Skip to main content
Linkly AI can pull up a document on your machine in a fraction of a second because of one thing it quietly does in the background: indexing. This page walks through every switch under Settings → Index — what it controls, what it defaults to — and gives you a path to follow when indexing misbehaves. If you’ve just installed the app and haven’t added your own documents yet, start with the Quick Start.

​
How Indexing Works

Indexing in Linkly AI is layered — it isn’t a case of nothing working until everything finishes. Each layer that completes adds another way to search.
1

Layer 1 — filenames (seconds)

Paths and filenames are written to the index the moment a file is discovered. This layer is close to instantaneous, which is why you can search by filename right after adding a folder.
2

Layer 2 — full text (keywords)

Document bodies are parsed into a local full-text index that supports keyword matching. Once this layer is done, searching for “budget sheet” matches documents whose body contains those words.
3

Layer 3 — semantic (vectors)

A local embedding model splits the text into chunks and turns them into vectors. Once this layer is done, searching for “how much did we spend last year” can surface a finance report whose title contains none of those words.
On top of those three layers sit two optional content extractors, whose job is to turn files that contain no text into files that do, and then feed the result into the three layers above:
  • Text recognition (OCR): reads text out of images and scanned PDFs
  • Audio and video text extraction (ASR): turns spoken content in audio and video into text

​
A Tour of Settings → Index

Open Settings → Index and you’ll find four blocks, top to bottom:

​
Reading the Status Detail

​
The Index Settings

​
1. Performance Mode

What it does: controls how much compute background indexing is allowed to take. Default: Auto (Recommended)

​
2. Power Saving

What it does: pauses the power-hungry heavy work while a laptop is running on battery. Default: off
It only takes effect when the toggle is on and the machine is actually running on battery — both at once.
Paused while on battery:
  • Image OCR
  • Converting scanned PDFs to structured Markdown
  • Semantic indexing (vector generation)
  • Audio and video transcription
  • Cloud library auto sync
  • Database housekeeping
Unaffected, still running: filename indexing and full-text (keyword) indexing. So even in Power Saving, new files remain searchable. When to touch it: laptop users should just turn it on and forget about it. On a desktop it makes no difference either way.

​
3. Sleep AI Models When Idle

What it does: releases the memory held by local AI models after a period of inactivity, and reloads them automatically when they’re needed again. Default: on Local models (semantic embedding, OCR, speech recognition) are unloaded after roughly 300 seconds (5 minutes) of inactivity to free memory. With it off, the models stay resident — around 1 GB of memory for all three combined. When to touch it: if memory is tight (8 GB or less), leave it on.

​
4. Image Text Recognition (OCR)

What it does: uses a local OCR model to read text out of images, and to convert scanned PDFs into structured Markdown. Default: on What it does: when on, lets you find documents using natural, descriptive language. Default: on

​
6 / 7. Audio Transcription and Video Transcription

What they do: use a local speech recognition model to turn spoken content in audio and video into text, then feed that text into the full-text and semantic indexes. Default: both off ⚠️
Off doesn’t mean unindexed. Media files are still registered in the index — you can find them by filename, but not by what’s said inside them.
When to touch it: turn them on when you have meeting recordings, podcasts, or course videos you want to search.

​
Optimizing and Rebuilding the Index

​
Index Storage Optimization

Rebuilds the vector index in a more compact form. This is a near-free win:
  • Disk usage drops by roughly 80% (measured shrink of up to 11.6×)
  • Search gets faster
Quitting the app partway through is safe — nothing is corrupted, and it resumes on the next launch.
If the index is already in the optimal state, the button does nothing — it just reports Your index is already optimized, and there is no way to force a re-run.

​
Rebuild Index

This is a destructive operation, which is why the button is red. It will:
  • Delete: the full-text index directory, all vectors, and the per-document indexing progress markers
  • Keep: your original files, the text already extracted into cache, and the model files already downloaded
Precisely because the text cache and the models survive, a rebuild is usually much faster than the first index — it skips the slowest stages, parsing and OCR. While a rebuild is running, searches return empty results (they don’t error).
In these four situations the button is refused and nothing happens:
  • OCR or media transcription jobs are still running in the background
  • Indexing is paused (from the tray menu)
  • A scan or an index pass is currently in progress
  • A rebuild is already running

​
Index Status

Status has 5 possible values: Watching is the end state you want to see.

​
How Do I Pause Indexing?

Pause Indexing lives only in the system tray menu — right-click the L icon in the tray to open the menu, where you’ll find ⏸ Pause Indexing / ▶ Start Indexing (left-clicking opens the AI chat window, not the menu).

​
FAQ

Because audio transcription and video transcription are off by default. While they’re off, media files are still registered in the index (hence the filename match), but nothing spoken inside them has been transcribed.Go to Settings → Index and turn on Audio transcription and Video transcription. The app downloads a 161 MB speech recognition model, then works through your existing files automatically — no rescan and no rebuild required.One exception: if a media file sits next to a matching .srt / .vtt subtitle file, or was transcribed earlier and the cache is still around, its content is searchable even with the toggles off.
First check whether Image Text Recognition (OCR) is switched off under Settings → Index.Note that this one toggle governs both image OCR and text extraction from scanned PDFs — even though the label only mentions images. Turn it off and scanned PDFs become filename-only, with no body text.If the toggle is on, look at the Text Recognition (OCR) row in the status detail: Preparing means the model is still downloading or loading, so give it a moment.
The most common cause is that indexing has been paused. That switch exists only in the system tray menu — right-click the tray icon to open the menu, and if you see ▶ Start Indexing, it’s currently paused; click it to resume.How to tell: the settings page shows Idle with a small Indexing paused line below it.If it isn’t paused, double-check under Settings → Folders that you’ve actually added your own directories — the onboarding flow only indexes the bundled sample documents.
Work through these in order:
  1. Does the Semantic Search row in the status detail under Settings → Index read Disabled? If so, the toggle is off — turn it on (it backfills automatically, no rebuild needed).
  2. Does it read Preparing or Queued? Vectors are still being built; wait for it to finish.
  3. If it’s neither, go to Settings → AI → Embedding model and check whether the model is still downloading or failed to download — the search interface gives you no indication of this at all, and search silently degrades to keyword-only. If it failed, click Retry on the card.
Pause Indexing is only in the system tray menu; the settings page doesn’t offer it. Right-click the L icon in the tray → ⏸ Pause Indexing. Resuming happens in the same place, where the item reads ▶ Start Indexing.Rebuild Index can’t run while indexing is paused — resume first.
Go to Settings → Index → Maintenance and click Index storage optimization. It rebuilds the vector index in a more compact form, cutting disk usage by roughly 80% at the cost of about 1% of recall — and search gets faster in the process.If the button reports Your index is already optimized, your index is already in the new format, there’s nothing left to reclaim, and a re-run can’t be forced.
Not yet. A rebuild is the last resort, and searches return empty results while it runs. Try this order instead:
  1. Confirm the relevant toggles are on (OCR / Semantic Search / audio and video transcription)
  2. Confirm indexing isn’t paused (tray menu)
  3. Confirm the status has actually reached Watching rather than still sitting at Indexing
Only once all three are ruled out is it worth going to Settings → Index → Maintenance → Rebuild. A rebuild doesn’t touch your original files, the already-extracted text cache, or the model files, so it’s usually much faster than the first index.
No. Performance Mode only controls how many threads indexing uses and how readily it yields — not whether it runs at all. Choosing Balanced just makes it slower and less intrusive.To actually stop it, use Pause Indexing in the tray menu.Also worth knowing: Auto looks at keyboard and mouse idleness only (over 120 seconds counts as idle) — it ignores battery state and CPU load. On a pure Wayland Linux desktop that signal isn’t available, so Auto behaves identically to Balanced.
That’s expected. Power Saving takes effect only when the toggle is on and the machine is running on battery. Plugged in — or on a desktop with no battery at all — it does nothing.Which is exactly why you can leave it on year-round, with no need to flip it on before you head out and off when you get home.
From the gentlest measure to the strongest:
  1. Set Performance Mode to Balanced (takes effect immediately; jobs already running pick up the new budget)
  2. Temporarily turn off Audio transcription and Video transcription — media transcription is the heaviest work there is, and it competes with OCR for the same single global slot
  3. Still sluggish? Use Pause Indexing in the tray menu and resume when the machine is free

​
Further reading