Register your rule engine as a Custom HTTP Tool

From AgentsHub → Tools, click Add HTTP Tool. Point it at your internal endpoint (for example POST /rules/check) and declare its JSON schema so the agent can call it safely. Agents do not learn to parse undocumented APIs — they pick from a typed catalog, and the schema is your control surface for what is and is not allowed during a run.

If you operate multiple lines of business (commercial, Medicare Advantage, Medicaid), register one tool per LOB or pass a line_of_business parameter and switch behavior server-side. Either pattern keeps the audit trail clean and lets SEEMR learn LOB-specific routing.

{
  "tool_name": "health_rule_check",
  "endpoint_url": "https://rules.internal/v1/check",
  "http_method": "POST",
  "auth_method": "bearer_passthrough",
  "parameters_schema": {
    "type": "object",
    "properties": {
      "plan_id":    { "type": "string" },
      "claim":      { "type": "object" },
      "context":    { "type": "object" }
    },
    "required": ["plan_id", "claim"]
  }
}

VDF AI stores the tool in tool_catalog with tool_type='http' and merges it into the agent catalog — visible only to the owner or to your domain.

Define an Intent Template for "adjudicate-claim"

Intent templates teach Networks v3 how to decompose a request. For adjudication, the template names the sub-intents (eligibility, medical necessity, plan benefit lookup, exclusion check, fraud signal) and the tools each sub-intent is allowed to call. The template is the unit of governance — change the template and you change the whole network's behavior, with a versioned audit trail.

Keep templates small and composable. A complex prior-authorization workflow is just an adjudicate-claim template that conditionally invokes a prior-authorization template when a procedure code requires it. Reuse beats one-off pipelines every time.

• Eligibility → member_lookup + policy_status
• Medical necessity → health_rule_check + RAG of clinical guidance
• Exclusion → RAG of plan documents + structured exclusion lookup
• Fraud signal → vector search on prior denials

Author the System Prompt for the Rule Checker Agent

The system prompt encodes how the agent must reason, justify, and cite. Keep it deterministic and auditable. The goal is not creativity — it is reliable evidence packaging. The prompt should specify the exact JSON shape of the output, the citation requirement, the escalation conditions, and the explicit "do not" guardrails.

You are a Health Insurance Rule Checker.
Always:
 - Quote the plan clause and CMS reference verbatim.
 - Call health_rule_check before producing a decision.
 - Return JSON: {decision, reason, citations[], confidence}.
Never:
 - Speculate beyond the rule set.
 - Reveal PHI outside the requested fields.

Vectorize plan documents and CMS guidance with VDF Data

Use VDF Data's Vector DB Builder to chunk and embed plan booklets, schedules of benefits, and regulator guidance into a pgvector index. Bind that index as a rag_vector_query tool inside the Network.

Compose the Network and let SEEMR govern routing

In Network Labs, drop the Rule Checker Agent, the RAG retriever, and the Custom HTTP Tool onto the canvas. Bind them to the adjudicate-claim intent template. SEEMR then governs which model handles which sub-intent — high-stakes adjudication on your private high-capability model, retrieval on a small, energy-efficient SLM.

Monitor every adjudication in Live Execution Monitoring

Every claim flows through Network Flow with timing, success rate, and per-node logs visible in real time. Reviewers can pause auto-scroll, replay a run, or escalate to a human.