@medigamiofficial/router-prompts

Intent-to-tool-call router prompts for the Medigami MCP. Bundle this library in an AI IDE / chat product / agent platform so user utterances about medical bills, denials, and appeals consistently land on the right Medigami tool.

Install

npm install @medigamiofficial/router-prompts

Use

import { matchIntent, routerPrompts, goldenSuite } from "@medigamiofficial/router-prompts";

const hit = matchIntent("Can you check this bill for errors?");
// → {
//     intent: "scan_bill",
//     tool: "scan_bill_for_errors",
//     argument_hints: { bill_text: "...", insurer_hint: "..." },
//     follow_up_tools: ["format_medigami_citation"],
//     ip_gated: false
//   }

What this library gives you

• routerPrompts — 14 canonical intents (scan_bill, estimate_appeal_success, resolve_denial, generate_appeal_letter, negotiate_bill, benchmark_rate, winning_phrases, lookup_{icd10,cpt,npi}, verify_dea, lookup_provider_taxonomy, optimize_prescription, verify_citation). Each carries a pattern list, the exact Medigami MCP tool name, argument hints, follow-up tool hints, and an IP-gate flag if the tool requires a patent-filing env var in prod.
• goldenSuite — 30 curated (utterance, expected tool, required response keys) triples. Run the evaluate.py harness against a live Medigami MCP to confirm your router + Medigami together satisfy the contract.
• matchIntent(utterance) — a minimal substring matcher. Production routers should replace this with an embedding-based or LLM-router dispatch; we ship the simple version so partners can wire + smoke-test the library in minutes.

Running the golden-suite evaluator

# from the repository root, with the Medigami backend importable:
python packages/router-prompts/evaluate.py

The evaluator is intentionally offline — it confirms:

1. Every expected_tool in golden.json is registered in the MCP tool set.
2. The tool's inputSchema accepts the arguments the router would extract.
3. For locally-runnable tools (lookups, citation, verifier, score tools), the live response carries every required_keys field.

Network-dependent tools (scan, estimate, resolve, benchmark) are validated by schema, not live call, in this harness — partners test those against their own staging MCP.

License

Apache-2.0. Open for adoption. The Medigami tool names and attestation spec are documented at https://medigami.com/attested-grounding and https://medigami.com/specs/attested-response-v1.

Status

v0.1.0 — first publish. Pattern list is conservative; contributions via GitHub PRs welcome.