From Paper to a Searchable Archive Without Uploading a Single Page

Somewhere in your home or office there is a box. Contracts, warranty cards, printed invoices, handwritten meeting notes, a letter from an institution that will ask you about it in eighteen months. Paper is the worst possible storage medium: it is not searchable, it is not backed up, and it degrades. Digitising it properly is a weekend of scanning and about twenty minutes of actual work — provided you never upload any of it.

This is the workflow. Every step runs in the browser, and the intermediate files stay on your disk.

Step 1: capture with the end in mind

Nothing in software recovers a bad scan. Two minutes spent here saves an hour of correction:

• 300 DPI minimum for text. Below roughly 200 DPI the characters lose the detail recognition depends on, and no setting will bring it back.
• Flat and evenly lit. A phone photo is fine if you fill the frame and avoid your own shadow. Skew beyond two or three degrees degrades accuracy quickly.
• One document type per batch. Grouping receipts together and contracts together lets you use the same recognition settings for a whole run instead of per page.

Step 2: recognise the text

Run each page through OCR with the language set to what is actually on the page. Mode matters more than anything else here:

• Single block for letters, contract pages and anything that is one uniform column of prose.
• Sparse text for receipts, forms, invoices and anything where text sits scattered around the page without paragraphs.
• Binarise scans that came out grey or patchy; leave clean colour captures in none or grayscale.

Then audit rather than trust. Raise the confidence filter until only the suspect words remain and check those against the image. On a decent scan that is a handful of words per page — proper nouns, dates, and amounts, which is exactly the set you wanted to verify anyway.

Step 3: name and store as text, not as images

The most important decision in the whole workflow: store the extracted text, not just the scan. Keep the image as the archival original, but make the text the thing you search.

A simple, durable convention beats any tool:

• Filenames that carry meaning: 2024-03-11_utility-invoice_northend-energy.txt. Date first sorts chronologically forever.
• One folder per year, or per category if you prefer — consistency is what matters, not the scheme.
• A short header at the top of each file: source, date, and one line saying what it is. Future you will not remember.

Step 4: make it searchable

With the text on disk, you have two levels of search, and they solve different problems.

Exact search — your operating system’s file search — finds the invoice with a specific reference number. It is instant and it is the right tool when you know the string.

Semantic search is what you want when you do not know what you are looking for. Loading the documents into the on-device document search tool lets you ask “which of these mentions an early termination fee?” and get the right page back even though no document uses that phrase. The embedding model runs locally, so the contents of your archive are still never uploaded.

Step 5: keep the originals

Do not throw the paper away immediately, and do not throw the scans away ever. The workflow is: scans are the masters, extracted text is the index, and semantic search is how you reach it. That gives you three independent ways to find anything, none of which requires trusting a third party with the contents.

What this replaces

The cloud version of this workflow exists and it works — it just requires you to upload contracts, medical paperwork and financial records to a service that will hold them under terms you have not read. For documents with your name, your address, your account numbers and your signature on them, that is a strange trade to make for convenience.

Running it locally costs nothing, works offline once the models are cached, and leaves you with files you control. Start with one folder — the utility bills, or the warranty cards — and see how long it actually takes. It is usually shorter than the time you have already spent looking for one of them.