An AI assistant that answers only from your real data

Most "AI assistants" hallucinate because they can't see your real data — they're a generic model wearing your logo. The fix isn't a bigger model, it's retrieval: ground every answer in documents that actually exist, and cite where it came from.

Documents, docs, and a running system's own knowledge get chunked and embedded into Postgres (pgvector), retrieved at question-time, and answered with citations back to source — so "the assistant said so" is always checkable, not just trusted.

The same underlying knowledge also exports as a wiki-linked markdown vault — one page per fact, live counts, cross-links — so a human can browse the exact brain the assistant answers from, in Obsidian, with zero extra tooling. Plain files are the ultimate portability: any human can read them, any agent from any provider can too.

Ingest your real documents — PDFs, docs, wikis, whatever you actually have — chunked with source anchors, embedded in Postgres, and served through a chat interface your team reaches where they already work, with every answer traceable back to its source. Sized to what you actually need: no vector-database vendor to babysit for a dataset that fits comfortably in Postgres.

I haven't shipped inside every proprietary platform — where that's true, I say so upfront rather than performing depth I don't have. What I have shipped, repeatedly, is the retrieval architecture itself: chunking, embeddings, citation-grounded answers, and knowledge that stays inspectable instead of becoming a black box.