Inspiration
Standard RAG is broken. Every time you ask an AI a question, it digs through raw document chunks and re-derives the answer from scratch. There's no memory. No structure. No compounding knowledge.
Andrej Karpathy put it well — AI systems should build knowledge, not just retrieve it. That idea became NeuroWiKi: what if every source you read automatically compiled into a living, interlinked wiki that grows smarter over time?
What It Does
NeuroWiKi ingests any source — URLs, PDFs, docs, plain text, and uses AI to extract structured wiki pages with automatic wikilinks between concepts. Instead of searching chunks, you query a compiled knowledge base that improves with every new source added.
Key features:
- Auto wiki generation — AI extracts 1–5 interlinked pages per source
- Knowledge graph — visual map of how concepts connect
- Hybrid recall — vector + graph search over your wiki
- Contradiction detection — consistency agent flags conflicting knowledge
- Diary & captures — daily journaling tied to your knowledge base
How We Built It
- Next.js 15 (App Router) for the full-stack framework
- Google Gemini (multi-model fallback chain) for wiki extraction and Q&A
- HydraDB — hybrid vector + graph store for knowledge retrieval
- SQLite — local source metadata, page health, wikilinks, query history
- React Force Graph — interactive knowledge graph visualization
- Multi-agent pipeline: ingest → absorb → consistency → lint agents run on every source --- ##Challenges The hardest problem was knowledge merging. When two sources describe the same concept differently, the absorb agent must decide: create a new page, merge into existing, or flag a contradiction. Getting this right without losing nuance took significant prompt engineering.
Incremental recall was also tricky — making sure new pages don't fragment existing knowledge but instead strengthen it required building a slug-based identity system with alias resolution.
What We Learned
That the gap between "search" and "understanding" is architectural, not just a model problem. Building a system that accumulates knowledge rather than queries it fresh each time changes what AI assistants can do.
