User guide

Contents

• Explaining
• Preview a page before reading
• Follow-up questions
• Reunions
• Concepts and spellings
• Review
• Weekly digest
• Graph
• History
• Export, backup and restore
• Site rules
• GitHub, Notion and Google Docs
• Pages that look private
• Keyboard and shortcuts

Explaining

Select up to 200 characters, then use any of these:

The card shows the answer as it is written. While it is being written, Stop ends it. If the request fails, the card offers Try again and, when you have several models configured, Try with… for each of the others.

Each answer carries a tag:

• Defined in source – the page defines the term, and the quote the model gave was found in the page text.
• External knowledge – the model explained it from what it knows.

Explanations are written in the language you choose in settings (16 languages, or "follow the interface"). Technical terms keep their original form. Formulas are typeset, and tables and nested lists are shown.

If the term is close to one you already looked up, the model is told so and can explain the difference instead of starting over. If it is very close, Harkback asks "Is this the term you looked up 3 days ago?" and records your answer.

Preview a page before reading

Click the toolbar button, press Alt+Shift+P, or choose Harkback: Preview this page in the right-click menu, and a panel offers to scan the page. It names the model that would read the page and says whether it is local or remote. Nothing is sent until you press Scan. The page text then goes to your model once, which lists the page's key terms. Your own records sort them, on your computer:

Preview beside a term shows the explanation you were given earlier, with no model call. For a new term the model writes a short explanation. Nothing from a preview is recorded; to keep a term, select it on the page and explain it as usual. The model sees only the page text, never your concept list, and a sensitive page only uses a local model. A scan reads the first 16,000 characters of a page; for a longer page the panel says how much was left out, so terms further down are not listed. In a private window every term counts as new, because your records are not read there.

Follow-up questions

Ask more opens a question box under the explanation. Each question stays above its answer, and the whole conversation is saved with the explanation. While an answer is written, the Send button becomes Stop. A follow-up answered before the explanation has been recorded (for example while the "same term?" question is open) is saved as soon as the explanation is.

Reunions

When a page contains a term you have looked up before, Harkback underlines the first occurrence. Hover to open the card.

A reunion is shown only if:

• it is not the page where you looked the term up;
• at least reunion gap days have passed since you last touched it (default 3);
• the term is not muted;
• if the term is an ambiguous abbreviation (SAM, RL, MoE, GP, or one that two concepts share), the page also contains another term from the same field;
• the page has not already used up its limit (default 3 per page). Terms you were confused about come first.

A page that only mentions a variant of something you understood (for example QLoRA, when you understood LoRA) reminds you of the concept it varies.

Card buttons: I remember, Explain again (from another angle), Compare usages (the model contrasts the two contexts), Don't show again.

Pages that load more text, or a single-page app that moves to another page without reloading, are rescanned on their own.

Concepts and spellings

LLM, LLMs, the LLM, large language model and Large-Language Models are one concept; so are β-VAE and beta-VAE, fine-tuning and fine-tuning (with a ligature), and résumé and resume. Rules:

• Accents on Latin letters are ignored; there is no stemming except for English plurals.
• An abbreviation matches its full name when the name has at least three words. Two different full names for the same abbreviation in one field stay separate concepts.
• Chinese names, and Japanese names written only in kanji, of three or more characters are underlined wherever they appear. A two-character name is underlined only where it is a word of its own on the page; a one-character name never is. Names with kana or Hangul need two characters.

A term you select in another language joins the concept you already have when the model's card names it, so "注意力机制" and "attention mechanism" end up as one concept. When the model knows the term's name in your explanation language, it appears under Suggested names on the concept page. It is only recognised on pages once you Accept it, and a name you Reject is not suggested again. On a concept page you can also add an alias, merge two concepts that are the same (this cannot be undone), remove a wrong relation, or mute the concept. Adding an alias that already belongs to another concept is refused; merge instead.

Review

Review (n) on the History page opens the queue of concepts that are due. The toolbar icon shows the same number.

For each concept you see its name first. Try to remember it; you can type what you remember in the box. Then press Show explanation and grade yourself:

Each button shows how long until the term returns if you pick it. The wait comes from FSRS-7, a model of memory: every term has two stabilities (how long it takes you to forget it, for a fast and a slow trace of the memory) and a difficulty, and each answer updates all three. A term you remember easily is spaced out quickly; one you forget comes back soon. A wait is never longer than a year. A term you looked up but never graded returns after 1 day; Understood or Confused on the explain card, pressed right after a look-up, does not count as a review. Marks made in an earlier version's review, a day or more after the look-up, count as Good (Got it, Remembered, a recalled reunion) and Again (Still confused). Muted concepts are never due.

Order. Confused terms come first, then shaky ones (last answered Hard), then the most overdue. A term that builds on another (its prerequisite, see the graph) comes after that prerequisite when both are due, and the card says which due terms build on it. Rejected relations are ignored.

Where to start. When a term is confused or shaky, its concept page and the review card (after you show the explanation) suggest the prerequisites to learn first (three are shown, with a count of the rest): ones you are confused or shaky about, or have never had explained, followed down through weak ones to the weakest foundation. A prerequisite you understand ends the search there. Muted terms and rejected relations are ignored.

Reunions count. Pressing Remembered on a reunion hint on another page counts as a Good answer, so the next review of that term moves out.

Target recall in settings (Review) is how likely you want to be to remember a term when it comes due; the default is 90%. A higher number means more reviews.

Keyboard. Space shows the explanation, 1–4 grade (Again to Easy), S skips. These keys are ignored while you are typing in the answer box.

Check my answer

If you typed an answer, Check my answer asks your model whether it matches the stored explanation. It shows a short verdict and highlights a suggested grade; it never grades for you, and it never suggests Easy. The button names the model and says whether it is on this computer or a remote service. It sends the term, your answer and the stored explanation, only when you press it, and it follows the same rules as an explanation: a concept from a sensitive source only goes to your local model. Each check counts against your rate limit. Nothing from the check is saved; only the grade you pick afterwards.

To turn it off, clear Offer "Check my answer" in review under Review in settings. The button then never appears and nothing is sent.

Weekly digest

Digest on the History page shows one week, Monday to Sunday: how many terms you met (new and revisited, compared with the week before), how many answers were remembered, hard or forgotten, the days you were active, the terms you are still confused about, where you read most, and the connections between concepts that appeared that week. Use Previous week and Next week to move. It is built on your computer from your records and calls no model.

Graph

Graph on the History page draws your concepts as a map. Fill colour shows how well you understood a concept (green understood, pale yellow shaky, yellow confused, blue new); a dashed ring marks a concept that something points to but you have not explained yet. Blue arrows run from a concept to its prerequisites, purple arrows from a variant to what it varies, and dashed lines join related concepts.

Drag to move, scroll to zoom, click a concept to open it. The search box filters the graph to matching concepts and their neighbours. Only the 150 best-connected concepts are drawn, and weak or rejected relations are left out.

History

Everything is grouped by concept, newest first, and updates live while other tabs record. Search covers names, quotes, explanations and follow-up conversations. Select lets you delete several entries at once. Each entry names its source; when the source has a web address (its page, or its arXiv or DOI page) the title is a link that opens it in a new tab, and the same link appears on the review card. A source you marked sensitive shows a badge with Mark as normal (two clicks, because it allows its content to go to remote models again).

Deleting an entry removes it from the app and clears its text from the stored log. Exported backups cannot be recalled.

Export, backup and restore

Backups and exports never contain settings or API keys. In settings, Leave sensitive sources out of backups and exports removes everything that came from sensitive sources, and concepts known only from them.

The file format is documented in Data format.

Site rules

Under Sites in settings, each rule has a pattern and three switches, plus an optional model:

• Pattern – a domain (example.com, which covers subdomains) or a URL prefix (https://example.com/docs, which matches that path and below).
• Auto-scan, Sensitive, Disabled – each has three states. Click a box to cycle: dash (not set: follow a broader rule, or off when no rule sets it), ticked (on), empty (off for sure, even if a broader rule turns it on).
• Model – use this model on matching pages instead of the default.

When several rules match a page, each setting comes from the most specific rule that sets it: a URL prefix beats a domain, and a longer pattern beats a shorter one. So you can mark a whole site sensitive and switch that off for one public path, or the other way round.

For example, with these two rules:

every page on example.com is sensitive except those under /public. If the second rule's box is left at a dash instead, it does not say anything about sensitivity, so /public follows the site and is sensitive too. A dash with no broader rule behind it means off, the same result as an empty box; the difference only shows when another rule also matches. A rule that says a site is not sensitive settles it for the pages of that site, except a page that says it is private (see Pages that look private): that one still asks you once.

Suggested research sites. The Suggested research sites button under the rules adds auto-scan rules for bioRxiv, medRxiv, PubMed, SSRN, OpenReview and ACL Anthology. It skips sites you already have a rule for and leaves those rules unchanged. Press Save changes and the browser asks for site access once for all of them. PubMed pages hold abstracts, not full text, so there is little to scan there beyond the abstract. If you deny the access prompt, the rules are still saved but automatic scanning does not run.

Regional research sites. The Regional research sites button works the same way for CNKI, J-STAGE, CiNii, KCI, SciELO, Redalyc and HAL. It was tried on articles at J-STAGE, SciELO, KCI, Redalyc and CiNii, where terms can be selected and explained and the lookup is recorded; a page with a DOI is one source by that DOI, and the others by their address. CNKI articles are mostly behind a login and HAL asks the browser to pass a check first, so those pages work only once you are through them.

YouTube captions

Add YouTube under Sites with the + YouTube captions button, press Save changes and allow site access when the browser asks. Nothing runs on YouTube before that. Then open a video and turn on captions (CC).

• Hints. A term you have looked up before is underlined in the caption line, and hovering the underline shows what you understood and where. The captions move on after a few seconds, so a hint is only there while its line is.
• Explain. Pause the video (the caption stays), select the term in the caption with the mouse (double-click a word, or drag across several) and press Explain or Alt+Shift+E. The explanation uses the captions you have already watched as its context, so a definition the speaker gave earlier can be quoted ("defined in source"). Text from before you opened the page, or after the point you have reached, is not available.
• Selecting. YouTube normally makes captions impossible to select, and a drag moves them. While the video is paused Harkback lets you select them, and a click or double click on the captions no longer resumes the video or switches to full screen. While it plays, nothing changes.
• Records. The lookup is a normal record with the playback position. In History, Review and the exports the position shows as 12:34, and the source link opens the video at that moment. A video is one source however you reached it (watch, youtu.be, a playlist or a start-time link).
• Privacy. Only the caption line on screen is read; no caption file is downloaded. The underlines are drawn by Harkback, not in YouTube's page. A Sensitive rule for youtube.com makes explanations use a local model, as on any site. The Preview button is not offered on video pages.
• Limits. Auto-generated captions misspell technical terms and a misspelt term will not match. Only youtube.com/watch is supported; in full-screen mode the player hides Harkback's hints (use theater mode).

GitHub, Notion and Google Docs

Add the site under Sites and press Save changes, as for any site; allow site access when the browser asks. Nothing runs there before that. These pages are then read by their own layout, so only the document's text is scanned, not the menus around it, and a document is one source however you reached it.

• Not covered. Code files and other GitHub pages, Notion database views, and the Google Docs editor. The editor draws its text instead of writing it, so there is nothing to read: the toolbar button (or Alt+Shift+E) says so and offers Open readable view, which opens the mobile view (/mobilebasic) of the same document in a new tab, in the same Google account. Select text there. Both views are read-only, so to write you go back to the editor tab. The published (/pub) and mobile views also work if you open them yourself. The Docs preview view has not been checked.
• Long Notion pages. Notion only draws the blocks near the screen, so a very long page may be read in part.
• When a site changes. Harkback finds the text by the site's current layout. If it no longer recognises the layout, the page is read as any other page would be until Harkback is updated.

Pages that look private

Many pages on these sites are private. Harkback cannot see who may open a page, so it goes by what the page suggests, and asks when it is not sure. Before anything is sent to a model:

• A private GitHub repository (the page says so) and a Notion or Google Docs page that is not published are held back. On Notion and Google Docs, a published page (*.notion.site, a Google document opened through its published link) is treated as public and is not held back.
• You are asked once, in the card when you first explain something there, or in the small lock at the corner of the page. Local only marks the source sensitive, so explanations use your local model and fail if you have none. Send anyway goes to your usual model. Tick Remember for this site to make the answer a rule for the whole site.
• The answer is yours and is kept. It is recorded with the source, so you are not asked again about it (nor about the issues of a repository you chose for). In a private window nothing is recorded, and the answer holds until you close or reload the page.
• The lock. A page that is sensitive, or would be asked about, shows a small lock. On a sensitive page, press it to mark the page normal again; on a page that would be asked about, press it to answer before you select anything. If a site rule is what makes the page sensitive, the lock says so and cannot change it: edit the rule under Sites. A page that is fine shows nothing.
• The page preview (the toolbar button) shows the same question instead of offering to scan, and scans only after you have answered. Underlining terms you know needs no model and never asks.
• A repository's issues are as sensitive as the repository. If you mark an issue itself as fine, that overrides it for that issue, unless you mark the repository sensitive again afterwards.

What a page suggests about itself is only a hint. If a page must never leave your computer, mark its site or path sensitive under Sites; that is always followed.

Keyboard and shortcuts