with a first download of about 15 MB.

So that “the same” is not just a claim, the build compares all 20,966 pages produced by the offline path with the published pages, byte for byte. The negative test: one page is changed, and the difference is caught right away. A test in a real Chrome turns the server off and then opens the list, search, filters, documents, articles, and collection while offline.

Known limits

• Real installation has only been tested on an Android phone by the project owner. Safari and iOS have not been tested.
• How fast the first save is on a slow phone has not been measured on a real device.
• Circulars from the Director General of Taxes are practically unavailable from official sources that can be fetched, so they are not in the built-in data. That is why the personal collection exists.