Agent-Scoped Knowledge Graph & Tri-Hybrid Retrieval

Exact terms find names. Semantic search finds meaning. Relationships reveal how things connect.

Compose governed knowledge relationships, then search by lexical terms, semantic meaning, and graph relationships inside the same application authority boundary.

BM25 / Lexical exact names, identifiers, terminology
Vector / Semantic related evidence with different wording
Graph / Relationships governed multi-hop paths
Connected does not mean authorized.
Knowledge boundary

Start with the knowledge boundary the application needs

Application scope

A supplier account-change application needs supplier, legal entity, bank account, payment instructions, policy, and prior account-change relationships.

Active authority

The same application view narrows by initiating subject, workload, tenant, legal entity, taxonomy, purpose, classification, and policy state.

Reusable without widening

One reusable knowledge application can serve different users and purposes without making the graph a broad permission grant.

Relationship foundation

Build knowledge relationships from semantics you already define

Relationships can come from governed data models, API contracts, connector definitions, documents, taxonomy, ownership, classification, and lineage.

Supplier to Legal Entity
Supplier to Bank Account
Payment Instructions to Policy
Case to Evidence and Lineage
Three retrieval paths

Use all three retrieval paths for different parts of the same question

Lexical

Resolve exact supplier names, contract identifiers, account suffixes, policy codes, and technical terminology.

Semantic

Find evidence whose meaning is related even when documents and systems use different language.

Graph

Follow explicit relationships through supplier, legal entity, account, payment instructions, policy, and prior cases.

Relationship authorization

Follow relationships without inheriting permissions

An authorized supplier node does not automatically authorize restricted fraud investigation or privileged legal review just because an edge exists.

Candidate eligibility

Authorize candidates before they compete for rank

Ineligible information should not consume candidate slots, reranking slots, diversity budget, cache entries, or context budget.

One scope contract

Keep all three retrieval paths inside the same authorization boundary

The architecture is not search everything three ways, merge it, then filter. Scope applies before an item can influence ranking or Context Assembly.

Lexical scopeVector scopeGraph traversal scopeFusion scopeReranking scopeContext Assembly scope
Incremental scope

Build useful graph scope around real applications

The primary unit can be the knowledge scope required for an application, not a complete enterprise-wide graph.

Operational truth

Keep relationships separate from operational truth

The graph can establish a relationship while Operational Collections, source systems, or governed APIs remain authoritative for current transactional state.

Map one governed knowledge view and retrieval boundary.

Use an architecture review to identify application scope, active authority, lexical/vector/graph retrieval paths, candidate eligibility, and Retrieval-to-Context Lineage.