An agentic knowledge base that answers questions with citations.

Kibi

RoleSolo founder — design, engineering, product
Timeline2026
StackNext.js, TypeScript, Supabase, pgvector, Claude

The problem

Notes, conversations, and documents pile up faster than anyone reorganizes them. The knowledge is in there somewhere — the retrieval is the part that never gets built.

What I built

Kibi is an agentic knowledge base: paste or upload content, an AI extracts structured entries — insight, decision, action item, concept, reference, note — and a retrieval agent answers questions against them with citations back to the source.

  • Retrieval is a real pipeline, not a keyword search. pgvector with an HNSW index; every stored vector records which embedding model produced it, so the model can be swapped later without breaking old data.
  • Prompt injection is resisted structurally. Retrieved content is data, and only ever enters the user turn — never the system prompt. The model can't be redirected by something it retrieved.
  • Row-level security on every table. Retrieval can't leak across accounts even if application code has a bug.
  • A real bug, root-caused properly. Magic-link sign-in was silently failing for some users. The cause turned out to be a mail client's link-prescanner opening (and burning) the single-use token before the person ever clicked it — found by reading the actual auth logs, not by guessing.
A diagram. Pasted or uploaded content goes through an AI step that extracts entries (insight, decision, action item, concept, reference, note). The entries go into a pgvector store, and a retrieval agent answers with citations. Row-level security wraps the store and the agent.
How Kibi turns content into a cited answer. A diagram, not a screenshot: extraction, a vector store, and retrieval, with retrieved text kept out of the system prompt.

Where it stands

Phase 1, built for a single user.

Got a hard problem? I’d like to hear about it.