Review a supplier transaction.
Supplier records, contracts, invoices, delivery evidence and procurement policy.
Agent-Scoped Knowledge Graph & Tri-Hybrid Retrieval
Find the evidence.
Follow the links.
Keep the boundary.
Search by exact terms, semantic meaning and business relationships—within the knowledge your application is allowed to use.
Give agents a governed view of enterprise knowledge. Carry the active authority through every retrieval path, and preserve why each piece of evidence belongs in the result.
LexicalFind “INV-1042” and exact policy terms.
WordsVectorConnect “delivery proof” to “goods accepted.”
MeaningGraphFollow invoice → PO → receipt.
RelationshipsEligible candidates onlyCombine evidence. Keep its origins.Source revisions · retrieval paths · scope
Context Assembly →Check freshness, trust and sufficiency.
A governed path into contextApplication-scoped knowledgeEligibility before rankingRelationship-path evidenceSource authority retained
Governed retrieval across your adoption path
Govern existing agents
Retrieve eligible evidence for the task your existing agent brings to Frozion.
Grow tools into workflows
Use lexical, vector or relationship-based retrieval within the workflow’s authority boundary.
Build native agents
Native agents find supporting evidence while retaining its sources and scope.
Build systems of engagement
Application cases can connect permitted documents and related business records.
01 / Start with the application
Enterprise knowledge.
A deliberate scope for each agent.
The enterprise may know more than a particular application, user or workload should be able to use.
An Agent-Scoped Knowledge Graph defines the relationships and source classes useful to an application. Active identity, delegated authority, tenant, purpose and source permissions narrow that view for each execution.
Investigate within its mandate.
Permitted risk indicators, investigation cases and restricted evidence under the appropriate authority.
Show the permitted self-service view.
Supplier-owned submissions, status and policy guidance appropriate to that supplier.
These can be different application views over shared enterprise semantics. Using the same agent does not give every user the same knowledge.
Build from the relationships
you already define.
Use governed data models, API contracts, connectors, documents and taxonomy to compose useful knowledge scope. Ownership, classification and lineage stay attached.
Start with one application’s entities and relationships. A complete enterprise ontology is not a prerequisite.
- Materialized relationships
- Persist entities and edges where a local representation is useful.
- Semantic relationships
- Reuse links among model types and business concepts.
- Live source routes
- Resolve related information through a governed API or connector while the source retains ownership.
A relationship is not current operational truth. A graph can connect a supplier to a payment instruction. The authoritative source still establishes the current value, status and revision.
02 / Continue the invoice review
Words. Meaning. Relationships.
One inspectable evidence set.
INV-1042 already has reusable Feature Store facts. Now find the delivery evidence and governing policy that help the application review it.
Compare the retrieval paths, then change the conditions. See what is eligible, why it was found and what Context Assembly still needs to check.
Compare retrieval paths for a fixed set of candidate sources.
Tenant A · SUP-204 · Purpose: invoice reviewAll paths use the same authority
Exact references
Resolve the invoice identifier and the explicit “invoice review” policy terms.
2 eligible candidatesRelated language
Find “goods accepted” as delivery proof and conceptually related review requirements.
2 eligible candidatesBusiness connections
Follow eligible invoice, purchase order, receipt and legal-entity policy relationships.
2 eligible candidates6 path appearances → 3 unique sources
Combined view deduplicates the same source revision and retains every contributing retrieval path. Display order follows the review’s evidence groups; it is not a comparison of raw scores.
Invoice reference
INV-1042 · invoice r5
The lexical path locates the exact invoice. The semantic and graph paths add delivery and policy evidence.
3 unique eligible sources in the combined example
CANDIDATES FOUND
The evidence can move into Context Assembly.
Resolve current records and reusable Feature Store facts, verify the receipt, and check the complete review contract. Retrieval alone does not establish readiness or approve the invoice.
- Execution & authority
- RT-1042-A · Tenant A · AP reviewer · invoice review · scope revision 12.
- Retrieval versions
- Lexical index L7 · vector index V9 · graph view G4 · source revisions retained.
- Eligibility & exclusions
- All displayed candidates are eligible under the active scope. No protected candidate details are exposed by this view.
- Selection & attribution
- 6 eligible path appearances, 3 unique sources. Deduplicate by source ID and revision, retain path attribution; fixed display order.
- Context handoff
- Selected evidence retains source identity, revision and permitted relationship paths. Context Assembly adds operational facts and FS-1042-A, then checks sufficiency.
- Cache & model use
- A retrieval record should identify cache reuse, retrieval configuration and transformations when used.
Lexical is good at
“Where is this identifier?”
Exact names, supplier numbers, contract references, policy codes and domain terminology.
Semantic is good at
“What describes this idea?”
Related meaning when the request and the evidence use different language.
Graph is good at
“How is this connected?”
Explicit business relationships and multi-hop paths that explain relevance.
03 / Eligibility before candidate competition
Three search paths.
The same authority on each.
Connected does not mean authorized.
A permitted supplier node does not authorize every relationship or document connected to it. Check the starting entity, traversed relationship, destination and related content against the active scope.
Resolve the active scope.
Carry subject and workload identity, delegation, tenant, taxonomy, resource, purpose, classification, residency and current source permissions.
Permission checks across every search path.
Apply equivalent eligibility semantics to lexical matches, semantic candidates and graph traversal before they enter retrieval competition.
Fuse eligible candidates.
Combine and rerank inside the authorized candidate set. Restricted material must not consume a candidate, ranking or context slot.
Reused content stays within scope.
Source permissions, revisions and lifecycle state remain associated with chunks, embeddings, summaries, index entries and cached results.
A broader query keeps
the same permission boundary.
Synonyms, generated subqueries and graph seed expansion can improve recall. They change what the application looks for, while the active authority still determines what may participate.
Explore Identity & SecurityWhat was permitted at ingestion
may not be eligible now.
Reclassification, deletion, a changed assignment or revoked access can alter eligibility. Re-evaluate current scope when using derived artifacts or resuming work; an old index entry is not enduring permission.
Explore durable executionEligible evidence still needs a trust role. An authorized document can be stale, misleading or written to manipulate a model. Retrieved content remains evidence; embedded commands do not become application instructions.
04 / Preserve the reason, not just the result
Know why it was found.
Know what the decision used.
A graph-derived result can retain its seed, relationship path and destination. A Retrieval Authorization Manifest can record the authority, versions, candidate generators, eligibility outcomes and selection behavior behind the evidence.
Carry those references into Context Assembly and the resulting decision, policy evaluation and human review.
Permitted relationship path
Why AP policy v4 applies
- Invoice INV-1042belongs_to
- Legal entity LE-Agoverned_by
- AP policy v4Active policy resolved from its authoritative source
What stays attached
- Seed, eligible edges and destination
- Source IDs, revisions and retrieval method
- Active scope and permission revision
- Selection, exclusions and context handoff
Finding candidates is one step.
Preparing a decision is another.
A result can be eligible and relevant, yet stale, conflicting, redundant or insufficient. Context Assembly makes those conditions explicit before the next decision.
Explore Context Store & Assembly- Authoritative records
- Resolve invoice r5 and PO r2 through governed queries or source APIs. Search hits do not replace current transactional state.
- Reusable calculated facts
- Consume FS-1042-A: 2.0% PO variance, $148,500 exposure and 0 prior matches. Check freshness; avoid repeating the calculation in each agent.
- Consistent review inputs
- Bind records, facts, verified receipt and active AP policy into a versioned package. Agents, policies and human reviewers receive their permitted views of the same input basis.
05 / Questions, answered
How the pieces
work together.
A shared knowledge foundation can support many applications while each execution keeps its own boundary.
Not necessarily. An application knowledge view can serve one or more agents over shared enterprise semantics. The active user, workload and purpose further constrain that view. Separate physical graphs are an implementation choice, not the definition of the boundary.
They answer different questions. Lexical retrieval finds exact terms and identifiers. Vector retrieval finds related meaning. Graph retrieval follows explicit relationships. Combining eligible results can improve evidence coverage, while keeping each method’s contribution visible. None alone proves context sufficiency or authority to act.
The architectural requirement is eligibility before competition. An ineligible result can consume a candidate slot, influence reranking or displace useful evidence even if it is hidden later. Each retrieval path must honor the active scope before candidates influence selection and Context Assembly.
It establishes a relationship used for discovery and reasoning. Current account details, invoice status or active policy may still need to be resolved from Operational Collections, authoritative systems or governed APIs. Preserve source identity and revision so the application can distinguish relationship evidence from current state.
Derived chunks, embeddings, summaries and index or cache entries remain subject to the source’s scope and lifecycle. Current permission and classification state must be evaluated when evidence participates. A previously indexed or cached copy must not become a route around changed access.
Existing content systems, lexical indexes, vector infrastructure, graphs and APIs can retain their roles. Frozion connects them through application knowledge scope, source permissions, eligible candidate generation and evidence lineage. The supported integration and enforcement behavior must be established for the chosen systems.
Eligible candidates with source references, versions, retrieval attribution and permitted graph paths. Context Assembly adds authoritative records and calculated facts, checks freshness, trust, conflicts, budget and sufficiency, and preserves the versioned package used by the next agent, policy or human decision.
Start with one application and its knowledge boundary
Connect the knowledge.
Keep the scope.
Bring the sources, relationships and permissions your workflow depends on.
We’ll map the retrieval paths and the evidence behind the decision.