NotebookLM-Lite is an open-source NotebookLM-style AI knowledge workspace for local files, grounded RAG chat, notes, interactive Studio artifacts, and long-form podcast generation. It is built for people who like the Google NotebookLM workflow but want a hackable, self-hosted project with their own model endpoints, their own document pipeline, and a clear backend architecture.
NotebookLM-Lite is not affiliated with Google NotebookLM. The goal is to bring most of the NotebookLM learning and research experience into an open project: source-grounded answers, citations, mind maps, flashcards, quizzes, reports, data tables, infographics, podcast/audio overviews, notes, native slide deck workflows, and future video workflows.
The demo uses doc/L9.md as the sample source and shows model configuration, source upload, cited RAG chat, interactive Mind Map, Flashcards/Quiz, Data Table, Podcast, and the Video Overview placeholder. Model waiting time is shortened for README pacing.
- Native PPT / Slide Deck generation is now available in NotebookLM-Lite. You can generate a deck from selected notebook sources, confirm the outline, confirm the per-slide prompt plan, generate slide images, regenerate or edit a single slide, and export the result as PPTX.
- The current NotebookLM-Lite PPTX export is image-based: each slide in the exported PPTX is a full-page static image. This keeps the generation path reliable and easy to preview, but individual PowerPoint elements are not yet native editable shapes.
- Fully editable PPT generation has already been implemented in LRriver/AIPPT. NotebookLM-Lite will integrate that editable PPT workflow next; you can try the complete editable-PPT generation experience today in LRriver/AIPPT.
NotebookLM-style tools are becoming the default interface for working with long documents, research notes, study material, product manuals, papers, and course content. NotebookLM-Lite focuses on the same core loop:
- Add sources to a notebook-like knowledge base.
- Ask questions grounded in selected sources.
- Save useful answers and notes back into the knowledge base.
- Generate structured outputs such as maps, quizzes, reports, tables, infographics, and podcasts.
- Keep every generated artifact downloadable and inspectable.
That makes the project useful as a NotebookLM alternative, a RAG notebook starter, an AI study assistant, a document Q&A system, and a backend reference for multi-step AI workflows.
- NotebookLM-like three-panel workspace: sources on the left, grounded chat in the center, Studio artifacts and generated files on the right.
- Multi-format source ingestion: upload or paste sources such as PDF, DOCX, TXT, Markdown, HTML, CSV, JSON, and YAML.
- Document intelligence pipeline: Docling extracts document text, Chonkie chunks it, and SeekDB persists sources, chunks, notes, jobs, and artifacts.
- Hybrid RAG retrieval: SeekDB combines lexical BM25-style recall, optional embeddings, optional rerank, selected-source filtering, and citations.
- Streaming grounded chat: ask questions over selected sources, render Markdown answers, and keep citations attached to retrieved chunks.
- Notes as first-class knowledge: create notes, save chat answers as notes, and convert notes back into sources for later RAG.
- Interactive Studio artifacts: generate Mind Maps, FAQ, Flashcards/Quiz, Reports/Study Guides, Data Tables, SVG Infographics, Podcast scripts, and Slide Decks from selected sources.
- NotebookLM-like study UX: flashcards flip, quizzes show feedback and scores, mind maps expand, tables render as tables, and artifacts can be downloaded as Markdown/JSON/SVG where applicable.
- Native Slide Deck workspace: generate a deck outline, confirm/edit it, generate a per-slide image prompt plan, confirm/edit it, render slide images, regenerate or edit one slide, and export an image-based PPTX.
- Podcast and Audio Overview workflow: structured podcast scripts are generated with Pydantic validation and can be expanded toward controllable durations up to 30 minutes; speech generation is optional when a compatible audio model is configured.
- Unified model runtime: LiteLLM manages text generation, embeddings, rerank, and OpenAI-compatible speech profiles. OpenAI-compatible, Anthropic, Gemini/GenAI-style, and custom gateway deployments can be configured from
config.yaml. - Extensible backend boundaries: FastAPI owns resource APIs, while longer workflows are isolated for future LangGraph orchestration, Deep Research, richer Slide Deck automation, and Video Overview support.
| Capability | Status | Notes |
|---|---|---|
| Source upload and pasted text | Available | File sources and text sources are stored in SeekDB. |
| RAG chat with citations | Available | Supports selected sources, Markdown answers, and streaming response UI. |
| Notes | Available | Create notes, save answers as notes, and convert notes into sources. |
| Mind Maps | Available | Structured tree output with expandable frontend viewer. |
| FAQ | Available | Structured FAQ generated from selected sources. |
| Flashcards and Quizzes | Available | Flip-card viewer plus quiz feedback, score, and retake. |
| Reports and Study Guides | Available | Structured summaries, sections, and takeaways. |
| Data Tables | Available | Rendered as real tables and downloadable as artifact data. |
| Infographics | Available | Structured SVG infographic generation and safe image rendering. |
| Podcast / Audio Overview | Available | Script generation is core; audio download is available when speech config is valid. |
| Runtime model configuration | Available | Separate text, embedding, rerank, speech, image, and edit profiles. |
| Deep Research | Placeholder | API/job boundary exists so future research output can be saved as a source. |
| Slide Deck / PPT | Available | Native Slide Deck workspace with two confirmation steps, generated slide previews, single-slide regenerate/edit, and image-based PPTX export. |
| Video Overview | Placeholder | Studio card is present and shows an in-development notice. |
frontend/
React + Vite workbench
SourcePanel, ChatPanel, NotesPanel, StudioPanel, ArtifactViewer, SlideDeckWorkspace
backend/
FastAPI routes for sources, chat, notes, artifacts, podcast, slide decks, config
Docling parser + Chonkie chunking
SeekDB repository and vector-store adapter
LiteLLM provider for text, structured output, embeddings, and rerank
Podcast workflow with Pydantic structured output
Native slide deck workflow with generated image assets and PPTX export
Key backend choices:
- FastAPI for CRUD, list, download, runtime config, and status APIs.
- Docling for document parsing.
- Chonkie for chunking, with a simple fallback.
- SeekDB as the project knowledge repository.
- LiteLLM as the unified model access layer.
- Pydantic structured output for Studio artifacts and podcast scripts.
- Native Slide Deck workflow integrated into NotebookLM-Lite state, jobs, artifacts, model config, and downloads.
- Python 3.10+
- Node.js 18+
- A text model endpoint configured in
config.yaml - Optional: embedding, rerank, and speech model endpoints
pip install -r requirements.txt
cp config_example.yaml config.yamlEdit config.yaml and fill in the profiles you want to use:
api.models.text_model: chat, RAG answers, Studio artifacts, podcast scriptsapi.models.embedding_model: optional vector retrievalapi.models.rerank_model: optional rerankapi.models.audio_model: optional speech/audio generationapi.models.image_model: slide image generationapi.models.edit_model: single-slide image editingstorage.seekdb_path: local knowledge databasedocuments.chunking: Chonkie/simple chunking settings
Keep config.yaml local. It is ignored by git and should not contain committed keys.
Start the backend:
python -m backend.mainThe API runs at http://localhost:8000.
cd frontend
npm install
npm run devOpen http://localhost:5173.
- Open the workbench and check the settings panel.
- Upload a source file or paste text in the Sources panel.
- Select one or more sources.
- Ask source-grounded questions in Chat.
- Save useful answers as notes or sources.
- Generate Studio artifacts:
- Podcast
- Mind Map
- FAQ
- Flashcards / Quiz
- Report / Study Guide
- Data Table
- Infographic
- Slide Deck / PPT
- For Slide Decks, open the dedicated workspace, generate and confirm the outline, generate and confirm the prompt plan, render slide images, optionally regenerate or edit one slide, then export/download the PPTX.
- Download generated artifacts as Markdown, JSON, SVG, transcript, PPTX, or audio where supported.
Backend:
PYTHONPATH=. pytest -q
PYTHONPATH=. python -m compileall -q backendFrontend:
cd frontend
npm run build
npm run test
npm run test:e2e -- --project=chromiumOpenSpec:
openspec validate notebooklm-slide-deck-heavy-integration --strictNotebookLM-Lite already covers the core NotebookLM-style loop: sources, cited chat, notes, Studio study artifacts, data tables, infographics, podcast scripts/audio, and native Slide Deck generation with image-based PPTX export. The next major areas are:
- Real Video Overview generation instead of the current placeholder card.
- Stronger Deep Research workflows that can search, synthesize, and save research reports as sources.
- Editable PPTX export, integrating the native-shape workflow already available in LRriver/AIPPT.
- More export targets, including CSV/Sheets-style table export and richer artifact sharing.
- More source types, such as URLs, audio, images, YouTube, and Google Drive-style integrations.
The Slide Deck integration is intentionally native to NotebookLM-Lite:
- It uses the same source selection, LiteLLM/runtime model configuration, SeekDB persistence, job APIs, artifact list, and download flow as the rest of the app.
- The workflow has two human confirmation points: outline confirmation and prompt-plan confirmation.
- Phase 1 exports a PPTX where each slide is a generated full-slide image. This is a PPTX export, not a promise that every PowerPoint element is a native editable shape.
- Image generation and editing use the same NotebookLM-Lite runtime model profiles as the rest of the app; credentials stay in local ignored config files.
- The integrated workflow is native to this repository and does not require running a separate AIPPT server.
Apache 2.0