A local-first study workspace for programming books. Read the PDF, take notes, ask a book-grounded mentor and run the code, all on one screen and all on your machine.
Learning from a technical book usually means juggling a PDF viewer, a notes app, an editor and a chatbot, and that chatbot answers from its training data rather than from your book.
DevBook AI brings these into one tri-pane workspace:
| 📖 Reader | 📝 Notes | 🤖 Copilot |
|---|---|---|
| Read the PDF, highlight any passage or code listing | Collect quotes with page references; run code blocks inline | Ask questions answered only from the book, with clickable page citations |
- 🔒 Private by design. Books, notes, embeddings and the model all stay on your machine. No accounts, no API keys, no telemetry.
- 💸 Zero running cost. Inference runs locally through Ollama.
- 🎯 Grounded answers. Every answer cites pages. If the book doesn't cover something, the copilot says so instead of guessing.
- 🧪 Run what you read. Execute Python and JavaScript snippets in an isolated Docker sandbox and see the output right under the code.
- PDF upload with structure-aware parsing: headings, page numbers and code blocks are preserved
- Local vector index (ChromaDB +
nomic-embed-text) - Streaming, book-grounded chat with a mentor persona
- Clickable
[p. 42]citations that jump the reader to the source - Highlight → Explain via AI / Send to Notes
- Per-book Markdown notes, auto-saved as plain
.mdfiles - Sandboxed code runner for Python & JavaScript (no network, resource-limited)
- Health panel for Ollama, models and Docker, with one-line fixes
- Full Markdown Studio (CodeMirror editor, multiple notes per book)
- Practice Bug generator: get a subtly broken version of a snippet to debug
- Active-recall quizzes per chapter, and note review against the book
- EPUB support, library-wide questions, model picker
- One-click desktop app (Tauri), spaced repetition, more runner languages
flowchart LR
subgraph Browser["Next.js app · localhost:3000"]
R[Reader] --- N[Notes] --- C[Copilot]
end
subgraph API["FastAPI · localhost:8000"]
ING[Ingest<br/>Docling → chunks]
RAG[RAG<br/>retrieve → prompt → stream]
RUN[Runner]
DB[(SQLite)]
VEC[(ChromaDB)]
FS[(PDFs & notes)]
end
OLL[Ollama<br/>chat + embeddings]
DOCK[Docker<br/>sandbox]
Browser -->|HTTP / NDJSON| API
ING --> FS & VEC & DB
ING & RAG --> OLL
RAG --> VEC & DB
RUN --> DOCK
- Ingest: Docling turns the PDF into a structured document. Its hybrid chunker splits it along headings without breaking code listings. Each chunk carries its page range and heading path.
- Index: chunks are embedded locally with
nomic-embed-textand stored in ChromaDB. - Ask: your question retrieves the most relevant passages from the open book, and a local model (default
qwen2.5-coder:7b) answers from those passages only. Citations come from retrieval metadata, so page numbers are never invented. - Run: code blocks execute in throwaway Docker containers with no network access, no host mounts and strict CPU/memory/time limits.
| Layer | Technology |
|---|---|
| Frontend | Next.js (App Router), TypeScript, Tailwind CSS, shadcn/ui, react-resizable-panels, react-pdf |
| Backend | FastAPI, Python 3.12 |
| Parsing | Docling + HybridChunker |
| Storage | ChromaDB (vectors), SQLite (library & chat history), plain files (PDFs & notes) |
| Models | Ollama: qwen2.5-coder:7b (chat), nomic-embed-text (embeddings) |
| Sandbox | Docker (python:3.12-alpine, node:22-alpine) |
Everything runs locally, so your hardware decides which model you can use and how fast it answers.
| Tier | Hardware | Suggested model | Experience |
|---|---|---|---|
| Minimum | CPU only, 16 GB RAM | qwen2.5-coder:3b |
Works, ~5–10 tokens/s |
| Recommended | NVIDIA GPU with 8 GB VRAM, or Apple Silicon with 16 GB | qwen2.5-coder:7b |
Smooth, ~40–60 tokens/s |
| High | 12–16 GB+ VRAM | qwen2.5-coder:14b |
Stronger reasoning |
Plan for about 10 GB of free disk for models and images. On 16 GB machines, consider capping Docker Desktop's WSL2 memory (.wslconfig) to keep things responsive.
⚠️ DevBook AI is under active development. The steps below describe the intended setup; the application code is landing incrementally.
- Ollama
- Docker Desktop (only needed for the code runner)
- Python 3.12+
- Node.js 22+
ollama pull qwen2.5-coder:7b
ollama pull nomic-embed-text
docker pull python:3.12-alpine
docker pull node:22-alpinecd backend
python -m venv .venv
# Windows: .venv\Scripts\activate | macOS/Linux: source .venv/bin/activate
pip install -e .
uvicorn app.main:app --host 127.0.0.1 --port 8000NVIDIA on Windows: install the CUDA build of PyTorch first for faster parsing:
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
cd frontend
npm install
npm run devOpen http://localhost:3000, drop in a PDF, and start studying.
The backend reads optional environment variables:
| Variable | Default | Description |
|---|---|---|
DEVBOOK_CHAT_MODEL |
qwen2.5-coder:7b |
Ollama chat model |
DEVBOOK_EMBED_MODEL |
nomic-embed-text |
Ollama embedding model |
DEVBOOK_OLLAMA_URL |
http://127.0.0.1:11434 |
Ollama endpoint |
DEVBOOK_DATA_DIR |
./data |
Where the library, index and notes live |
DEVBOOK_TOP_K |
6 |
Passages retrieved per question |
DEVBOOK_RUN_TIMEOUT |
10 |
Sandbox time limit (seconds) |
DevBook-AI/
├─ backend/ # FastAPI: ingestion, retrieval, chat streaming, notes, code runner
├─ frontend/ # Next.js: reader, notes, copilot
└─ data/ # created at runtime (gitignored): SQLite, ChromaDB, PDFs, notes
DevBook AI makes no outbound network requests during normal use. The only downloads are ones you start yourself: model pulls through Ollama and Docker image pulls. Your books and notes are regular files in data/ that you can back up, move or delete at any time.
Contributions are welcome once the MVP lands. Until then, feel free to open an issue with ideas, bugs or book-parsing edge cases (tricky PDFs are especially useful).
- Fork the repo and create a branch:
git checkout -b feat/your-idea - Use Conventional Commits (
feat:,fix:,docs:…) - Open a pull request describing the change and how you tested it
MIT © 2026 ZeroElemental
The models you run are covered by their own licenses (for example, Qwen2.5-Coder and nomic-embed-text are Apache-2.0). You're responsible for having the right to use the books you import.