AI & Documents · Updated October 6, 2026
How does semantic search work in PDFDepth?
Semantic search is a toggle in the left sidebar’s Find panel. The first search builds a one-time index — page text and text annotations are chunked and embedded with a Gemini embedding model, cached locally in IndexedDB — after which queries return meaning-ranked hits with a page number and snippet that jump to the exact line.
Keyword search vs meaning search
Classic Find is literal: "termination" finds "termination" and misses "either party may end this agreement". Semantic search ranks by meaning, so the clause surfaces anyway — the index embeds the document’s text as vectors, embeds your query the same way, and scores each chunk by similarity blended with keyword overlap. You still get a page number and a text snippet per hit; the difference is what reaches the list.
The one-time index
Enabling Semantic and searching builds the index once per document: the extractor pulls each page’s text plus any text annotations, chunks it, and embeds the chunks in batches — the status line counts "Indexing n/N" as it goes, and unchecking the toggle cancels the build. The finished index lives in your browser’s IndexedDB keyed by the file’s hash, so reopening the same file later skips the whole build. It uses your own Gemini API key, set in Settings.
Reading and jumping to results
Hits come back ranked — up to eight, each with a snippet and an "AI" badge — and the status line reports the count. Clicking a result jumps to its page, zooms in, and highlights the matched line the same way Ctrl+F does; the top hit auto-jumps so the best answer is already on screen.
Scanned pages and honest limits
A page with no text layer contributes nothing to the index — the build records those pages as empty — so scanned documents need the OCR toggle first (it runs OCR so Find, and therefore the index, has text to work with). The index is keyed to the file’s exact bytes: edit the document and the next semantic search builds a fresh index. And embeddings mean the page text goes to the embedding API, same as the other AI tools.
Doing it in PDFDepth
Left panel → Find → check "Semantic". Type the way you would ask a colleague — "the retention clause", "where they talk about drainage" — and click a hit. For follow-up questions rather than lookups, the AI chat and summary tools read the same document.
References
Everything in this article runs in the live PDFDepth workspace — the same app the workflows above were verified against.
Key takeaways
- Semantic = ranked by meaning; the same Find box does literal search with the toggle off.
- The first search builds a per-document index (chunks → Gemini embeddings → IndexedDB); later sessions reuse it.
- Hits carry page, snippet, and an AI badge; clicking zooms to the matching line.
- Scanned pages are invisible until OCR runs — flip the OCR toggle on scanned documents first.
Related questions
No — embeddings come from the Gemini API, which needs your API key and a connection. The finished index, once built, lives locally in IndexedDB.
Do it in PDFDepth
In the app: Left panel → Find → Semantic
Try it free — no signup, no upload
Related articles
Put it into practice
Everything above runs inside the PDFDepth workspace — free for 14 days, no credit card.
Start for Free