Bedrock
A self-hosted, multi-tenant platform that turns a company's industrial PDFs — catalogs, datasheets, CAD, manuals — into searchable, cited, conversational knowledge, with confidence as a first-class UX element.
Overview
Extraction without trust is useless, so Bedrock makes confidence visible: every extracted item carries a 0–100 score and a bounding box, rendered as a three-bar signal so a user knows what to double-check.
Upload flows into a fire-and-forget pipeline decoupled from the request — queue-ready — lands in a denormalized Postgres search spine, and comes back as synthesized answer cards with citations. It's built to run on the customer's own infrastructure: row-level security handles tenant isolation, and the app layer enforces capability and corpus scope on top.
Highlights
- Postgres full-text search as the deliberate spine — no external search service.
- Repository, storage, and extractor swap-seams so pieces are replaceable.
- White-label theming; the first user is bootstrapped as owner, everyone else invite-only.