Show HN: Bowerbird – Open-source NotebookLM for any model, with checked quotes

An open-source NotebookLM alternative where your files stay yours and every quote is checked.

Add PDFs, books, web pages, YouTube videos and recordings, then ask questions. Answers cite the passages they draw on, and each quoted phrase is looked up in the source it cites. Notebooks are plain folders on your disk, search, OCR and speech run on your machine, and answers come from any model you choose.

Each citation in an answer carries the words it quotes. Bowerbird looks those words up in the source the citation names: in the cited passage, or in another passage of the same source, since a quote can run across a passage boundary. A quote that isn't there, or that the answer attributes to the wrong source, is marked unverified, and the footer counts how many were found. A quote shortened with an ellipsis counts only when each of its pieces is long enough to stand on its own and the pieces sit close together.

The check shows that the quoted words are in the source, not that the answer reads them correctly. Click a citation to open the passage with the quoted words highlighted, or, for a PDF, the printed page, and judge for yourself.

Notebooks, the search index, OCR, transcription and the voices of audio overviews stay on your machine. Bowerbird sends no analytics and doesn't check for updates. This is everything that goes out:

API keys go only to their own provider. Choose an Ollama model running on your machine, and questions and passages stay there too.

Sources: PDF, Word (.docx), EPUB, Markdown, plain text, web pages, YouTube captions, and audio recordings, which are transcribed on your machine with Whisper. Scanned PDFs are read with OCR on your machine (PP-OCRv5, Latin script including accented letters; Cyrillic and other scripts aren't read yet). Pictures in web pages, Word documents and EPUBs are kept beside their captions.

Search by keyword and by meaning: passages are found with SQLite FTS5 (BM25) and with EmbeddingGemma vectors, fused by weighted reciprocal rank, so a question worded differently from the source still finds it. Questions can be asked in a different language from the sources.

Notes: save answers to notes, edit them, and add a note back as a source.

Reports: a study guide, briefing, timeline, FAQ, or a report you describe, written into a note with citations.

Study material: flashcards (with Anki export), quizzes, mind maps (SVG), data tables (CSV), slide decks (PPTX), infographics (SVG, PNG), and audio overviews in four formats, voiced on your machine with Kokoro.

Finding sources: search the web through DuckDuckGo, SearXNG, Brave, Tavily or Exa, and add what you pick.

Runs where you work: a web app served from your machine, a desktop app for Linux, macOS and Windows, and an Android app. Light and dark themes.

Builds for Linux (.deb, AppImage), macOS 13.4 or later (Apple Silicon .dmg), Windows (portable .zip) and Android (64-bit ARM .apk) are attached to each release. The macOS app is signed ad hoc, not notarized, so macOS blocks its first launch: allow it under System Settings, Privacy & Security, Open Anyway. To build from source, see Build and run.

A source is extracted to text with its heading tree: PDF bookmarks or font sizes, HTML and Markdown headings, EPUB and Word structure. Scanned PDFs and recordings are first read by OCR or Whisper in the background.

The text is split into passages and indexed in SQLite with FTS5. The source is searchable as soon as that finishes.

In the background, each passage gets an EmbeddingGemma vector. Until a passage has one, keyword search answers alone.

A question's keyword hits and nearest passages by meaning are fused, each hit brings its neighbors in the same section, and a source that ranks among the candidates gets its best passage even if it misses the top 8. Deep search also lets the model pick whole sections from the outline.

The model answers with citations to passage labels such as B12, and each quoted span is matched against that passage's text.

Models

The default model is deepseek-flash. Any provider that genai supports works, including OpenAI, Anthropic, Gemini, DeepSeek, Mistral, Groq, xAI, OpenRouter and Ollama. In Settings, enter the model name (prefix it with the provider, as in groq::model-name, to choose the provider yourself) and paste the provider's API key.

Keys are saved in settings.toml, readable only by your user account: ~/.config/bowerbird/ on Linux and macOS, %APPDATA%\bowerbird\ on Windows, or the path in $BOWERBIRD_CONFIG. A key set as an environment variable, such as DEEPSEEK_API_KEY, works too.

The local models download on first use into ~/.local/share/bowerbird/models (%LOCALAPPDATA%\bowerbird\models on Windows): EmbeddingGemma 4-bit for search (200 MB) when the first source is added, the OCR models (12 MB) for the first scanned PDF, and the speech models when they are first used. Their licenses, including the Gemma Terms of Use that cover EmbeddingGemma, are in THIRD_PARTY_NOTICES.md.

Each notebook is a folder in ~/Bowerbird, or in $BOWERBIRD_HOME:

<notebook>/ notebook.toml title, answer style, sources (order, letters) sources/ originals as added, plus OCR text and transcripts notes/<id>.md Markdown notes chats/<id>.jsonl one conversation: questions, answers, citations studio/<id>.json flashcards, quizzes and other study material guide.json overview and suggested questions .bowerbird/index.sqlite search index and vectors, rebuilt when missing

Copying the folder backs up or moves the notebook.

Needs Rust 1.95 or newer and Node.js 22.17 or newer. On Debian or Ubuntu, install the build tools and the desktop app's libraries first:

sudo apt-get install cmake clang libwebkit2gtk-4.1-dev libayatana-appindicator3-dev librsvg2-dev

scripts/fetch-pdfium.sh # PDFium, which reads PDFs, into vendor/pdfium export BOWERBIRD_PDFIUM_DIR="$PWD/vendor/pdfium/lib" npm --prefix ui ci npm --prefix ui run build cargo run --release -p bowerbird-server

The server prints a link with an access token. Open it in a browser. It listens on 127.0.0.1:7477; --addr changes that, and anything other than a loopback address exposes your notebooks to the network, protected only by the token. Links added through the server may only point at the public internet, not at the server's own machine or its local network, unless it runs with --allow-private-links. Leave that off when anyone else can reach the server, directly or through a reverse proxy. The desktop app allows such links.

The bowerbird command works on the same library:

cargo run --release -p bowerbird -- add history treaty.pdf https://en.wikipedia.org/wiki/Peace_of_Westphalia cargo run --release -p bowerbird -- ask history "Who signed the treaty at Osnabrück?"

Desktop app

On Linux, install Tauri's prerequisites, then:

ui/node_modules/.bin/tauri build --config crates/desktop/tauri.conf.json

The .deb and AppImage land in target/release/bundle/. On a Mac with Apple Silicon, run scripts/fetch-pdfium.sh first and add --bundles app,dmg; the .dmg lands in target/release/bundle/dmg/.

The Windows app cross-compiles from Linux with cargo-xwin (needs clang 19 or newer, lld and llvm):

scripts/windows.sh

The portable folder lands in dist/bowerbird-windows: Bowerbird.exe with DirectML.dll, pdfium\pdfium.dll and the licenses beside it.

scripts/android.sh

Builds a 64-bit ARM APK. Needs the Android SDK and NDK, found through ANDROID_HOME and NDK_HOME. The APK is signed with the release key that crates/desktop/gen/android/keystore.properties names (storeFile, relative to that folder, keyAlias and password; kept out of git), or with the debug key when that file is missing.

cargo test --workspace npm --prefix ui run check ui/node_modules/.bin/playwright install chromium # once cargo build -p bowerbird-server && npm --prefix ui run test:e2e BOWERBIRD_LIVE=1 cargo test -p bowerbird-core --test live BOWERBIRD_LIVE=1 cargo test -p bowerbird-core --lib -- "ocr::" "embed::" "retrieve::"

The end-to-end tests and the live test ask the default model real questions, so they need DEEPSEEK_API_KEY. The end-to-end tests serve the UI from ui/build, so build it first. The live test also checks YouTube captions, web page extraction and DuckDuckGo; the last line downloads the OCR and search models and reads a scanned page.

node ui/scripts/readme-screenshots.mjs <url> <token> docs/screenshots rebuilds the demo notebook and the screenshots above. Run it against a server whose --library is an empty folder; it asks the configured model real questions.

.github/workflows/ci.yml, on every push and pull request: formatting, clippy with warnings as errors, the workspace tests, the UI type check, and cargo deny for licenses (deny.toml) and advisories.

.github/workflows/live.yml, daily: the live tests and the end-to-end tests. Needs the DEEPSEEK_API_KEY repository secret.

.github/workflows/release.yml, on v* tags and manual runs: the Linux, macOS, Windows and Android builds; a tag publishes them as a GitHub release. It also runs weekly without macOS, whose runner minutes cost ten times as much. The APK gets the release key from the ANDROID_KEYSTORE (base64 of the keystore) and ANDROID_KEYSTORE_PASSWORD repository secrets.

License

MIT. The software by others in the releases, and the models the app downloads, are listed with their licenses in THIRD_PARTY_NOTICES.md. ARCHITECTURE.md describes how the code fits together, CONTRIBUTING.md how to send a change, and SECURITY.md how to report a vulnerability.