raphaelsty/knowledge

What it is

Knowledge is a personal library site that collects what influential people in AI, science, and software find interesting. It's an “opinionated reading list” of 450+ contributors — researchers like Geoffrey Hinton, Yoshua Bengio, Yann LeCun, and Andrej Karpathy, plus builders like Sam Altman and François Chollet. The tagline: No ads, no algorithm, just what they read.

Core idea

Each contributor has their own page that looks like a curated bookshelf: their tweets, GitHub stars, papers they wrote, videos they appear in. You browse someone’s page like you'd flip through a friend's bookmark folder. Click a heart on any card and the document gets saved into your personal library — indexed and searchable. A talk from last week is still worth watching this week; a paper from two months ago still teaches what you came for.

Search

Type a query and ColBERT (a semantic search model) searches the actual contents of every document, not just titles. It ranks results by how well the words match. You can search one library, several at once, or the whole shared corpus of ~440,000 documents.

The API / MCP server

The site exposes an MCP server (a way for AI agents to interact with the service) at /mcp with fifteen tools:

• Search & discover (7 tools): query a single library, search across libraries, find libraries by description, find similar docs, get latest docs, get chronological feed, find docs shared between libraries.
• Catalog (5 tools): list all libraries, list sources for a library, list tags, get one library's metadata, get one document by URL.
• Authenticated (3 tools, require a bearer token you mint at your profile page): view your saved docs, view your activity timeline, save a document.

You can wire this MCP server into Claude, Cursor, or any agent that speaks MCP.

How it works

A pipeline runs all day, walking through each person's sources in a continuous loop: GitHub stars, X posts, Hacker News submissions, arXiv papers, Hugging Face likes, Reddit, Stack Overflow, Wikipedia references, YouTube channels, Zotero libraries, and any blog you can point it at via RSS or sitemap. Each document gets cleaned, tagged, and written to Postgres. A separate indexer daemon picks up new rows and embeds them with ColBERT, so search stays current without blocking the main pipeline.

When you type a query, the API serves ranked results from a next-plaid index sitting on local disk. Your browser then does a second pass with an unquantized ColBERT running in WebAssembly (WASM) to re-rank the results on your device — the project describes this as “an original trick.”

Tech stack

• sources/ — Python (fetchers and orchestrator)
• api/ — Rust (search, ingest, auth, MCP)
• web/ — plain HTML and JavaScript (no framework)

The API is a single Rust binary. The whole thing runs on one Hetzner CX33 VPS in Helsinki: 4 vCPUs, 8 GB RAM, ~$15/month. No CDN, no managed Postgres, no Cloudflare proxy. The 3.8 GB ColBERT index sits on local disk.

The project started four years ago on a cherche backend and now runs on next-plaid and pylate-rs, the same search stack behind ColGREP (a semantic code-search tool).

Hosting and license

• Free to use and read.
• To self-host: clone the repo, set five environment variables, point a domain at the box, push to main.
• Licensed under PolyForm Noncommercial 1.0.0 — free to use, modify, and self-host for non-commercial purposes. Contact the author for commercial use.

Why it helps

The pitch: most platforms compete for attention with infinite feeds, ads, and manipulative recommendation algorithms. Knowledge does the opposite — small, finite libraries you can return to. Use it to research a topic across experts: search 450+ libraries at once for "speculative decoding" and get curated context instead of random Google noise. Browse Karpathy's GitHub stars, LeCun's papers, Hinton's interviews — all in one place. As the author puts it: "Stop doomscrolling X. The site compresses someone's year of tweets into a static page you can read once and close."