A local RAG (Retrieval Augmented Generation) app for Windows, macOS, and Linux. It indexes your notes, books, and code into one SQLite database with hybrid vector + full-text search (Reciprocal Rank Fusion), so you can find things by meaning, not just exact words. Keyboard-first, and everything runs on your machine.
Inspired by Sebastian Hutter’s local-rag. No Ollama, no API keys: the embedding model runs in-process from a .gguf file you import or download from the built-in catalog.
- Private: indexing, parsing, embedding, search, and OCR run inside the app. Your documents and their text are never uploaded. See network activity for the three requests the app does make.
- Search by meaning: a blue-green deploy note matches “how do we ship changes safely” without using those words.
- Index what you have: Obsidian vaults, document folders (Markdown, PDF, DOCX, HTML, TXT, CSV, JSON, YAML, XML, SQL, shell scripts, XLSX, PPTX, Jupyter notebooks, EPUB), Calibre libraries, and code repositories with commit history.
- Bring your own model: import or download a
.gguffrom Settings. - Scanned PDFs: install the Tesseract OCR plugin in Settings; photo pages become searchable.
- Manage the library: browse files and chunks, delete a source, selected chunks, or a whole library.
- MCP server: expose search, read, and collection tools to any local MCP client.
Windows 10/11 · Linux (AppImage, deb) · macOS (universal) · not code-signed, see first-run notes
Normal Windows builds need AVX2; PCs from before ~2013 will not start. Use the legacy installer (or legacy portable zip) and pick BGE Small EN v1.5 Q8_0 in Settings. Other quantized models are much slower without AVX2.
Build from source or package a release: CONTRIBUTING.md.
Not code-signed, so your OS may warn on first launch.
Windows - SmartScreen
- Click More info
- Click Run anyway
Or right-click the .exe -> Properties -> check Unblock -> Apply.
macOS - Gatekeeper
Right-click the app -> Open (once), or run xattr -d com.apple.quarantine /path/to/vectile.app.
Linux
The .deb pulls in GTK4, WebKitGTK 6.0, and libgomp1. The AppImage needs chmod +x.
| Source | Collection Type | What Gets Indexed |
|---|---|---|
| Obsidian | system | Vault files: .md notes with frontmatter, tags, and wikilinks |
| Project folders | project | Any folder of documents, each file parsed by its extension (.md, .pdf, .docx, .html, .txt, .csv, .json, .yaml, .xml, .sql, .sh, .xlsx, .pptx, .ipynb, .epub) |
| Code repositories | code | Git repos: tree-sitter splits each function and class into its own chunk (cAST split-then-merge); commit history is indexed as its own source |
| Calibre | system | Ebook metadata + content: title, author, tags, series, publisher, description, and EPUB/PDF text |
- Launch vectile.
- Get an embedding model: import one in Settings (copied to
models/), download one from Settings → Model → Get a model, or drop a.ggufintomodels/yourself. - Add sources in Settings: Obsidian vaults, project folders, code repositories, Calibre libraries.
- Open the Index view and index a collection, or everything at once. Unchanged files are skipped.
- Press ⌘K / Ctrl K and search.
Deleted files drop out of the index automatically. Auto-reindex rebuilds on a timer; start-on-login opens the app with your session.
Settings → Connect starts a local MCP (Model Context Protocol) server at 127.0.0.1:31123; it accepts local connections only.
Search, read, and list tools are always on; index and prune need Allow write tools.
Search returns a snippet and a chunk id, not a whole passage. The assistant pulls more with vectile_get_chunk or vectile_read_source, finds exact text with vectile_grep, and lists collection contents, so your whole library is never loaded.
Point any MCP client (Claude Desktop, Claude Code, …) at http://127.0.0.1:31123/mcp. Streamable HTTP is default; switch to SSE in Settings if a client needs the older transport. Settings shows server status, offered tools, and per-client setup.
Two searches run per query.
- Full-text search matches exact words in an FTS5 index. Fast and literal.
- Vector search embeds the query and finds passages whose vectors point the same way, so "how do we ship changes safely" matches a blue-green deploy note.
A fast index picks candidates, then full vectors are compared and re-sorted by distance. Both result lists merge by rank, not score: Reciprocal Rank Fusion.
Filters: collection, source type, path, sender/author, date range. Top-k sets the result count.
Query embeddings are cached per model, so repeat searches skip the model; ranking is always fresh. The cache clears on reindex, prune, or model switch. Settings → Cache inspects and clears it.
config.json, in your per-user config folder:
- Windows:
%AppData%\vectile\config.json - macOS:
~/Library/Application Support/vectile/config.json - Linux:
~/.config/vectile/config.json
| Key | Default | Description |
|---|---|---|
| embedding_model | bge-m3 | Embedding model name |
| active_model | (default model path) | Path to the active .gguf model |
| embedding_batch_size | 32 | Chunks per embedding call |
| chunk_size_tokens | 500 | Chunk size in whitespace-separated words |
| chunk_overlap_tokens | 50 | Overlap between chunks |
| obsidian_vaults | [] | Paths to Obsidian vaults |
| obsidian_exclude_folders | [] | Names to skip in vaults |
| project_exclude_folders | [node_modules] | Names to skip anywhere inside project folders |
| repository_exclude_folders | [] | Names to skip anywhere inside repositories |
| calibre_exclude_folders | [] | Names to skip inside Calibre libraries |
| calibre_libraries | [] | Paths to Calibre libraries |
| repositories | {} | Collection name to repo or directory paths; directories are scanned recursively for git repos |
| projects | {} | Collection name to document paths |
| disabled_collections | [] | Collection names to skip during indexing |
| skip_cloud_placeholders | true | Skip cloud-only placeholders (OneDrive, iCloud, Google, Synology), not download them |
| git_history_in_months | 6 | Months of commit history to index |
| git_commit_subject_blacklist | [] | Skip commits whose subject starts with any of these strings |
| search_defaults.top_k | 10 | Default number of search results |
| search_defaults.rrf_k | 60 | Reciprocal Rank Fusion parameter |
| search_defaults.vector_weight | 0.7 | Weight for vector similarity |
| search_defaults.fts_weight | 0.3 | Weight for full-text search |
| gui.auto_reindex | false | Enable periodic re-indexing |
| gui.auto_reindex_interval_minutes | 60 | Minutes between auto-reindex runs |
| gui.start_on_login | false | Launch at login |
| gui.mascot.show_idle | true | Show Vexter at rest between runs |
| gui.mascot.show_searching | true | Show Vexter in the sidebar while a query runs |
| gui.mascot.show_indexing | true | Show Vexter while a library rebuilds |
| mcp.enabled | false | Serve MCP tools to local assistants on launch |
| mcp.port | 31123 | Port the MCP server listens on (127.0.0.1 only) |
| mcp.allow_write | false | Let assistants call the index and prune tools |
| mcp.transport | streamable-http | MCP transport: streamable-http (/mcp) or sse (/sse) |
| ocr.enabled | true | Run OCR on PDF pages that come back with no text |
| ocr.languages | [eng] | Language codes; codes with no matching data are ignored |
Parsing, chunking, embedding, search, and OCR run in-process. Document text, chunks, vectors, and library metadata never leave the machine. No server, account, or cloud sync.
Three outbound requests, and only these:
| When | What | Host |
|---|---|---|
| Every launch | The latest release tag, compared with your version to show an update badge. Sends no identifiers. No setting turns it off; blocking the host just means no update notice. | api.github.com |
| You press download in Settings → Model → Get a model | The .gguf you picked |
huggingface.co (file bytes from its CDN, *.cdn.hf.co) |
| You install the Tesseract OCR plugin in Settings | The platform archive and the release SHA256SUMS it is verified against |
github.com (assets from release-assets.githubusercontent.com) |
Both downloads happen only when you ask: import a model you already have, or skip the OCR plugin, and nothing is fetched. Links open in your default browser, not in-app. Settings → Connect starts a loopback-only MCP server, off by default.
No analytics, telemetry, crash reporting, or background uploads. The privacy page has the same detail.
| Component | Choice | Notes |
|---|---|---|
| Language | Go 1.26+ | Wails v3 desktop app |
| UI | SolidJS + TypeScript + Vite | Tailwind CSS v4 |
| Database | SQLite (modernc.org/sqlite) + sqlite-vec + FTS5 | Pure Go, no cgo; single file |
| Embeddings | llama.go (llama.cpp) | In-process .gguf; bge-m3 by default; no Ollama |
| Code parsing | go-tree-sitter | cAST split-then-merge; one chunk per function, class, or method |
| go-pdfium (WASM/Wazero) | No cgo needed | |
| DOCX | archive/zip + encoding/xml | Word document extraction (.docx, .dotx) |
| XLSX | excelize | Spreadsheets, one ## Sheet section per worksheet |
| PPTX | archive/zip + encoding/xml | Slides, one ## Slide section per slide |
Apache License 2.0, copyright (c) 2026 d3uceY. See LICENSE.
Bundled third-party code keeps its own license: vendored third_party/llama-go and the llama.cpp
sources it wraps are MIT.