Knowledge AI for Obsidian

简体中文 · English

An AI assistant for your Obsidian vault: ask questions and get cited answers grounded in your notes, generate study guides / timelines / mind maps / slide decks from any folder.

Inspired by Google NotebookLM but fully local-first and provider-agnostic — bring your own API key, or run embeddings entirely on-device.

Status: pre-1.0. Public release of a previously private project; expect rough edges. Issues and PRs welcome.

Features

• Grounded RAG chat with inline [N] citations that jump straight to the source chunk in your vault.
• Hybrid retrieval — BM25 (Chinese-tokenized via segmentWords) + vector embeddings + document-level summary search, merged with Reciprocal Rank Fusion.
• Multi-format indexing — Markdown, PDF (text layer), DOCX, XLSX, PPTX. Extraction runs in a Web Worker so the UI stays responsive on large vaults.
• Image understanding — Tesseract OCR (chi_sim + eng by default) and optional LLM vision for image-bearing notes.
• Structured artifact generation — summary, study guide, timeline, FAQ, briefing, mind map, and slide deck (PowerPoint export).
• Provider-agnostic — any OpenAI-compatible endpoint (DeepSeek, Moonshot/Kimi, GLM, Qwen, OpenAI, …). Local embeddings via @xenova/transformers (multilingual-e5-small by default).
• Save chat answers as notes — one click turns a cited answer into a permanent vault note.
• Internationalization — Chinese (zh-CN) and English (en) UI; auto-detects Obsidian's language.

Screenshots

Screenshots coming with v0.4.0.

Installation

From Obsidian Community Plugins

Passed Obsidian's automated review; the listing is live at community.obsidian.md. In-app availability follows once the community registry syncs. Once it appears in the in-app browser:

1. Open Obsidian → Settings → Community plugins → Browse.
2. Search "Knowledge AI" and install.
3. Enable the plugin.

Manual install (current method)

1. Download main.js, manifest.json, styles.css, and the two ort-wasm*.wasm files from the latest release.
2. Copy them to <your-vault>/.obsidian/plugins/notebook-ai/.
3. Reload Obsidian and enable Knowledge AI under Community plugins.

The plugin is desktop-only (mobile is out of scope due to embedding worker + WASM requirements).

Quick start

1. Add a provider — Settings tab → Providers → Add. Paste your OpenAI-compatible base URL and API key. For DeepSeek use https://api.deepseek.com/v1; for Moonshot use https://api.moonshot.cn/v1; for OpenAI use https://api.openai.com/v1.
2. Assign tasks — Settings → Task assignment. Pick a chat model, optionally an embedding model, optionally a summary model. Tasks default to your first provider's default model.
3. (Optional) Enable vector retrieval — Settings → Vector retrieval. Toggle on; pick local (default: multilingual-e5-small, downloaded on first use, ~110 MB) or external API embeddings.
4. Create a Notebook — Settings → Notebooks → Add, pick a folder. Or right-click any folder in the file explorer → "Create Notebook from this folder".
5. Chat — Click the ribbon icon (or open command palette → "Open Notebook AI chat"). Ask questions, follow citations to the source.
6. Generate artifacts — Switch to the "Artifacts" tab in the chat view; pick a kind (summary / study guide / timeline / …) → it streams.

See docs/USAGE.md for advanced topics (query expansion, summary backfill, OCR, custom system prompts, multi-folder notebooks).

Configuration tips

• Large vaults (> 2 000 documents) — turn on document-level summaries: open a notebook card → click "Backfill summaries". This precomputes a per-document summary so the retriever can pick the right document first, then dig into chunks. Default concurrency is 1 (kind to rate-limited Chinese providers); raise to 3-4 for DeepSeek/OpenAI.
• Chinese content — leave the default tokenizer alone; it falls back to character n-grams for CJK so BM25 still works without a Chinese segmenter dependency.
• Cost control — vector retrieval is local by default. The only API calls are: chat completions, optional reranking, optional document-summary generation, and optional API embeddings.

Privacy and data

• No telemetry. The plugin never phones home.
• All indexes and chunks live under <vault>/.obsidian/plugins/knowledge-ai/data/ (JSONL files; compacted on plugin reload).
• API requests are sent directly to the provider you configure. Your API key is stored in the same plugin folder.

Development

npm install
npm run dev        # incremental build (watch)
npm run build      # production build
npm test           # vitest
npm run type-check # tsc --noEmit

To test the build against a real vault:

node scripts/deploy.mjs <path-to-vault>

See CONTRIBUTING.md for the contribution workflow.

Acknowledgements

• Obsidian — the platform.
• @xenova/transformers and ONNX Runtime Web — on-device embeddings.
• pdfjs-dist, mammoth, xlsx, @xmldom/xmldom — file extraction.
• MiniSearch — BM25 retrieval.
• Tesseract.js — OCR.
• Google NotebookLM — UX inspiration.

License

MIT — see file for full text.