Getting Started

Getting Started

From a fresh install to your first semantic search and your first Dry-Run action.

The layout

Three panels:

• Sidebar (left) — indexed roots with gardener health badges, indexing progress and pause, the indexing queue, per-stage throughput, image search, the AI status panel, and Settings.
• Explorer (center) — breadcrumb and file list of the current folder, with each file's extracted sense and index status. Switches to search results when you search, and offers a meaning-tree view.
• Chat (right) — the agent, scoped to the current folder. Proposed file operations render as a Before → After card.

The AI status panel shows one indicator per server slot — embedding, reasoning, vision, transcription, video description — so you can see at a glance which one isn't answering.

Step 1 — Add a folder

In the sidebar, click add a folder and pick a root (e.g. Documents). SenseTree immediately starts:

• a crawler that walks the tree (the past), and
• a watchdog that reacts to live changes (the present).

Add as many roots as you like; remove any of them later. Technical folders — venv, node_modules, app bundles, DAW sample packs — are detected and indexed as single opaque blocks, so you don't have to curate. See Indexing Pipeline.

Step 2 — Choose your models

Open Settings. Only the first slot is required.

Slot	Gives you	Default
Embedding	Search itself	Built-in local multilingual-e5-small — works out of the box
Reasoning	Chat agent, action plans, file qualification	llama3.1:8b on Ollama, enabled
Vision	Image captions, scanned-PDF OCR	moondream, disabled
Transcription	Speech in audio/video	disabled
Video	What a video shows	disabled

For the LLM slots, point base_url at your Ollama or LM Studio server, or open the catalog to browse live benchmarks, pick a quantization that fits your VRAM, and download in one click. Test connection verifies each endpoint, and the AI status panel keeps a live indicator per slot.

Changing the embedding model — or its dimensions — triggers a full re-index, because vectors from different models aren't comparable. Decide this one early. Everything else can be changed at any time with no re-index.

Step 3 — Let it index

Watch progress in the sidebar, or open the queue modal to see the current file, the AI stages it's traversing, what's pending and what failed.

Every file ends up with a sense — a couple of sentences saying what it is — and a searchable extract:

• Documents — real text; scanned PDFs get their pages rendered and OCR'd by the vision model.
• Images — a caption from the vision model, then qualified.
• Audio / video — transcribed and/or visually described, then chunked like any document.
• Anything unreadable — described from its name, folder, neighbours and the model's best guess.

You can pause at any time: the queue freezes, the local model is unloaded, and it resumes where it left off.

Step 4 — Search by meaning

Type a concept — "insurance renewal", "trip to Korea", "budget spreadsheet". Results are ranked, snippeted, and scopeable to the current folder.

Search is hybrid: meaning and exact words, so an IBAN, a serial number or a surname works as well as a description. See Semantic Search.

Two more views: the meaning tree (which branch of my tree is about this?) and image search (find pictures by what they look like — click Index images once first).

Step 5 — Chat, and safely reorganize

Open the chat and ask a question about the folder. The agent searches, reads files, cross-checks, and answers with clickable citations — you see each tool call as it happens.

Ask for an action instead — "group these invoices by year" — and you get a Dry-Run plan: a Before → After list you can review operation by operation, uncheck what you don't want, then Approve or Discard.

Nothing touches disk until you approve. Apply is transactional: if any operation fails, everything already done is rolled back. Deletes go to a local trash, not oblivion. See AI Chat & Agent.

Step 6 — Fix what it got wrong

Open a file's detail drawer to see its sense next to its extract. If the sense is wrong, you can rewrite it — or ask the agent to ("the descriptions in this folder are wrong, fix them"), which produces a Dry-Run plan of corrections.

A corrected sense is pinned: re-indexing never regenerates over it, and the file is re-embedded so the fix takes effect in search too, not just on screen.

Where's my data?

Everything under %APPDATA%\com.virgi.sensetree:

• settings.json — your configuration,
• a SQLite database — file catalog, queue, extracted senses, folder profiles, agent memory, transaction log,
• a LanceDB directory — the vectors,
• models\ — downloaded local models (embedding, reranker, CLIP) and the ONNX runtime,
• trash\ — files deleted through an approved plan.

Nothing is uploaded anywhere. See the FAQ and AI Server Protocol for the exhaustive list of outbound traffic.