Project
LiveUltraInstinct
A self-hostable second brain — retrieval-augmented AI chat over your own knowledge, hybrid semantic search, real-time streaming responses, and a Chrome extension for one-click capture from any page.
UltraInstinct is a self-hostable personal knowledge management system — a “second brain” that is genuinely searchable. Where most PKM tools give you a folder of files and a search box, UltraInstinct wraps your notes in a retrieval layer that can actually answer questions across everything you’ve saved, citing the sources it drew from.
The pitch: your notes are useless until they’re reachable. I’d accumulated thousands of notes across a handful of tools, and the honest truth was that I never found anything unless I remembered roughly when I wrote it.
Context
The problem with notes is retrieval. Writing them is easy; finding them again is not. Keyword search fails the moment you don’t remember the exact phrasing, and folder hierarchies assume you knew how you’d want to find something before you wrote it. Retrieval-augmented chat solves this the way a librarian does: by reading everything once, indexing it, and letting you ask questions instead of guessing at filenames.
Architecture
The system is a full-stack RAG application built around ChromaDB vector indexing and PostgreSQL:
- Hybrid semantic search — vector similarity over ChromaDB combined with structured queries over PostgreSQL, so “the email I sent about the YC application” works whether you remember the concept or the exact wording.
- Retrieval AI chat — streaming answers generated from the retrieved context, with real-time SSE streaming so responses feel immediate rather than processed.
- One-click capture — a Chrome extension that captures any webpage into the system without leaving the page you’re on.
- A self-hostable core — everything runs on infrastructure you control, from the database to the vector store (pgvector alongside ChromaDB).
Engineering details
The unglamorous parts took most of the effort: soft-delete so nothing is ever irrecoverably gone, duplicate detection so re-captured pages don’t bloat the index, and automated weekly digests that surface what’s been captured and what deserves re-reading.
State management across the streaming chat, the extension, and the main app runs on Zustand; the vector indexing and retrieval chain are RAG pipelines over ChromaDB and pgvector. Every part of the system exists to answer one question faster: “I know I saved something about this — where is it, and what did it say?”