How Satyashodhak works

Watch it search

Every page of 553 books was split into 43,219 passages. Each one was turned into a vector ahead of time, a list of 384 numbers that captures its meaning, and stored in a Neon Postgres database. Every dot is one of those passages.

Where each part runs

Your browser

Turns your question into a vector, and checks the sentences in what comes back

Neon

A Postgres database in Singapore holding the passages and their precomputed vectors. It finds the closest ones using pgvector, alongside keyword search.

Vercel

Runs the app in Mumbai, and passes your question between your browser, Neon and Gemini

Gemini

Writes the optional summary

The model on your device

Model

all-MiniLM-L6-v2 ↗, compressed to run in a browser

Download

About 23 MB from Hugging Face on your first search, then kept by your browser

Runs with

Transformers.js, inside this page

Its job

Turning text into numbers. It doesn’t write anything.

Reading the confidence score

• 80+Strong match. May answer your question.
• 60+Related material. May not answer it.
• <60Hidden, unless you choose to see it.

The score is mostly how close a passage’s nearest sentence is in meaning, with a small bonus for sharing your question’s keywords. It isn’t a probability, and a high score doesn’t mean the text answers you or is true, so read the passage.

How to read the map

The 3D view is a UMAP projection, which squeezes 384 dimensions into three, so nearness there is only a rough guide. On the disc, each passage’s similarity to your question, calculated from the original 384 numbers, is drawn as its distance from the centre; its angle means nothing. Scores come from the original numbers too, not from anything you see in the picture, and the examples are precomputed.

The summary

With Contextual summary on, the app sends your question and up to five passages to Google’s Gemini 3.5 Flash-Lite, which writes a short summary with numbered citations. It’s told to use only those passages and to decline if they don’t answer. If no passage reaches 60, Gemini isn’t called. The passages themselves are never rewritten. Turn it off to see the excerpts alone.

What leaves your browser

• Your question and its vector go to the app on Vercel, which searches the database on Neon.
• With the summary on, your question and the passages also go to Google.

What it can’t do

• The model works best in English.
• It reads text only. To see diagrams and tables, use Open exact page.
• It can rank things wrongly, so a good answer may score low, and a poor one high.

Source

Publicly accessible material from the National Science Digital Library at NIScPR. Rights remain with the authors and publishers. Browse the NSDL collection ↗