How to Search Scanned PDFs and Images Locally with OCR

July 28, 2026 · 5 min read

A scanned contract, a photographed receipt, an old record saved as an image — to a computer, these are just pictures. There is no text to match, so ordinary search comes up empty even when you know the exact phrase is in there somewhere.

The fix is OCR (optical character recognition): reading the words inside an image so they become searchable. This guide explains how to search scanned PDFs and image files, why doing the OCR step on your own device matters for confidential documents, and how to set it up without uploading anything.

Why scanned PDFs are invisible to normal search

When you scan a document or snap a photo of a receipt, the result is an image of a page. Visually it looks like text, but the file stores pixels, not characters. Your operating system's search, and most PDF viewers, can only match text that already exists in the file.

That is why a search for a clause you know is in a signed agreement returns nothing: there is no underlying text layer to match against. The same goes for:

• Scanned contracts and signed agreements
• Photographed or scanned receipts and invoices
• Old records digitised on a flatbed scanner
• Screenshots and photos of whiteboards or forms

Until the words are extracted, these files are effectively a black box. OCR is what opens that box.

What OCR actually does

OCR analyses the image, recognises the shapes of letters and numbers, and produces a machine-readable text layer. Once that layer exists, the document behaves like any other searchable file.

A good OCR step handles the messy reality of real documents:

• Scanned PDFs where each page is a full-page image
• Image files such as PNG, JPG, and TIFF
• Mixed documents that combine typed text with scanned inserts or signatures

OCR is not perfect — faint scans, unusual fonts, and handwriting can trip it up — but for typed contracts, invoices, and printed records it is usually accurate enough to make previously unsearchable files findable. The important design decision is where that processing happens.

Why cloud OCR is risky for confidential scans

Many free 'convert scanned PDF to text' tools work by uploading your file to a remote server, running OCR there, and sending back the result. For a shopping list that is fine. For confidential material it is a problem.

When you upload a scanned document you lose direct control over it. You are trusting a third party's retention policy, security, and staff. For regulated professionals this can conflict with obligations around client confidentiality and data handling.

Consider what tends to be scanned in the first place:

• Legal contracts and case files
• Medical and personal records
• Tax documents, bank statements, and receipts
• Signed HR and compliance paperwork

These are exactly the files you would least want sitting in someone else's cloud. The safer approach is to keep the entire process — reading the image and building the index — on your own machine. Our guide to searching PDFs without uploading them covers this trade-off in more detail.

How to search scanned PDFs entirely on-device

Local OCR removes the upload step. The recognition runs on your computer's own processor, the extracted text stays on your disk, and nothing leaves the device. You can literally unplug the Wi-Fi and it still works.

With Fossick, the workflow is straightforward:

1. Point it at the folders that hold your documents, including scanned PDFs and images.
2. It runs OCR automatically on image-only PDFs and picture files such as PNG and JPG — no separate conversion step.
3. It builds a private, on-device index of the extracted text.
4. You search, and matching scans appear alongside your normal documents.

All of this happens locally. Indexing, OCR, and search run on your machine; your files are never uploaded. The only thing that touches the internet is licensing. If you want the broader picture of how offline indexing works, see our overview of offline document search.

Beyond exact keywords: finding scans by meaning

OCR makes the text searchable, but you still have to know what to search for. That is harder than it sounds when a scan has a generic filename like scan_0043.pdf and you only half-remember the contents.

Fossick adds semantic search on top of the OCR text. Instead of matching exact keywords, it finds documents by what they mean. You can type the gist of what you are looking for — the lease amendment about the parking spaces — and it surfaces the right scan even if those precise words never appear.

This matters for OCR output specifically, because recognition is imperfect. A slightly misread word can break an exact keyword match, whereas meaning-based search is more forgiving of small errors. If you want the distinction spelled out, read semantic search vs keyword search.

To be clear about what this is and is not: Fossick is search, not chat. It finds and ranks your own documents. It does not summarise them, answer questions about them, or generate text. There is no LLM and no generative AI involved.

Setting realistic expectations for OCR results

OCR quality depends on the source. A few practical notes:

• Clean, typed scans at 300 DPI or higher generally OCR well.
• Faded, skewed, or low-resolution scans will have more recognition errors.
• Handwriting is far less reliable than printed text and should not be assumed searchable.
• Very large libraries take time to index the first time, because every page image has to be processed once. Searches after that are fast.

Because semantic search tolerates imperfect text better than exact matching, minor OCR mistakes rarely stop you from finding a document. But if a scan is genuinely unreadable to the human eye, OCR will struggle too. Rescanning at higher quality is the fix in those cases.

The practical payoff is simple: a folder of scanned contracts and receipts that used to be a dead zone becomes a set of files you can actually find on demand — without any of it leaving your computer.