Youssef

note-lm

Self-hosted research notebook: ingest PDFs, videos, YouTube and pages; transcribe in the background; chat with citations across all sources

Code on GitHub

Language
TypeScript
Licence
MIT
Last push
1 Oct 2026
Stars
1

What it is

Everything runs in one installed Windows program. Your data remains on your disk; remote AI is opt-in and requires an explicit provider connection.

How it works

  1. Versioned evidence. Every import is an immutable source version. Claims anchor to that version with a page (or audio time range) and the quoted text. Re-importing a changed document appends version 2, nothing is overwritten, and a deterministic scan flags claims whose quote moved or vanished as review proposals. The app flags changed inputs; it never decides your conclusion is false, and every accept/reject keeps its history.
  2. Everything inside one program. A Tauri shell spawns the engine and talks NDJSON over stdio; background workers persist through pause, resume, cancel, and recover after a crash, resuming downloads from the partial bytes on disk. The release gates prove this by killing the engine mid-work and restarting it.
  3. Local-first AI, cloud strictly optional. With zero configuration, chat, embeddings and transcription resolve to local llama.cpp / whisper.cpp. Remote providers need an explicit connection plus a key stored in the OS keyring, never in a file, never in logs. An offline mode blocks remote calls at call time; local calls are never blocked.
  4. Portable research packages. Export a notebook as a hash-verified package, with or without the original files, and restore it anywhere. Claims, versions, review proposals and calculations ride along.
  5. Reproducible calculations. Ask for a sum over a CSV and the engine parses the sheet with strict validation and computes it deterministically. An LLM never does arithmetic here, and a file that isn’t a clean sheet gets a typed error, not a guess.

Results

1. Bring the source in 2. Save the evidence 3. Review what changed
Import PDFs, pages, CSV tables, audio, or video. Select a passage and save a claim pinned to its source version. Re-importing creates review proposals; the original claim remains traceable.
Local storage, local search, local models by default. Open every citation back in the reader. Compare claims and sources in the evidence matrix.