Orbit

Inspiration

Most note apps are just dumping grounds. You save a thought, a link, a PDF — and then it just sits there. Weeks later you’re scrolling, hoping something rings a bell.

We wanted something that keeps notes in motion. Second Brain is all the hype right now, We brought them to everyone. Capture rough ideas, let them connect, and actually ask questions against what you’ve saved — without losing track of what came from you vs the web.

That’s Orbit.

What it does

Orbit is a note-first personal knowledge system.

You dump a rough note, link, PDF, DOCX, Markdown, or TXT. Orbit keeps the original, interprets it, researches when it’s worth it, and compiles everything into a provenance-aware wiki.

From there you can:

Browse a growing personal wiki
See connections in a knowledge graph
Ask questions with citations and confidence
Capture thoughts over Discord DMs
Get optional proactive nudges (weekly synthesis, forgotten ideas, contradictions, research updates)

No onboarding quiz. No perfect formatting required. Just capture and go.

How we built it

Frontend: Next.js dashboard (notes, wiki, research, graph, ask, proactive, profile)
Backend: FastAPI pipeline for intake, interpretation, research, wiki compilation, and Q&A
Data: PostgreSQL + pgvector for storage and hybrid retrieval
Research: Selective web research (Wikipedia / Tavily) with auditable source decisions
LLM: OpenAI-compatible or local Ollama for structured interpretation and answers
Auth / capture: Discord OAuth + Discord bot for the same personal space across web and DMs
Infra: Docker Compose with Redis, MinIO, and a production path via Caddy HTTPS

The core loop is simple: save → understand → research if useful → compile wiki → answer from evidence.

Challenges we ran into

Keeping answers grounded in real evidence instead of letting the model freestyle
Deciding when research is worth it vs when a note should just stay personal
Preserving provenance so user notes, attachments, external sources, and synthesis stay visually distinct
Handling long PDFs and articles without blowing model context budgets
Making Discord + web feel like one brain without leaking data across users
Building proactive routines that are useful, not spammy

Accomplishments that we're proud of

A zero-question first-run experience that still feels powerful
Provenance-aware wiki blocks with clear origin (user / attachment / external / system)
Selective research that records every query, score, and acceptance reason
Citation-backed Ask with confidence and knowledge paths
Discord capture that opens the same personal space as the dashboard
Full export + deletion so you actually own your data

What we learned

“Second brain” products fail when they invent. Trust comes from sources and provenance.
Notes should stay immutable. Synthesis can rebuild; originals shouldn’t get rewritten.
Research is a product decision, not a default. Privacy and noise matter as much as enrichment.
Hybrid retrieval (lexical + vectors) beats either approach alone for personal knowledge.
Shipping a working vertical slice beats designing the perfect architecture on paper.