abhyasa

All releases

abhyasa (अभ्यास) is Sanskrit for “steady practice”. It is a tutor for people who want to learn well and keep their minds active. It asks you questions, and it never gives the answer first. You answer by voice or in text, and you can think aloud. The tutor grades your reasoning, not only the result.

abhyasa keeps a learner model for each concept: how deep you know it, how well you recall it now, and when you must review it. It also learns how you learn. If code helps you more than equations, it uses code to teach a weak concept again. Sometimes it uses a different way on purpose, so that you get better at that way too.

The app needs macOS 26 on Apple Silicon. The tutor uses a language model. You choose it from a list in Settings, for example Claude, GPT or a local model. Your answers go to that model. Speech to text runs on your Mac by default. All other data stays in your vault on your Mac.

Download

Click the Download for macOS button at the top. Then do these steps:

1. Unzip Abhyasa.zip.
2. Move Abhyasa.app to Applications.
3. Open Abhyasa.
4. Wait for the first start. The app uses uv to get Python and the abhyasa package one time. This step takes some minutes.

The app needs uv. If you do not have uv, install it with brew install uv. Each release has a Developer ID signature. Apple notarized each release.

Homebrew

brew install --cask vedicreader/tap/abhyasa

The cask also installs uv. To get a new version, run brew upgrade --cask abhyasa. To remove the app, its settings and its Python environment, run brew uninstall --zap --cask abhyasa. The zap does not remove your vault, because other tools use it too.

How it works

Each question goes through the same six steps:

1. Pick. The code chooses the next concept and the depth. A due review comes first. Then a new concept comes, if you know its prerequisites. The app shows the reason.
2. Ask. A model writes one question at that depth, in one way to learn, such as code or an analogy. It also writes a correct answer and a rubric. You do not see them.
3. Struggle. You answer. If your answer is wrong, the tutor shows your first error. It gives a small hint, but not the fix. You can ask for hints, one at a time.
4. Grade. The model grades your answer and your think-aloud. The code grades choice questions and code exercises.
5. Update. The result and your confidence change the floor, the ceiling and the next review date of the concept.
6. Teach. After a miss, the tutor teaches the smallest step that fixes your first error. Then it asks a check question.

When you say “I don’t know X”, X becomes a concept, and it comes first the next time. The Strengths page (⌥G) shows your strong concepts, your weak spots with their reasons, and the ways that you learn best. “Practice my weak spots” starts a session on your weakest concepts.

Learn a paper, a topic or a link as a walk

A walk takes you through one source, one idea at a time. You can make a walk from a paper in your vault, a text that you paste, a topic, or a link. Use Add, and choose what you have.

The tutor reads the source and picks the main ideas. A walk has 12 ideas by default. You can change this number in Settings, under Walks. Before you start, the plan page shows each idea and the kind of check for it. You can change the kind for an idea there.

Each idea is one step. A step shows the passage from the source, word for word. It also shows a short note from the tutor and a figure from the source, if the source has one. Each step has one check:

• Short answer. You write the idea in your own words, and the tutor grades it.
• Demo. You predict what a demo will do, lock in your prediction, and then change the inputs.
• Order. You put the parts of the idea in the right order.
• Code. You write a small Python function, and its tests run.

The tutor picks the kind that fits the idea and your profile. The step shows the reason. Reading never waits for a check. You can read the passage and press “Skip this idea” at any time. A skipped idea stays unseen in your record. A pass or a miss on a check changes your record, as an answer in a session does.

On a walk step, ⌥S sends your answer, or runs the tests. When the check is closed, ⌥S goes to the next idea. ⌥W asks for a hint. In a demo, ⌥V keeps the state of the demo. ⌥F still shows all the keys.

Code runs in a closed process

The tests of a code check run in a closed process with no network. Each test has 2 seconds, and the whole run has 10 seconds. You can change these limits in Settings, under Code runner. On macOS, a sandbox-exec profile stops the network and stops writes outside a temporary folder. On other systems, the guard is best-effort, so run abhyasa in your own container if you need more. abhyasa installs no libraries. Its code checks use only the Python standard library.

A walk from a link

On Add, choose “A link”. These links work: a web page, a documentation site, a PDF, an arXiv link or id, and a GitHub repository or file. A video link does not work. A link to your own computer or to a local network does not work either.

abhyasa fetches the link only when you press “Make a walk”. It reads only the pages in the folder of the link, after any redirect. It reads at most the page cap. The default is 8 pages. You can set the number on Add, up to the cap in Settings, under Links. The cap is 20 by default.

Rabbit holes

A passage can use an idea that you do not know yet. The step then shows a “This needs” line with each idea that it needs. An idea that is not learned yet has a “Learn beside” button.

You have three choices:

1. A quick primer. A short text that you read in the side pane.
2. A short side walk. A small walk on that idea, in the side pane.
3. Its own session. A normal session on that idea, in a new tab.

Press ⌥B to open the offer. Then press 1, 2 or 3. A primer or a side walk opens in a pane beside the step. The paper stays on the left. On a narrow window, the side walk opens in the same tab, with “Back to the paper” at the top.

• ⌥B opens the offer for the first idea that you do not know.
• ⌥Z goes back to the paper and closes the pane. Esc also closes it.
• A side walk can open another side walk. The limit is 2 levels, and you can change it in Settings, under Walks. Past the limit, only “Own session” is left.
• The check that was open on the paper stays open. The answer that you typed there stays in the box. When the side walk ends, the paper tells you what you learned on the way.

The pane opens only pages of abhyasa. A pane address that points to another site is ignored.

Open in Leela

abhyasa does not install libraries for you, and it runs only short exercises that use the standard library. A library or a command-line tool is better to learn in Leela. Leela has the tool environments, code lookup and an agent tutor.

When you add a link, tick “It is a library or a command-line tool”. This makes a tool walk. A tool walk asks short-answer and order checks here, and it has no code check.

A step that needs code has an “Open in Leela” link. The link opens Leela in a new tab, on the same walk and step. abhyasa never calls Leela. It reads the file that a running Leela writes, and it makes a link only to an address on this computer.

If Leela does not run, the step says so, and it links to the install page.

Settings

Settings has every value that you may want to change. The values are in 11 groups:

Each number has a range. The page shows the range and the default next to the value, for example 3 to 30 · default 12. A value outside its range is not kept. If your settings file has such a value, abhyasa uses the nearest end of the range and tells you on the page. Where a value is not the default, “Use the default” puts the default back. “Use the defaults for this group” does the same for the whole group. “Find a setting” filters the rows as you type.

Some guards have no setting: the closed process with no network, the check of link addresses, plain HTTP for web pages, the check of a pane address, and the rule that abhyasa answers only this computer. A change to a setting works on the next question or the next walk.

Keyboard shortcuts

All keys are for the left hand. ⌥ is the Option key. To see all the keys, press ⌥F, or press ? outside a text box.

If you use Sanketa, it shows the same keys when abhyasa is in front. abhyasa writes the pack ~/.config/sanketa/packs/abhyasa.toml when it starts. It does not write over a pack that you made.

Run from source

Install the package and its dependencies:

uv sync

Start the web app on http://127.0.0.1:5077:

uv run abhyasa

The app answers only requests from this Mac. These variables change the defaults:

Other ways to use it

The web app, the terminal and the notebooks use the same tutor and the same data.

The web app

The web app is made from blocks in abhyasa/blocks:

Each page shows where you are, above its title. Esc goes up one level, ⌥T goes to Today, and Back and Forward work.

ramabana

In the ramabana terminal agent, the /learn extension gives you the tutor next to your code. To turn it on, make the file ~/.config/ramabana/extensions/abhyasa.py with this line:

from abhyasa.rama import setup

Notebooks

In leela or Jupyter, learn shows a question card. exercise gives a code exercise, and its tests run in the kernel.

from abhyasa.nb import learn, exercise
learn('reciprocal-rank-fusion')
ex = exercise('reciprocal-rank-fusion')

Where the data is

All data is on the learning shelf of your vault. Each concept, attempt and session is one document in the vault. Thus your other tools can search it. The tests use a vault in a temporary folder. They never open your real vault.

Demos

Some ideas are easier to see than to read. A demo is one interactive figure from rupaka. It asks you to predict first. It hides the result until you lock in (⌘↩). Then the controls work, and the last line asks a new question.

• The tutor makes a demo after two misses on a concept with the same kind of error. It also makes one for a concept with visual: true in its meta. The model writes the demo as data. rupaka checks it in a headless browser. You do not wait for it.
• After a miss on a concept that has a demo, the session shows the demo under the lesson.
• “Show me a demo” (⌥E) on a concept page shows the newest demo, or makes one.
• abhyasa keeps each prediction as an attempt with kind='demo'. A prediction changes the learner model of the concept, as an answer does. A prediction with a reason counts at depth 3 (explain). A prediction without a reason counts at depth 2 (recall).

abhyasa.demo.import_gallery(learning) keeps the rupaka gallery demos, such as vowel sandhi, for the concepts that match.

Example

This example runs offline. temp_learning makes a shelf in a temporary vault, and Fake takes the place of the model.

from abhyasa.store import temp_learning
from abhyasa.llm import Fake
from abhyasa.tutor import Tutor

lr = temp_learning()
lr.put('Ranking', 'Order items by a score.', 'ml')
fake = Fake({'You grade one answer': dict(correct=True, grade=8, first_error='', misconception='', feedback='Yes. Only the order counts.')},
            dflt=dict(question='Why can you compare ranks from two systems, but not their scores?', answer='Ranks share one scale.', rubric=['ranks share one scale']))
t = Tutor(lr, fake, context=False)

Ask the first question. why gives the reason for the choice, and method gives the way to learn that the question uses. A new concept starts at depth 3 (explain).

q = t.next()
assert q.why.startswith('next on the frontier: its prereqs are in place')   # a note on the way to learn can follow
assert q.depth == 3
print(q.question)

Answer with high confidence. A solid pass at depth 3 changes the floor to 3.

r = t.answer(q.id, 'A rank is a position, and positions use the same scale in each list.', confidence=80)
assert r.outcome == 'solid'
assert lr.concept('ranking').floor == 3