Home

SenseTree Wiki

SenseTree is a local-first semantic overlay for your Windows file system: search your files by meaning, chat with a folder through an agent that can read and search them, and let it propose reorganizations you approve — with the AI models you choose, on the machines you choose.

This wiki is the complete documentation.

New here?

• Installation — download, install, and automatic updates.
• Getting Started — index your first folder, run your first search.
• FAQ — privacy, data location, updates, MSI vs EXE.

Using SenseTree

• Configuration — every setting, field by field, and where settings.json lives.
• Models & Providers — the five model slots, the live catalog, hardware picks.
• Semantic Search — searching by meaning, scoping, the meaning tree.
• Image Search — finding pictures by what they look like (CLIP).
• AI Chat & Agent — the agent, its tools, and the Dry-Run approval model.
• Gardener — auditing a tree for duplicates, clutter and empty folders.
• Prompts — retune every AI stage without recompiling.
• MCP Servers — plug external tools into the agent.

Under the hood

• Embeddings — the engines, the models, and what text actually gets vectorized.
• Retrieval & RAG — hybrid search, rank fusion, cross-encoder reranking.
• Indexing Pipeline — crawler, watchdog, worker, and every AI stage.
• Media: Audio & Video — transcription and visual description.
• AI Server Protocol — every HTTP request SenseTree makes, with its exact payload.
• Architecture — module map, data stores, IPC surface, concurrency.
• Building from Source — dev setup and the release process.
• Troubleshooting — the issues you're most likely to hit.

---

What makes it different

Real files, not a vault	An overlay on your actual folders. Nothing is hidden or relocated behind your back.
Meaning and keywords	Dense vectors find the idea; BM25 finds the exact serial number. Fused, then reranked by a cross-encoder.
Every file gets a sense	Documents are extracted, scans are OCR'd, images captioned, media transcribed, and unreadable files described from their context. Nothing is left unindexed.
You propose, I plan	Structural changes always arrive as an interactive Before → After diff. Apply is transactional with rollback; deletes go to a local trash.
Bring your own models	Five independent slots — local engine, Ollama, LM Studio, vLLM, a LAN box, or an API.
Private by construction	Fully offline by default. Embeddings computed in-process; LLM calls go only to endpoints you configure. No telemetry.

At a glance

• Hybrid RAG — dense + BM25, Reciprocal Rank Fusion, cross-encoder reranking, contextual retrieval at index time.
• Agentic chat — a ReAct loop with six built-in tools (search, read, list, read senses, propose actions, remember), plus any MCP server you add, and durable memory across conversations.
• Multimodal indexing — vision captions, rendered-page OCR for scanned PDFs, speech transcription, and video description.
• Live model catalogs — MTEB, OpenCompass and the Ollama library, with quantization picking and VRAM fit.
• Signed auto-update — the app checks, downloads and installs new versions itself.

Current version: 3.9.0. Installers and changelog on the Releases page.