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.
Answer
Retrieved sources
click to expandembedding query
hybrid retrieval · pgvector + FTS + trigram
synthesizing with citations
Under the hood
- embed — OpenAI
text-embedding-3-small(1536d) - retrieve — Postgres function
hybrid_search_vectra: HNSW semantic over content + summary + hypothetical questions, fused via RRF with English FTS and trigram match - synthesize —
gpt-4o-miniwrites the answer using only the retrieved chunks; cites each claim with[^N] - graph — parent/child edges between chunks within each document for traversal queries (not shown in this UI)