Knowledge Bases

Knowledge Bases are workspace-scoped document collections. Unlike memories (which are created from agent interactions), knowledge bases hold static reference material: documentation, policy documents, product specs, etc.

Schema

knowledge_bases: {
  id: uuid PK
  workspaceId: uuid FK → workspaces (cascade delete)
  name: text
  description: text | null
  createdAt: timestamp
  updatedAt: timestamp
}
 
knowledge_entries: {
  id: uuid PK
  knowledgeBaseId: uuid FK → knowledge_bases (cascade delete)
  content: text
  embedding: vector(1536)
  metadata: jsonb    // { source, page, author, ... }
  createdAt: timestamp
}

Difference from Memories

API Endpoints

Search

The search endpoint currently uses ilike keyword matching. It is designed to be upgraded to pgvector cosine search once embeddings are generated on entry insert: the embedding column is already present in the schema.

// Current implementation
.where(and(
  eq(knowledgeEntries.knowledgeBaseId, baseId),
  ilike(knowledgeEntries.content, `%${query}%`)
))

Frontend

The knowledge UI is at /knowledge. It lists all bases in the active workspace. Selecting a base opens an entry list with search. Knowledge bases span all spaces: they are workspace-wide shared resources.