Changelog

​

FilDOS 1.0.0-beta.1

​

File manager

• Three views — virtualized list (resizable, reorderable columns), grid (cached thumbnails at several icon sizes), and gallery.
• Full operations — create, rename, copy, cut, paste, duplicate, and trash, with inline rename and an undo stack for every mutation.
• Drag-and-drop — internal move/copy, plus OS drag-in and drag-out to Finder / Explorer.
• Live filesystem watching — external changes appear in under a second.
• Per-folder view memory — sort, view mode, and icon size are remembered per folder.
• Delete = OS Trash — deletions go to the system Trash / Recycle Bin; FilDOS keeps no trash of its own.

​

Tags & recents

• Colour-coded tags in six brand accents, stored in SQLite.
• Tag Files view — every file carrying a tag, wherever it lives.
• Persistent recents across sessions.
• Tags, recents, and views follow files through renames and moves.

​

Search by meaning

• On-device semantic search over text, code, PDFs, DOCX, and images.
• Hybrid ranking — vector + BM25 fused with Reciprocal Rank Fusion, then an on-device cross-encoder rerank.
• Image search via CLIP — text and images embedded into one space.
• Scope chips — Everywhere / This folder / Folders / Docs / Images / Audio / Video / Code / Tag, plus time filters.
• Embedding model catalog — MiniLM, BGE, GTE, multilingual E5, and CLIP.

​

Background indexer

• Cost-classed, duty-cycled pipeline that stays out of your way and becomes useful within minutes.
• Byte-range PDF transport for book-size documents with bounded memory.
• Codebase-aware indexing (docs-only for checked-out repos).
• Ambient mode — keeps indexing from the menu bar / system tray after the last window closes.
• Skips dotfiles, system folders, node_modules, caches, and build output.

​

Ask AI

• Chat with your files through a fully on-device LLM (node-llama-cpp).
• @file / #folder mentions and /commands (e.g. /find).
• File actions — create, copy, move, rename, list, read, delete-to-Trash — via function calling, with your approval and full recoverability.
• Bring your own model — a catalog from 0.6B to 35B (Llama, Qwen, Gemma, Phi, Mistral, DeepSeek, and more), or paste any GGUF from Hugging Face.
• Saved conversations you can reopen and continue.

​

Canvas

• A GPU-rendered knowledge graph of your files built from embedding similarity, extracted entities (opt-in NER), tags, and temporal sessions.
• Louvain communities, a detail panel that explains connections, and an mtime time scrubber with replay.

​

Cloud drives

• Browse Google Drive and Dropbox alongside local folders through the same interface (built on opendal).

​

Privacy

• Hide from AI — exclude specific files, folders, or extensions from indexing and chat.
• Everything is local — embeddings, search, the graph, and chat all run on-device; nothing leaves your machine unless you connect a cloud drive.

​

Foundations

• Electron + React + TypeScript with a strict process boundary and a Result-typed IPC contract.
• SQLite via Node’s built-in node:sqlite + Drizzle ORM — zero native dependencies for the data layer.
• Tested with Vitest (unit + integration) and Playwright (e2e) on Ubuntu, macOS, and Windows.