github.devotd.
NewAI

lfnovo/open-notebook v1.15.0 — a self-hosted NotebookLM alternative

lfnovo/open-notebook is an open-source implementation that lets you host Google NotebookLM yourself, supporting research organization and chat with multiple AI providers and local models, plus multi-speaker podcast generation.

Repositorylfnovo/open-notebook ↗

lfnovo/open-notebook is an open-source implementation that lets you host Google NotebookLM yourself, supporting research organization and chat with multiple AI providers and local models, plus multi-speaker podcast generation.

GitHub description: An Open Source implementation of Notebook LM with more flexibility and features

stars
★ 40.4K
Forks
4.7K
Language
TypeScript
License
MIT
Latest release
v1.15.0
Read original source ↗View on GitHub ↗

The latest release of lfnovo/open-notebook is v1.15.0. It is an open-source implementation of Google NotebookLM, the note-based AI service, that you host yourself. It bundles materials and model conversations into notebook-sized workspaces so you can research and organize.

The README describes this repository as a more flexible and feature-rich implementation than NotebookLM. It organizes multiple sources — PDFs, videos, audio, web pages — into a single notebook, lets you chat grounded in those sources, and generates podcasts with multiple speakers. It provides both full-text search and vector search, cites sources for every answer, and can be called from external programs through a REST API. Because you can use several AI providers and local models together, you can choose yourself which model handles the material. The code is written in TypeScript and the license is MIT.

However, the given record does not include the list of changes contained in v1.15.0. To know what this version fixed, you have to check the release notes directly. There are also no measurements of how large the input material can be or how accurate the answers are. The repository was created on October 21, 2024 and has 40414 GitHub stars, but the momentum score in the collected research record is 6.0 out of 10, a middling level. The feature list is based on the README description, and actual performance and the improvements in this version must be confirmed through the release record and direct testing.