Rajgruha

The question

In a digital-first world, what is a physical public library for?

India has plenty of public libraries, and digital natives barely use them. eGranthalaya alone lists 5,380 libraries and 2.8 crore catalogue records. Of the 27 people I surveyed, 12 had ever used a government public library.

Systems mapping

Nothing faces the public. That’s the gap.

Money, laws and software all stop at the branch. None of it is built to reach readers. eGranthalaya is the one place where library records already exist, so that’s where a way in for readers could start.

eGranthalaya works as designed. It was designed for librarians.

NIC built it with a dream of “a single window access system … for government libraries”, and for librarians it is one. Readers were never part of the brief.

On the ground, one catalogue kiosk was dark and another was on but had no internet, the catalogue address had to be typed by hand, and a shared login was printed openly on a sign.

It isn’t only Bengaluru. A 2024 study of Maharashtra’s public-library catalogue found its search problems were “mainly relevant to the system”, not to the people searching. When Madhya Pradesh put 528 colleges on eGranthalaya at once, so few students registered that the state had to run a campaign.

The insight

Two realistic options: redesign eGranthalaya, or open its records to everyone.

My first instinct was to redesign the portal. The map changed my mind: a better portal is still a staff tool, and every library stays an island. The records were never the problem. The missing piece was a layer that faces readers.

Redesigning the portal would be a band-aid. Public libraries need surgery.

The libraries

So I went to all kinds of libraries.

First a library walk hosted by IIHS and the Goethe-Institut, then my own visits across Bengaluru, Delhi and Gurgaon, looking at the same three things: discoverability, digital readiness and ambience.

PublicPrivateCulturalAcademicBookstores and spaces

12 places, 3 cities. Pick a photo, a label or a city.

• Bengaluru 9
  State Central LibraryCity Central, YelahankaCity Central, MalleshwaramNIDIIHSMythic SocietyGoethe-InstitutAlliance françaiseChampaca
• Delhi 2
  British CouncilMayday
• Gurgaon 1
  Coforge

Every book has an address. You’re expected to know how to read it.

And it isn’t the same in every library.

Some libraries shorten the number, some add their own prefixes, and some don’t use Dewey at all. Learning one library’s system doesn’t help you in the next.

Every library breaks somewhere different.

You’re in the library and want one book. Instead of browsing the shelves or asking the librarian, you go through the catalogue. Here’s how far it gets you.

At Yelahanka, borrowing starts on paper.

At Coforge, the number is inside the cover.

Curation

Some spaces have what a catalogue can’t: curation and a wider world.

Events

And some give you a reason to walk in.

Readers

They love books. That doesn’t extend to public libraries.

I surveyed 27 digital natives and interviewed 4. Their libraries were the ones they were enrolled in: all of them had used a college library. Otherwise, reading moved online or to a shop.

Why? They don’t know what’s near, and don’t expect much of it.

Digital platforms didn’t help.

15 of 27 had never tried a government digital library, and those who had mostly rated it 2 or 3 out of 5. “Loading time, unable to download.” “Do pirating websites count T.T”

So I showed them Oodi.

A walkthrough video of Helsinki’s central library, then the question of what they’d use a place like it for.

What it indicated

What those three things indicated.

The opportunity

We live with information abundance, and yet we seem more lost and unsure than ever.

AI summaries give us confidence we haven’t earned. Language models are probabilistic, so a summary can’t be trusted until it’s verified, and the world is filling up with them. When 22 public broadcasters tested four AI assistants on the news, 45% of answers had a significant issue.

Libraries as the intellectual and epistemic backbone of society.

Maybe that’s the answer to the question I started with. But how do we operationalise it?

Build it like public infrastructure, not like an app.

India already has a model. UPI lets any bank account work in any app, and ONDC lets a seller listed once reach every buyer app. The G20 calls this digital public infrastructure: shared systems built from minimal, modular blocks. Libraries fit it well. The records already exist, and they’re publicly funded.

So the work splits in two. The government builds the rails, the way NPCI built UPI: open specifications for catalogue records, live holdings, a registry of libraries and a standard code for every book. On top sits something like BHIM: a reference app that proves the rails work, with room for anyone else’s beside it. The records stay where they are, and membership and lending stay with the branch.

Other governments already open their catalogues this way. Singapore’s National Library Board has a catalogue search API with live availability at its libraries, and Australia’s Trove lets anyone with a free key build on the national collection. eGranthalaya’s records leave only through a staff login. There’s no public API to build on, and no standard feed that national collections like NDLI could gather them from.

Enter Rajgruha

A name from Ambedkar’s library. A mark from Buddhist chaitya arches and open books.

Rajgruha

• Near me
• मेरे आस पास
• ನನ್ನ ಸುತ್ತಲೂ
• എനിക്ക് ചുറ്റും
• என்னை சுற்றி
• నా చుట్టూ
• আমার চারপাশে
• माझ्या आजूबाजूला
• میرے ارد گرد
• મારી આસપાસ
• মোৰ চাৰিওফালে

• Near me
• मेरे आस पास
• ನನ್ನ ಸುತ್ತಲೂ
• എനിക്ക് ചുറ്റും
• என்னை சுற்றி
• నా చుట్టూ
• আমার চারপাশে
• माझ्या आजूबाजूला
• میرے ارد گرد
• મારી આસપાસ
• মোৰ চাৰিওফালে

Rajgruha

DM SansSee your reading lifeOneQR

• 
• 
• 
• 
• 

SatyashodhakAsk a question…How do vaccines protect the body?

• 000General
• 300Society
• 500Science
• 800Literature
• 900History

RajgruhaReader cardOneQR

✓Issued

Scan to find a bookಪುಸ್ತಕ ಹುಡುಕಲು ಸ್ಕ್ಯಾನ್ ಮಾಡಿकिताब ढूँढने के लिए स्कैन करें

Rails from the government, Rajgruha on top.

One catalogue, next, is the rails: what the government would build once, for everyone. Satyashodhak, the feed and curation are Rajgruha’s own, built on top. OneQR is both: a standard on the rails, and a feature in the app.

One catalogue

First, the biggest problem: one catalogue out of 76.

Even NIC can’t agree on the count: 76 clusters on its home page, 74 at login and 77 on its dashboard. That’s why a registry of libraries is one of the rails.

Once that exists, two opportunities appear. One search across every library.

Search the network and find it at your nearest library. Simple and location-specific, so it joins the habit: alongside, or ideally instead of, Amazon or LibGen.

OneQR

OneQR takes the reader from screen to shelf.

You don’t have to learn Dewey. Enter the library and you’re shown where the book is, with guidance to the shelf. Or show the QR and the librarian finds it for you. It also means instant checkout. This depends on Dewey being implemented and digitised on the ground. And it only works if every branch reads the same code, so OneQR should be an open standard on the rails, not a feature of one app.

Back to those four libraries.

Satyashodhak

How do we make libraries the de facto source of truth?

Satyashodhak quotes verified books instead of generating answers. It’s the second digital-native behaviour: just ask. It fetches excerpts from a vector index, so it can’t make up the text, and it surfaces material you’d never have found. This is retrieval-augmented generation (RAG) with the generation made optional, and the more it scales, the more useful it becomes as an alternative to today’s answer engines.

It started with one book, on my laptop.

The smallest test that could fail, all on my laptop. First, indexing: the book is cut into passages, a local embedding model turns each one into a vector, a list of numbers standing for its meaning, and a FAISS index stores them. Then retrieval: your question becomes a vector the same way, and FAISS returns the passages closest to it, with their pages. Nothing is written, so nothing can be made up.

One bookSplit into passages, once

Your questionEvery time you ask

Local embedding modelTurns text into a vector

FAISS indexStores a vector per passage, finds the nearest

The exact passageWith its page

Where it fell short.

Then a whole collection, live.

The National Science Digital Library had no way to get inside its text and answer questions semantically. So: 553 books, 43,219 passages, every page. The passages become vectors once, ahead of time. When you ask, a small model in your browser does the same to your question, so a search doesn’t need a paid AI call. The server compares it with every passage, adds a small boost for shared keywords, and keeps some variety across books. Back in your browser, each sentence of those passages is checked against the question, because one sentence can answer what a whole page blurs. Because of the feedback on bare excerpts, a short Gemini summary can sit on top. The excerpts are never rewritten.

Future: answers made of excerpts.

I haven’t tried building this yet, but here’s what I’d want. Instead of a list, the answer is a layout made of the authors’ sentences. The agent never writes them: it sends the layout as data, the way A2UI does, and each quote is only a pointer to a sentence in a passage. Perplexity writes and cites; this would let the authors speak.

At scale: millions of books.

NSDL is short teaching modules; a real book is about 500 passages. If I took this to HathiTrust’s 17.6 million volumes, that’s around 9 billion passages, roughly 7 TB of vectors. When I worked it through, search wasn’t the hard part: DiskANN searched a billion vectors on one machine in under 3 ms, so that’s about nine machines. Judging what comes back is.

That judge is where I’d experiment. System One models like Jev don’t write: they answer a question about an input with a probability, cheap enough for hundreds of candidates, and unable to reword an author. One recent paper found them no better than tuned rerankers, though Jev matched them, so it’s a test to run.

Using Rajgruha

First pass: a map of nearby libraries. It wasn’t the point.

It seemed less important than what people actually needed: which libraries are nearby, and the events that could lead them in. So: browse libraries and events, with a dynamic feed. It shows the libraries near you, their collections and what events they hold, along with distant but interesting ones.

Discovery first. Search second. Three ways to reach the same book.

A dynamic feed and a hybrid curation system, inspired by contemporary libraries, Coursera and reading lists. Systematised as the third way in.

Your reading life, across every library.

Every loan, every library you’ve been to and every book you’ve saved lands in one place, whichever branch it came from. That only works because membership is shared on the rails, not kept separately by each library.

Making it real

First, talk to officials.

The proposal rests on agreements I haven’t secured yet, so the next step is to go and ask. Three conversations decide whether it can work, and each has a short list of questions.

Then one pilot, small enough to say yes to.

A pilot needs both sides working. Rajgruha exists as a prototype, but librarians need their own MVP too: holdings, checkout with OneQR, and their events. Then the department publishes the rails for one cluster in Bengaluru, and the pilot runs for twelve weeks. It tests the rails as much as the app: Rajgruha is simply the first thing built on them. It reads the records that already exist, so branches change nothing about how they lend.

What I’d test, and how I’d know.

And what policy would need to change.

A pilot can prove the rails. Making them the norm needs a few rules to change, most of them about records and money.

The problem isn’t access. It’s structure, curation and trust.

I started with a catalogue interface, and the research widened the problem. The harder part was deciding which connections to share, which to keep local, and which part to build well enough to test. Still open: the librarian’s workspace, partner access to catalogue records, and a pilot with a real branch.