Provenance-grounded natural-language Q&A (§197)
POST a natural-language question about a supported U.S. healthcare-provider supply dataset. Returns a short answer plus the evidence fields available to the selected resolver, which can include source URL, observation timestamp, methodology, confidence, and limitations. Fields can be null when the audit-pack registry does not supply them.
Pipeline (SQL-injection-safe by construction):
- Pre-flight filters reject prompt-injection signatures, payment/claims questions, predictions, vendor comparisons, and single-provider PII fishing before any LLM call.
- Anthropic Sonnet maps the question to a strict whitelist intent (dataset slug + state code + intent kind) via tool-use; the input schema constrains output to the registered enums. Output is then runtime-validated.
- A typed resolver dispatches to hand-written, pre-existing data accessors. The classifier output never produces SQL — there is no query() call that concatenates LLM output into a query string anywhere in the pipeline.
- Available claim evidence is attached from the audit-pack registry; a second LLM call writes the summary constrained to the resolver's exact numbers, with a deterministic-template fallback.
County-scope honesty: NPPES has no public county-of-practice field. Queries scoped to a county return state-level supply with geo_widened: true flagged in the answer and an explicit limitation in the evidence trail.
Refusal codes: out_of_scope, predictions, payment_data, competitor_compare, individual_provider_lookup, prompt_injection, classification_failed. Refusals return HTTP 200 with status: "refusal" and a machine-readable refusal.code.
Demo (no auth, 10 queries / 24h per IP): POST /api/evidence-query/demo with the same body shape.
Request
- Base URL: https://fonteum.com
- URL: https://fonteum.com/api/v1/evidence-query
- Auth: HTTP bearer
Request body
Example request
{
"query": "How many active dermatologists practice in Texas, and how is that number sourced?"
}Response
Answer + evidence trail (or refusal).
Example response
{
"data": {
"answer": {
"methodology_version": "v2026.05.0"
}
},
"meta": {
"request_id": "req_abc123def456",
"freshness": [
{
"source_id": "oig-leie"
}
],
"freshness_policy": {
"max_staleness": "48h",
"max_staleness_hours": 48
}
}
}Changes
No recorded changes to this operation across all 1 revision of this API.