Enterprise · Agent integration

Agent integration

Use Enterprise routing to translate an agent question into an explicit retrieval request, then give the agent only authorized evidence and validate its citations.

Integration flow

Step	Owner	Contract
1. Establish scope	Application authentication boundary	Tenant, user, workspace, matter, document, and dataset permissions
2. Route	Enterprise query hook	Intent, retrieval plan, evidence plan, filters, limit, confidence, review signal
3. Retrieve	CoreCortex query API	Authorized ranked Evidence with trace identity
4. Act or synthesize	Application agent	Tool result or answer over supplied evidence
5. Validate and log	Application + Core contracts	Citation membership, abstention, request ID, trace ID, model identity

Call the routing hook

shellCoreCortex Enterprise

cat > request-scope.json <<'JSON'
{"matter_id": "matter-482", "jurisdiction": "federal"}
JSON

corec query route \
  "What authorities does this matter cite?" \
  --dataset authorities \
  --filters request-scope.json \
  --output routing-decision.json

Read the routing decision

The decision is advisory and inspectable. The application intersects its filters with authenticated policy before retrieval. A routing confidence score never grants access and never acts as answer confidence.

jsonCoreCortex Enterprise

{
  "apiVersion": "corecortex.io/enterprise/v1alpha1",
  "kind": "QueryRoutingDecision",
  "request": {
    "query": "What authorities does this matter cite?",
    "intent": "relationship",
    "retrievalPlan": "bounded_graph_then_evidence",
    "scope": "matter",
    "evidencePlan": "graph_path",
    "filters": {"dataset": "authorities", "matter_id": "matter-482", "jurisdiction": "federal"},
    "limit": 10,
    "explain": false
  },
  "routing": {"signals": ["matter_filter", "relationship_language"], "confidence": 0.85, "requiresAgentReview": true}
}

Translate routing into the query API

CoreCortex applies tenant and record policy at candidate generation and again before evidence leaves the retrieval boundary. The response gives the agent evidence IDs, source identity, content hashes, processor revisions, and trace information.

httpCoreCortex Core

POST /v1/corpora/{corpusId}/query
Authorization: Bearer <token>
X-CoreC-Tenant-Id: <tenant-uuid>
Content-Type: application/json

{
  "query": "What authorities does this matter cite?",
  "retrievalPlan": "bounded_graph_then_evidence",
  "intent": "relationship",
  "filters": {"dataset": "authorities", "matter_id": "matter-482", "jurisdiction": "federal"},
  "limit": 10,
  "explain": true
}

Supported deterministic routing intents

Intent	Detected signal	Suggested strategy
lookup	Structured identifier such as record key, external identifier, or version code	Lexical single passage
exact	Quoted or explicitly verbatim language	Lexical single passage
comparison	Compare, difference, changed, versus, or timeline	Hybrid multi-passage
relationship	Cites, connected, between, path, authority, treatment, or linked	Bounded graph then source evidence
semantic	Detopic natural-language question	Hybrid evidence retrieval

Add customer-owned synthesis

Set responsePlan when the agent needs a generated answer. The Synthesis page defines the request, response, citation, and abstention rules; the Evaluation page defines the answer tests.

Configure generated answers Test generated answers

Agent safety invariants

Authentication and tenant context come from the trusted application boundary.
Routing hints can narrow scope and choose a strategy; they cannot expand permissions.
Relationship paths require agent review and resolve back to source evidence before use.
The agent receives stable evidence identities and cites only supplied fragments.
Expected-empty and unavailable-source cases lead to explicit abstention behavior.
Request, trace, configuration, index, model, and judgment identities remain available for replay and evaluation.