Vector Vault 0.1.0
Vector Vault 0.1.0
A local-first desktop app that reads your documents, files them where they belong, and lets you find anything — documents and photos — by describing it in plain English.
Everything runs on your machine. The language model, the embeddings, your files: nothing is uploaded anywhere.
What it does
- 📂 Drop a folder into the chat — it reads every file, builds a knowledge base, and proposes where each one belongs, with a reason. Nothing moves until you click Apply, and every move can be undone.
- 💬 Ask your documents anything — answers cite the exact file and page. If your library doesn't cover it, it says so instead of making something up.
- 🖼️ Find photos by describing them — "a man lifting a baby", "sunset on a beach". No tags, no filenames; it looks at the pictures.
- 🗂️ Works while you're not looking — lives in the system tray, watches your Downloads folder, and files new documents automatically. Closing the window doesn't stop it; quit from the tray icon.
Install
- Download VectorVault-0.1.0-win64.msi below and run it.
- Windows SmartScreen will warn you — the installer is not code-signed yet. Click More info → Run anyway.
- On first launch, a setup screen detects your GPU and RAM, recommends a language model that actually fits your hardware, installs Ollama, and downloads the model with a progress bar. Internet is needed for this one step; after that the app is fully offline.
Requirements: Windows 10/11 (64-bit). ~1.2 GB for the app plus 1–9 GB for the language model depending on what your hardware supports. A GPU with ≥4 GB VRAM is recommended but not required.
Good to know
- The academic classifier ships with a sample B.Tech CSE curriculum. Edit
curriculum.yaml(see the README) to teach it your own subjects — it rebuilds its knowledge base automatically. - Scanned/photographed PDFs are detected and routed to the Review queue, but OCR isn't built yet — that's the next milestone.
- The AI never executes commands against your files. It proposes a plan; a logged, reversible engine applies it. Classification thresholds were calibrated by measurement, not guesswork.
Under the hood
Ollama (qwen2.5, hardware-matched) · bge-small embeddings · CLIP image search · SQLite + write-ahead move log · 114 tests
SHA256 (VectorVault-0.1.0-win64.msi):
A80626AC2896215F422F06087F49540CA2430B849FDF78B198F0B1B311B0D97C