Vectra operational intelligence · live demo
chunks edges 1536dembeddings

Ask the institutional memory of a business that doesn't exist yet.

Vectra indexes the unglamorous operational guts of a company — customer threads, supplier emails, SOPs, decision logs, product specs — into a hybrid vector + full-text + graph store, then retrieves and synthesizes answers with citations. The corpus below is for a fictional premium home-goods brand, Verdant Home Goods. The retrieval, ranking, and synthesis are real.

Under the hood

  1. embed — OpenAI text-embedding-3-small (1536d)
  2. retrieve — Postgres function hybrid_search_vectra: HNSW semantic over content + summary + hypothetical questions, fused via RRF with English FTS and trigram match
  3. synthesizegpt-4o-mini writes the answer using only the retrieved chunks; cites each claim with [^N]
  4. graph — parent/child edges between chunks within each document for traversal queries (not shown in this UI)