Role-scoped retrieval
The filter comes from trusted server state, so a prompt cannot reach documents outside the caller’s role.
Authorization chooses the corpus. Ranking chooses within it.
A permission-aware support agent authenticates the caller, turns their role into a vector filter, and retrieves only the handbook records they are allowed to read. The question text can never widen the filter, and a similarity score is never an authorization decision.
From handbook to grounded answer.
Visibility is attached when documents are embedded, and enforced again on every request.
- 01
Embed the handbook
Validate documents at startup and embed them with flat visibility metadata.
- 02
Build the store
Load the embeddings into a vector store your app owns.
- 03
Authenticate the caller
Check the bearer token and the knowledge:read permission before anything else runs.
- 04
Filter, then retrieve
Derive a vectorFilter from the principal’s role and attach createVectorContext to a per-request agent.
- 05
Answer with a source
The agent cites the handbook, or says the handbook is insufficient instead of inventing policy.
Built to refuse the wrong answer.
Cited answers
Every answer names its handbook source, or admits the handbook does not cover the question.
Closed failures
A bad token returns 401 before parsing, search, or completion. Invalid JSON, extra fields, and oversized bodies are rejected.
Your vector store
Start in memory, then move to pgvector, Qdrant, Chroma, or LanceDB for production.
A per-request agent with a role filter.
The filter is decided by the principal, not the question.
1import { Agent, createVectorContext } from '@anvia/core/agent'2import { vectorFilter } from '@anvia/core/vector-store'34export function createSupportAgent(principal: Principal) {5 const filter = principal.role === 'manager'6 ? vectorFilter.or(7 vectorFilter.eq('visibility', 'agent'),8 vectorFilter.eq('visibility', 'manager'),9 )10 : vectorFilter.eq('visibility', 'agent')1112 return new Agent({13 id: 'customer-support-rag',14 model: completionModel,15 instructions: 'Answer only from the retrieved handbook documents.',16 context: [createVectorContext({17 store: knowledgeStore, model: embeddingModel, topK: 4, filter,18 })],19 })20}The example uses an in-memory store, which is not a production persistence strategy. For strong tenant isolation, add namespaces, database policy, or separate indexes on top of filters.
Build it.
Start from the guide, run the example, or look at what else Anvia handles.
