DocLLM

Chat with your documents using AI that runs entirely on your Mac.

No account. No cloud. No data ever leaves your computer.

getdocllm.com

DocLLM is a free, open-source macOS desktop app that lets you have AI conversations with your documents — PDFs, Word files, Markdown, and plain text. Every AI model runs locally on your Mac via Ollama. Your documents are never uploaded anywhere.

• Chat with any document — PDF, Word (.docx), plain text, Markdown
• Page-level citations — every answer cites the exact page it came from; click to jump there in the built-in PDF viewer
• Multi-document workspaces — group related files; ask questions across all of them at once
• OCR for scanned PDFs — built-in Tesseract OCR handles documents with no text layer

• Resizable split-pane layout — sidebar, PDF viewer, and chat all freely resizable
• Pinned notes — save important answers per workspace
• Full-text history search — search across all past conversations (⌘F)
• Export conversations — save chat as .txt or formatted HTML

• Streaming responses — tokens appear in real time from Ollama
• Parallel embedding — chunks embedded in batches of 3; reduces ingest time by ~3×
• Model pre-warming — chat model loads into memory on startup for a fast first response
• ETA progress bar — shows estimated time remaining during document indexing

• Zero telemetry — no analytics, no crash reporting, nothing phoned home
• No accounts — no login, no email, no sign-up
• Single outbound connection — only talks to localhost:11434 (your local Ollama)

You drop a document
        │
        ▼
  Extract text                  (pdfjs-dist for PDFs, Tesseract OCR fallback)
        │
        ▼
  Split into chunks             (400-word overlapping windows)
        │
        ▼
  Embed each chunk              (Ollama — nomic-embed-text, batched × 3)
        │
        ▼
  Store embeddings              (in-memory + localStorage, no database)

You ask a question
        │
        ▼
  Embed question                (same embedding model)
        │
        ▼
  Cosine similarity search      (across every chunk in the workspace)
        │
        ▼
  Top-6 chunks as context       (threshold 0.3, globally re-ranked)
        │
        ▼
  Stream answer from Ollama     (mistral:7b, with page citations in the reply)

• macOS (Apple Silicon or Intel)
• Ollama — free, installs in 30 seconds
• ~4.5 GB free disk space (for AI models, downloaded once)
• 8 GB RAM minimum · 16 GB recommended for large documents

1. Go to the Releases page
2. Download the latest .dmg
3. Drag DocLLM to Applications
4. Launch it — the setup wizard guides you through everything

Prerequisites: Node.js 22+ · Rust · Ollama

# 1. Clone
git clone https://github.com/yourusername/docllm.git
cd docllm

# 2. Install dependencies
nvm use 22
npm install

# 3. Run in development
export PATH="$HOME/.cargo/bin:$PATH"
npm run tauri dev

# 4. Or build a distributable .dmg
npm run tauri build

The setup wizard runs once on first launch:

After setup, the app always opens straight to the chat interface.

• Drag and drop a file onto the app window
• ⌘ O — open file picker
• + button in any workspace

• Type your question and press Enter
• Click any source badge in the answer to jump to that exact page in the PDF viewer
• ⌘ K — focus the chat input from anywhere

• ⌘ N — new workspace
• Add multiple documents to a workspace and ask questions across all of them simultaneously
• Each workspace has its own independent chat history and pinned notes

docchat/
├── src/
│   ├── App.tsx                        # loading → onboarding → app state machine
│   ├── components/
│   │   ├── MainLayout.tsx             # Full layout + ingest pipeline
│   │   ├── Sidebar.tsx                # Workspace list + file upload
│   │   ├── ChatWindow.tsx             # Message thread with markdown + citations
│   │   ├── PdfViewer.tsx              # In-app PDF renderer (pdfjs, canvas)
│   │   ├── DropZone.tsx               # Drag-and-drop file target
│   │   ├── SettingsDrawer.tsx         # Model + Ollama URL settings
│   │   ├── Dashboard.tsx              # Usage stats modal
│   │   ├── NotesPanel.tsx             # Pinned notes per workspace
│   │   ├── HistorySearch.tsx          # Full-text conversation search
│   │   ├── ExportMenu.tsx             # Export chat as TXT / HTML
│   │   └── onboarding/                # 4-step setup wizard
│   ├── lib/
│   │   ├── pdf.ts                     # PDF text extraction + Tesseract OCR
│   │   ├── ollama.ts                  # Ollama REST client (health, pull, embed, chat)
│   │   ├── rag.ts                     # Embed → search → stream pipeline
│   │   ├── chunker.ts                 # 400-word overlapping chunker
│   │   ├── extractor.ts               # Routes files to correct extractor
│   │   └── vectorStore.ts             # Cosine similarity + localStorage persistence
│   └── store/
│       ├── appStore.ts                # Docs, workspaces, Ollama state
│       ├── settingsStore.ts           # Model selection, Ollama URL
│       └── usageStore.ts              # Per-workspace notes + usage counters
├── src-tauri/
│   ├── src/lib.rs                     # Rust: get_ram_gb command
│   ├── capabilities/default.json      # Tauri permissions
│   └── tauri.conf.json                # Window, CSP, bundle config
└── public/                            # Runtime deps (DO NOT DELETE)
    ├── pdf.worker.min.mjs             # pdfjs web worker
    └── tesseract/                     # Tesseract WASM engine + OCR model

All settings are in the gear icon in the sidebar. They persist in localStorage.

The model dropdowns are populated live from Ollama — only models you've already pulled appear.

• No network requests except to localhost:11434 (your Ollama, running locally)
• No telemetry — no analytics, no error reporting, nothing is tracked
• No accounts — no login, no email address required
• All data is yours — document text, embeddings, and chat history are stored in your browser's localStorage and Tauri's app data directory at ~/Library/Application Support/com.local.docllm/

• Chunk size: 400 words with 80-word overlap (sentence-aware splitting)
• Similarity threshold: 0.3 — chunks below this are dropped even if they rank in top-K
• Top-K per doc: 8 candidates per document, globally re-ranked, top 6 sent to the LLM
• Zero-chunk fallback: if no chunks pass the threshold, the LLM answers from workspace metadata only
• Multi-doc citations: source badges show document name + page when workspace has multiple files

npm run tauri dev        # hot-reload dev mode
npm run build            # frontend only (tsc + vite)
npm run tauri build      # production app bundle
npx tsc --noEmit         # type check only
npm test                 # run test suite
npm run test:watch       # watch mode
npm run test:coverage    # with V8 coverage

Tests live in src/test/ and use Vitest + Testing Library + jsdom. See CLAUDE.md for architecture details and Tauri-specific constraints.

All contributions welcome — bug fixes, new file format support, Windows/Linux ports, UI improvements.

1. Fork the repo
2. Create a branch: git checkout -b feature/your-feature
3. Make your changes
4. Type check: npx tsc --noEmit
5. Run tests: npm test
6. Open a pull request

• Windows support
• Linux support
• In-app model manager (download/switch models without a terminal)
• Image and diagram understanding
• Custom system prompt per workspace
• More embedding model options

MIT — free to use, modify, and distribute.