Text to your JSON schema, with evidence attached

What was tested

The evaluation is deliberately small enough to inspect. Twelve synthetic business-text cases freeze the input text, the caller's schema, the expected typed value, and the shortest source quote that supports it. The driver submits all cases in one bounded completion with temperature zero, parses one JSON array, and compares every field with ordinary equality. A malformed response, an unreachable model, a provider change, or an unreadable JSON result is UNKNOWN; the driver does not convert any of those failures into a zero or a pass.

The measured run used the shared model router task structured_extraction_api. Its configured chain was exactly [vllm], and the completion receipt named vllm. The source driver refuses any different chain before it sends the prompt. That isolates the cost and prevents a local outage from silently becoming a paid fallback. The command is reproducible from scripts/benchmark_structured_extraction.py; the complete public fixtures are available as eval-v1.json.

A pass means the value and source quote matched this frozen synthetic case. It does not establish production accuracy, generalize to another language or document type, or measure a marketplace service-level objective.

Proposed request and response

{
  "text": "Supplier ACME quotes USD 149.50 for Valve VX-200.",
  "schema": {
    "supplier": "string",
    "price_usd": "number",
    "sku": "string"
  },
  "infer_missing": false
}

{
  "fields": {
    "supplier": {
      "value": "ACME",
      "quote": "ACME"
    },
    "price_usd": {
      "value": 149.5,
      "quote": "149.50"
    },
    "sku": {
      "value": "VX-200",
      "quote": "VX-200"
    }
  },
  "unsupported_fields": []
}

The request would reject unknown schema types, nested executable expressions, schemas with more than 20 fields, and text longer than 8,000 characters. The response would preserve field order, return typed scalar values, attach one exact quote per supported field, and list unsupported keys. It would not browse links, execute markup, follow commands found inside the input, retain documents for training, or claim that extracted values are correct for legal, compliance, safety, credit, employment, or medical decisions.

Source quotes are evidence pointers, not proof of truth. If a supplier email says a price is $149.50, the quote can show where that number came from; it cannot prove the supplier's statement is accurate. A consuming system should retain the original authorized text and require human review wherever a wrong field would create meaningful harm.

The three-step path from benchmark to paid API

1. Review the contract. Download the public eval set, inspect the null and prompt-injection cases, and decide whether exact source quotes fit the consuming workflow.
2. Request evaluation access. Name one authorized text shape and a representative schema. The request is private and does not charge, subscribe, or authorize processing.
3. Qualify the metered listing. Only counted evaluation requests and a benchmark on the requested text shape justify activating a paid marketplace endpoint. A listing would expose fixed quotas, rate limits, retention terms, and the same UNKNOWN behavior before accepting payment.

Current state: there is no active RapidAPI listing and no public production extraction endpoint. The benchmark is evidence for a demand test, not a claim of availability. The product advances only if a buyer asks for this exact contract and the measured per-call cost clears the pre-registered self-funding gate.

What the public hub probe actually found

On 2026-08-09, the public RapidAPI search returned seven results for json schema extraction. Two listings were close schema-bound text-to-JSON comparables. Their public pages exposed three combined subscriptions, zero combined ratings, and monthly tiers ranging from free to $149.99. One listing exposed a popularity score of 8.1. These are observations of marketplace listings, not revenue, paid subscribers, or transferable demand for this proposal.

• JSON-Forge — AI Text to Structured JSON: public tiers observed at $0, $10, $29, and $79 per month; one public subscription; popularity 8.1.
• AI Text-to-JSON Extraction API: public tiers observed at $0, $14.99, $69.99, and $149.99 per month; two public subscriptions.
• RapidAPI listing-metric documentation: popularity is derived from request and user counts, while latency and service level use the latest 30 days. A popularity score is therefore not a subscriber or payment count.

The evidence is thin but readable. That is why this page is an offer test at $249 per month rather than an active paid endpoint or a revenue forecast. At the seven-day checkpoint, the page's reach and product-specific evaluation requests must be counted separately; our own probes and crawlers are never demand.

Scope, delivery, and cancellation

Provisioning would occur within one business day after the evaluation scope is accepted. A buyer could cancel before renewal; service would remain available through the paid period and then stop. No customer text has been accepted by this demand test, no checkout is active, and no contract is created by visiting or submitting the request form.

Request the $249/month evaluation scope   Download the public eval set