Knowledge overview¶
"Knowledge" is what the Website Agent knows about a client and can ground its answers in. This page is the high-level map: how a client's knowledge is organised, built, stored, and served, with links down into each sub-concept. For precise definitions of the terms used here, see the glossary.
The diagrams describe the target site model. S2 implements explicit ingestion
into new site:<UUID> corpora; legacy ingestion keeps domain keys. Runtime
selection of those new corpora is S3, and shared authored-document membership
is S6. See the S2 command and lifecycle
for the implemented boundary.
The lifecycle¶
The four stages¶
-
Identity — what the knowledge belongs to. A site is the client-scoped identity that owns exactly one corpus (its documents, FAQs, vectors, and graph) and one or more roots (the URLs it is placed at, which also derive the crawl seeds). Retrieval for a turn is scoped to the resolved site's corpus and nothing else. See the Client-Scoped Sites ADR and the glossary.
-
Ingestion — how the corpus is built and kept fresh. Website updates, managed documents, and knowledge FAQs become durable background operations. The content worker discovers, scrapes, and cleans pages; the document-loader is the controlled writer into the RAG stores. Supabase owns the workflow state. See the Knowledge ingestion control plane.
-
Storage — where the corpus lives. Ingested content is written through LightRAG into MongoDB (chunks and vectors) and Neo4j (the knowledge graph), every row scoped by
tenantId. See the Backend overview storage table and the IXRag package. -
Retrieval — how knowledge reaches an answer. A visitor's message is turned into a retrieval query, then
ixragruns hybrid vector + graph retrieval and reranking over the resolved site's corpus. See Query resolution for retrieval and the IXRag package.
Diagnostics cut across all four: rose-document-loader diagnose-tenant
inspects what RAG can actually read and flags corruption or cleanup candidates.
See Knowledge base diagnostics and, for exact-passage
evidence, Capture retrieved passages.
What ingestion covers¶
Ingestion spans from a producer trigger to the KB being updated — it owns the work in between and tracks its durable state. Authoring a site's identity is upstream of it, and retrieval at query time is downstream.
Managed documents and knowledge FAQs enter the same control plane but skip discover/scrape/clean — they go straight to the document-loader.
Where to go next¶
| I want to… | Page |
|---|---|
| Understand the workspace / site / corpus / root model and terms | Glossary |
| See how ingestion runs as durable operations | Knowledge ingestion control plane |
| Understand the retrieval storage layer | IXRag package |
| Learn how a message becomes a retrieval query | Query resolution for retrieval |
| Inspect what RAG can read for a tenant | Knowledge base diagnostics |
| Capture the exact retrieved passages | Capture retrieved passages |
| Read the ingestion architecture decision | Control plane ADR |