Knowledge — current implementation
Historical graph-design note: LightRAG/graph-backend text below is superseded by AGNT2-303; current retrieval uses words and vectors.
Teams upload documents into folders. Open folders and explicit agent grants decide what each agent can retrieve. This is distinct from conversational memory. Store-catalog sync remains a product scenario, not an implemented connector.
What technology we use
The current implementation uses the existing Postgres-backed word and vector indexes.
Documents are extracted, chunked, contextualized and indexed. Search combines words and vectors by reciprocal rank. The turn reads a bounded set of passages from the same search service as the human search path. Tenant and folder scope apply both to searching and resolving matches.
How it scales
Scaling is bounded by ingestion and retrieval budgets; a graph backend is not part of the current design.
The wall is ingestion, not storage
The expensive path is extracting and indexing a large corpus. Retrieval remains bounded and scoped.
Retrieval supplies evidence; it does not prove answer correctness. Current turn passages carry document name, ordinal, text and score. A durable evaluation workflow and source-faithfulness verdict are not established. No graph service is required.
| Source | Evidence |
|---|---|
| knowledgeSearch.service.ts | word + vector / RRF |
| knowledgeLookup.service.ts | team / folder scope; bounded turn passages |
| knowledge.module.ts | ingestion / retrieval wiring |
| research.md | dated graph comparison |