Second Brain Agent

Private

A full-stack agentic RAG pipeline that indexes an Obsidian knowledge base and answers queries with real-time reasoning, source attribution, and knowledge graph traversal.

Second Brain Agent knowledge graph visualization

What I built

Built in 2 hours at the NVIDIA Agents for Impact hackathon. A 4-stage agentic RAG pipeline over my personal Obsidian vault (69 notes with wikilink references).

Stage 1: Nemotron generates 3–5 expanded search queries from the user's question. Stage 2: Each query is embedded via NV-Embedqa-E5-v5 and searched against ChromaDB (top-5 per query, deduplicated). Stage 3: Nemotron reasons over retrieved chunks with extended thinking enabled — the reasoning trace streams to the frontend in real time via SSE. Stage 4: Structured synthesis with the answer, discovered cross-domain connections between notes, and an auto-generated markdown note with backlinks that saves directly into the vault.

The frontend streams status updates through each phase (expanding, searching, retrieving, reasoning, answering, connecting) with a collapsible thinking panel, source cards with relevance scores, and a knowledge graph visualization.

Tech stack

NVIDIA Nemotron 3 Super 120B (query expansion + reasoning + synthesis)NVIDIA NV-Embedqa-E5-v5 (passage/query-aware embeddings)ChromaDB (persistent vector store, cosine similarity)FastAPI + SSE (real-time token streaming)Next.js 14 + React 18 (chat UI with streaming state)Tailwind CSS

Key stats

4-stage agentic pipeline (expand → retrieve → reason → synthesize)

Real-time SSE streaming with reasoning traces

Cross-domain connection discovery across 69 notes

Auto-generated synthesis notes saved back to vault

Built in 2 hours at NVIDIA hackathon