Skip to main content
Ask for “receipts from tax season” or “photos from the Japan trip” and FilDOS finds files by what they contain. Text, code, PDFs, DOCX, and images are all indexed on-device — nothing is uploaded, and the index lives in a local database.
Semantic search results in FilDOS

Meaning and name search, fused into one ranked result list.

​
How search works

Hit ⌘ K and type a natural-language query. FilDOS runs a hybrid search and returns a single fused list — never separate “semantic” and “name” sections:
  1. Vector retrieval finds chunks whose meaning matches your query.
  2. BM25 keyword retrieval finds exact term and filename matches.
  3. The two lanes are fused with Reciprocal Rank Fusion — filename evidence anchors the rank, the semantic score is discounted against it, and a file that scores in both lanes gets a bump.
  4. The top candidates are reranked by an on-device cross-encoder for the final ordering, and the strongest hit is badged Best.
Scope chips let you narrow to the current folder, or to Folders / Docs / Images / Audio / Video / Code / Tag, and filter by time.

​
The indexing pipeline

From documents to embeddings

Files are extracted, chunked, embedded, and stored as vectors — all locally.

1

Extract

Text is pulled from plain text, code, Markdown, CSV, JSON, and similar files. PDFs stream through a byte-range transport (book-size files extract with bounded memory), and DOCX is parsed too. Files that can’t be extracted still get a filename-fallback chunk so name queries find them.
2

Chunk

Long documents are split into ~512-token windows so context isn’t lost at boundaries.
3

Embed

Each chunk is passed through the active model — running on WASM — to produce a vector. Images are embedded with CLIP into the same space as text, so a text query can match a photo.
4

Store

Vectors are stored as Float32 BLOBs in SQLite and searched with brute-force cosine similarity over an in-memory cache — plenty fast for a personal-scale index.

​
Choosing a model

Open Settings → AI & Search to pick an embedding model and download it. Models are cached in userData/models and loaded once per session. An optional cross-encoder reranker further sharpens results; like the Canvas NER model, it’s opt-in and never auto-downloaded.

​
The background indexer

The indexer crawls your configured roots (your home directory by default) in the background. It’s designed to stay out of your way:
  • Cost-classed queue — removals and text first, images next, PDF/DOCX last — so a fresh index is useful in minutes.
  • Duty-cycled around your activity: near full speed when you’re away, gentler while you’re working, and easier on battery.
  • Skips the noise — dotfiles, system folders, node_modules, build outputs, and caches. Checked-out codebases are indexed docs-only so source files don’t drown your results.
  • Keeps itself fresh via a filesystem watcher plus a periodic reconcile.
Progress lives in Settings → Indexing. You can pause, resume, or clear the index at any time.

​
Keeping things private

Use Hide from AI (a context-menu action, or Settings → Privacy) to exclude specific files, folders, or whole file extensions from indexing — the assistant and search never see what you don’t want them to.

​
Next

Ask AI

Turn search into a conversation with an on-device LLM.

AI Pipeline

The engineering behind extraction, embeddings, and fusion.