FAQ

FAQ

Is my data private?

Yes, by construction. Embeddings are computed on your machine (or on the server you chose). LLM calls go only to the endpoints you configure — localhost by default. No telemetry, no background uploads, no account.

The one way file content leaves your device is if you point a slot at an external cloud API.

AI Server Protocol → What leaves the machine lists every outbound request the app can make, including model downloads and catalog metadata.

Where is my data stored?

Under %APPDATA%\com.virgi.sensetree:

• settings.json — configuration (including any API keys, in plain text),
• a SQLite database — file catalog, indexing queue, extracted senses, folder profiles, agent memory, transaction log,
• a LanceDB directory — the vectors,
• models\ — downloaded local models and the ONNX runtime,
• trash\ — files deleted through an approved plan.

Deleting that folder resets SenseTree completely.

Does SenseTree move or modify my files?

Never without your explicit approval. It reads files to index them. Any move, rename or delete arrives as a Dry-Run plan you Approve or Discard, apply is transactional with rollback, and deletes go to a local trash rather than being destroyed. See AI Chat & Agent.

Will it update itself?

Yes. Since v2.0.0 the app checks for new releases at startup and offers to download and install them, verifying a cryptographic signature first. Manual reinstall still works too. See Installation.

What's the difference between the .exe and .msi?

Both install the same app. -setup.exe (NSIS) is a lightweight per-user wizard needing no admin — recommended for individuals, and the target of auto-update. .msi is enterprise packaging for mass deployment (GPO/Intune/SCCM), silent install, usually per-machine. See Installation.

Do I need a GPU?

No, but it helps. Embedding runs fine on CPU — that's the default. The LLM slots are much happier on a GPU. Per-hardware picks in Models & Providers.

Do I need Ollama or LM Studio?

Only for reasoning and vision. Semantic search over your documents works with the built-in local engine, no extra install. To chat, plan reorganizations or caption images, point those slots at a runner. Transcription and video description need their own server (speaches, whisper.cpp, vLLM…) — Ollama does not do transcription.

Why did indexing restart after I changed a setting?

Changing the embedding model or its dimensions invalidates every vector, so a full re-index is unavoidable. Changing the block/recursive slider or the folder_classify prompt makes SenseTree forget prior folder classifications, which a scan then redoes.

Everything else — reasoning, vision, media, prompts, retrieval, MCP, API keys — takes effect immediately with no re-index. See Configuration.

Which folders should I not index?

You don't have to micromanage it. The block/recursive classifier treats venv, node_modules, app bundles, DAW packs, caches and build artifacts as opaque blocks automatically, and AppData, Windows, $RECYCLE.BIN, __pycache__ and dotfolders are never touched. Tune the aggressiveness with the slider, or force a folder's mode by hand.

Can the AI read files outside the folders I indexed?

No. The agent's file tools (read_file, list_directory, read_semantics) and every path in a proposed plan are restricted to your indexed roots, with segment-boundary matching so a root named Docs doesn't authorise DocsEvil.

The exception is MCP servers you add yourself — those run their own code and are not bounded by SenseTree.

Why is a file missing from search results?

Usually: it isn't indexed yet, it's inside a block folder, or it was processed while the embedding endpoint was down (vectors are only written after a successful embedding). Retrieval & RAG walks through the causes.

Can it read scanned PDFs? Audio? Video?

Yes, all three, if the corresponding slot is enabled. Scanned PDFs have their pages rendered to images and read by the vision model. Audio and video are transcribed, and videos can additionally be described visually — both then chunked like any document, so a one-hour recording is searchable on any of its passages. See Media: Audio & Video.

Is it cross-platform?

Builds target Windows. The stack (Tauri + Rust + fastembed/LanceDB) is portable in principle, but Windows is what's built and tested today, and several paths (ONNX Runtime provisioning, path handling, the cmd /C MCP launcher) are Windows-specific.

How do I report a bug or request a feature?

Open an issue or a PR. Include your OS, the version (shown in Settings), and relevant logs — run the app from a terminal to see its tracing output.