Semantic search vs keyword search for your documents

July 24, 2026 · 5 min read

You remember the document. You remember roughly what it said. But you cannot remember the filename, and the exact words you're typing into the search box aren't the words the author used. So nothing comes up.

This is the everyday gap between semantic search vs keyword search. Keyword search matches the letters you type. Semantic search matches the meaning of what you type. For professionals sitting on years of contracts, reports, and scanned PDFs, that difference decides whether you find the file in ten seconds or give up and recreate it. This article explains how each approach works, where keyword search breaks down, and what to look for when choosing a tool.

What keyword search actually does

Keyword search is the model most of us grew up with. It looks for the literal characters you type and returns files that contain that exact string, often weighted by how often the term appears or whether it's in the filename.

It is fast, predictable, and excellent for one thing: finding an exact value you already know. A case number, an invoice ID, a specific surname, a code snippet. If you type it correctly and the document contains it as extractable text, keyword search will find it.

The catch is in those conditions. Keyword search only works when:

• You remember the precise words used in the document.
• The document stores those words as real, searchable text.
• Your spelling and the author's spelling match.

When any of those breaks, the file effectively disappears.

What semantic search does differently

Semantic search doesn't hunt for characters. It represents the meaning of text as a set of numbers — an embedding — and does the same for your query. It then ranks documents by how close their meaning sits to what you asked for.

The practical result: you can describe a document in your own words and still find it, even if none of your words appear in the file.

A quick example

Suppose you're looking for a report about staff leaving the company. A keyword search for "staff leaving" might miss a document that only ever says "employee attrition", "voluntary turnover", or "resignations in Q3". Semantic search understands those phrases point at the same idea and surfaces the file.

That's the core of semantic search vs keyword search: keyword asks "which files contain these letters?" while semantic asks "which files are about this?"

Where keyword and filename search fail

Most of the daily friction professionals feel comes from a handful of predictable failure points.

• Synonyms and jargon. You search "NDA"; the file says "confidentiality agreement." Nothing matches.
• Paraphrasing. You remember the conclusion, not the sentence. Keyword search needs the sentence.
• Filenames tell you nothing. scan_0142.pdf, Final_v3_REALfinal.docx, and Document(2).pdf are invisible to filename search because the name carries no meaning.
• Scanned documents. A photographed contract or a scanned PDF is just an image. There is no text to match unless OCR has been run first.
• Spelling and language variation. "Organisation" vs "organization," a mistyped client name, an acronym expanded differently each time.

Each of these is a case where the document exists, you know it exists, and keyword search still returns nothing. That's not a small annoyance when the file is a signed agreement you need in a meeting.

A side-by-side comparison

It helps to see the two approaches on the same tasks.

Finding an exact identifier

• Keyword: Strong. Type the invoice number, get the invoice.
• Semantic: Also works, but exact-match is genuinely a keyword strength.

Finding a document by what it discussed

• Keyword: Weak. Requires you to guess the author's exact wording.
• Semantic: Strong. Describe the topic and it ranks the closest matches.

Searching across inconsistent vocabulary

• Keyword: Weak. Every synonym is a separate search.
• Semantic: Strong. Related terms cluster together by meaning.

Scanned or photographed files

• Keyword: Fails without OCR.
• Semantic: Also needs the text extracted first — so OCR still matters, but once the text exists, meaning-based ranking makes it findable even from a vague memory.

The honest takeaway is that these are complementary. Exact-string lookups still belong to keyword search. But for the far more common problem — "I know I wrote something about this, where is it?" — semantic search is the better tool.

Why on-device semantic search matters for confidential files

Semantic search has a reputation for living in the cloud, because generating embeddings used to mean sending your text to a server. For anyone handling privileged or confidential material, that's a problem: uploading client files to a third party can breach obligations you've signed up to.

That trade-off is no longer necessary. Fossick runs the entire pipeline — indexing, embedding, and search — locally on your Windows or Mac machine. Your documents are never uploaded. You can pull the network cable out and it still works, which is the simplest proof that nothing is leaving the device.

It also handles the scanned-document problem directly, running OCR on scanned PDFs and images automatically, on-device, so a photographed contract becomes searchable without a cloud service touching it. If offline handling is your priority, see our guide to offline document search and searching PDFs without uploading.

One clarification worth making: Fossick is search, not chat. There's no chatbot, no generative AI, and it does not summarise or answer questions. It finds and ranks your files by meaning, then you open them and read the source yourself.

How to choose a search tool

When you're evaluating a document search tool, weigh it against how you actually work.

1. Does it understand meaning, or just match strings? If you frequently forget exact wording, semantic ranking will save the most time.
2. Where does processing happen? For confidential files, on-device processing isn't a nice-to-have. Confirm nothing is uploaded.
3. Can it read scanned documents? Automatic OCR determines whether half your archive is searchable at all.
4. What formats does it cover? PDFs, Word documents, plain text, and images should all be in scope.
5. Can you verify the privacy claim? "Works offline" is testable — disconnect and search.

Fossick is built around those answers and is free to download during the beta. If pricing is part of your decision, it's free during the beta with full pricing announced before launch; details will appear on the pricing page.