Skip to main content

RAG vectors

The vectors used for RAG (knowledge bases) and the document index are stored directly in HocNota's PostgreSQL database (pgvector extension), alongside the rest of the data — no separate service to configure. This means organization isolation applies to vectors exactly like to any other data: no organization can ever see or query another organization's vectors.

Admin → Data → Vectors

note

Don't confuse this with the neighboring RAG & Storage tab, which manages RAG storage spaces (see Managing RAG Storage Spaces) — a different topic from the pgvector vectors documented here.

The screen first shows a summary (total number of collections, total number of vectors), then the per-collection detail: technical name, linked knowledge base, vector count, last indexing date.

A collection is named kb_{id} after the knowledge base it serves. It's orphan when no existing knowledge base matches that id — typically after a knowledge base was deleted without its collection being cleaned up. The most common cause in practice: a leftover docidx_… collection created by a Document index node — normally deleted automatically at the end of the job (it never shows up here in that case), unless that job failed or was interrupted before cleanup.

  • Individual deletion — the row-level Delete button is only offered for orphan collections. This only removes the vectors; the knowledge base's documents (if it still exists) remain in place and will need reindexing (see Task Management).
  • Multi-selection — the checkboxes let you select several collections (including ones linked to an existing base) and delete them together in one batch action.
  • Purge orphans — a dedicated button, shown only when at least one orphan collection exists, deletes all of them at once.