Still scrubbing through lectures to find one slide?

Retrieve makes your lecture slides, PDFs and recording captions searchable on your phone. Type what you remember and get the module, the week, the slide, and the exact minute in the recording. Or ask a question and get a short answer that cites its sources.

  • Free
  • Works offline
  • Nothing leaves your phone

Revision shouldn’t start with a search party.

The answer is somewhere in fourteen PDFs and a two-hour recording. Some slides are images, so Ctrl+F finds nothing. Phone search can’t look inside any of it. So you scrub, skim and guess.

Without Retrieve

One question, four places to look

Blackboard · Week 6
14 files. Which lecture was Raft?
Lecture recording
1:52:04 at 1.5× and still scrubbing
Ctrl+F in Lecture 11.pdf
0 results (the slide is an image)
Group chat · 23:48
“does anyone know where he did log matching”

With Retrieve

One search, and you’re on the slide

raft log matching
Raft: log matching
If two logs contain an entry with the same index and term…
COMP26120 · Week 6 · Lecture 11 · slide 14
32:10In the recording: “…the log matching property…”

From a folder of slides to the right minute.

Add a week of material once. After that, every question takes a few seconds.

Step 1

Import

Pick slides, PDFs and caption files and file them under a module and week.

Step 2

Index

Retrieve reads every page on the phone, including slides that are only pictures of text.

Step 3

Search

Type what you remember. Narrow it with module: or week:4..6 when you know roughly where it was.

Step 4

Jump to it

Open the slide with your words highlighted, or the transcript at the moment it was said.

Built for the way you actually revise.

Your course is already organised by module and week. Retrieve keeps that structure, so every folder is also a search.

Search by meaning

Ask for “agreement between servers” and it finds the slide titled consensus. Keyword and meaning matches are combined.

A query language when you need it

Filter by module and week, search exact phrases, leave things out.

module:COMP26120 week:4..6 "amortised" NOT tutorial

Recordings, to the minute

Import the caption file from the lecture portal. Results open the transcript at the right timestamp.

See the slide itself

Results show the real page, with your search words highlighted. Swipe to the slides either side.

Organised like your course

Modules, then weeks, then documents, each with its own colour. Open a week and you’re already searching inside it.

Stays on your phone

No account and no cloud. Indexing and search run on the device, and it works in airplane mode.

Ask your lectures. Every answer shows where it came from.

Ask a question in plain words. Retrieve finds the passages that cover it and writes a short answer from them on your phone, with a numbered link to every slide and recording moment it used.

  • Answers only from your material

    The model only sees passages from your own slides and captions, not the internet.

  • Every citation is checked

    Tap a number to open that slide, or the transcript at that minute. A citation that doesn’t match a passage the model was given is flagged.

  • It says when it doesn’t know

    If your material doesn’t cover the question, you get “Not in your materials” and an offer to search instead.

8 of 8off-topic test questions refusedrefusal.csv · test
96%of answerable questions answeredrefusal.csv · test, tuned on dev
— of 30answers judged faithfulqa_eval.csv · being judged

Small enough for a phone. Measured, not guessed.

A C++ search engine, a distilled embedding model and a lot of benchmarking. Every figure here comes from a results file in the repo.

86%

of 101 slide questions find the right page in the top five, up from 81% with keyword search alone

recall.csv · hybrid vs BM25 · test:slides
3×

faster embeddings after distillation: 89 ms vs 270 ms per query on a Galaxy Tab A7

embed_latency_device.csv · median, CPU
11 MB

embedding model after int8 quantisation, down from 42 MB, with near-identical output

quantize.csv · cos ≥ 0.999
— ms

search latency at P99 on device, over 200+ queries

to be measured (spec §6)
KotlinJetpack ComposeC++ · NDK/JNIPyTorchLiteRT Read the code on GitHub ↗

Your lectures, searchable in your pocket.

Free on Android. Add your first module in a couple of minutes, then search it anywhere, even without signal.

AndroidAndroid 10 or later
Get it onGoogle Play
QR code that opens Retrieve on Google Play
Scan on your phone

Questions students ask first

Does anything get uploaded?

No. Your files are read, indexed and searched on the phone. Retrieve has no account and no server, and search works in airplane mode. The privacy policy has the details.

Which files can I add?

PDF slides and documents, and WebVTT caption files (.vtt) from your lecture recording portal. Slides that are only images are read with on-device text recognition.

Can it search lecture recordings?

Yes, through their caption files. Each 30 seconds of captions becomes a search result, so you land on the right minute. It doesn’t transcribe audio itself.

What do I type?

Whatever you remember, in your own words. If you know roughly where it was, add filters like module:COMP26120 or week:4..6. The search tips in the app use your own module codes as examples.

Can Ask make things up?

It can get things wrong, like any language model, which is why every sentence links to its source. It only sees passages from your own material, citations that don’t match those passages are flagged, and when nothing relevant is found it says “Not in your materials” instead of guessing.

Which phones does it run on?

Search runs on Android 10 or later, even on a 3 GB tablet. Ask runs a language model (Gemma 3 1B) on the phone, so it needs a more recent phone with more memory. The model (about 580 MB) downloads the first time you open Ask, and after that it works offline too.

Is it free?

Yes. There are no ads, subscriptions or in-app purchases.

Find it in seconds next time.