Skip to content

Releases: shawnbarnes-lab/tensorvault

Release list

0.1.0

0.1.0 Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 26 May 03:50
v0.1.0
b73a5b6

TensorVault v0.1.0 — Early Access

Private business knowledge assistant by TensorSpace LLC. Your documents indexed locally, answered by a local LLM, exported as PDF or DOCX. No cloud, no API keys, no telemetry.

⚠️ System Requirements — read before downloading

Component Minimum Recommended Optimal
OS Windows 10/11 64-bit Windows 11 64-bit Windows 11 64-bit
CPU 4 cores 8+ cores 8+ cores
RAM 8 GB 16 GB 32 GB
GPU NVIDIA 4 GB VRAM NVIDIA 8 GB VRAM NVIDIA 12 GB VRAM (RTX 3060 12GB+)
Disk 15 GB free SSD 25 GB free SSD 25 GB free SSD
Internet Required on first launch (~10 GB model download) same same

Asymmetric GPU/RAM fallback (the perk). TensorVault uses Ollama for both the LLM and embeddings. Ollama automatically splits the model between GPU VRAM and system RAM based on what is available — a 4 GB card uses all 4 GB and runs the rest on CPU. Same Gemma 4 model runs on every machine. Smaller GPUs just see slower inference because layers offload to CPU. At Optimal (12 GB VRAM), the entire LLM lives on the GPU and inference feels real-time.

Important: The installer itself is ~1.7 GB. On first launch, TensorVault downloads ~10 GB of models (Gemma 4 LLM + mxbai-embed-large embedder) via Ollama. First launch requires internet. After that, the app runs fully offline.

Installation

  1. Download TensorVault-Setup-0.1.0.exe from the Assets below.
  2. Run the installer.
  3. SmartScreen prompt — click "More info" → "Run anyway". (v0.1.0 ships unsigned; signed builds tracked for v0.1.1 via SignPath Foundation.)
  4. Launch TensorVault from the Start Menu.
  5. Wait 10-30 minutes on first launch while the LLM + embedding models download. Status bar shows live progress.
  6. Click My Docs to add your business documents.
  7. Click Ask to ask questions across your documents.

README.md and LICENSE are attached as separate files below for offline reference.

What's in v0.1.0

  • Document indexing: PDF, DOCX, TXT, MD, CSV, RTF (clean text PDFs work; OCR is on the v0.2 roadmap).
  • Semantic search with GPU-accelerated embeddings (mxbai-embed-large via Ollama, 1024-dim).
  • Grounded AI answers with inline citations, powered by Gemma 4. Auto-falls back to CPU on low-VRAM machines.
  • Business-tuned system prompt that handles lookup, summarize, compare, draft, and extract patterns.
  • PDF and DOCX export of any answer with citations.
  • Voice input via Whisper (auto-detects CUDA).
  • Fully offline after first-launch model downloads. No telemetry. No accounts.

What's NOT in v0.1.0 (planned for v0.2)

  • OCR for scanned PDFs — Tesseract not bundled. Text-based PDFs work fine.
  • Voice output / read-aloud — Piper TTS not bundled. Voice input still works.
  • Signed installer — SignPath Foundation application in progress.
  • Reranker — cross-encoder was dropped to stay under GitHub's 2 GB asset cap; v0.2 may add LLM-based reranking via Ollama.
  • Multi-user / shared corpus — single-user only for now.

Where your data lives

  • Indexed documents: %APPDATA%\TensorVault\user_docs\
  • Ollama model cache: %APPDATA%\TensorVault\ollama_models\
  • Nothing ever leaves your computer after the first-launch downloads complete.

MIT licensed. Source: https://github.com/shawnbarnes-lab/tensorvault
Published by TensorSpace LLC.

Assets 7
Loading