Structured Extraction LLM → validated JSON

Fields below 0.75 confidence land here, with the reason. In a real deployment these would block the write to your accounting system or CRM until someone approves or corrects them. The queue is stored in this browser.

1. Schema first. Each document type has a Zod schema: required fields, ISO dates, 3-letter currency codes, numbers without separators, at least one line item.
2. Extract. A free LLM (via OpenRouter, rotating models if one is busy) returns {"data": …, "confidence": {field: 0–1}} in JSON mode.
3. Validate and retry. If the JSON doesn't parse or fails the schema, the exact errors go back to the model ("invoice_date: must be YYYY-MM-DD") and it tries again, up to 3 attempts. Every attempt is shown.
4. Check. The model's confidence is only a starting point. Code checks the arithmetic (quantity × price, subtotal, total), that invoice numbers, emails and amounts actually appear in the text, and that dates make sense. A failed check caps that field's confidence.
5. Review. Fields under 0.75 go to the review queue with the reason. Latency, tokens and cost for every run are in the log (the models are free, so cost is $0).

Source code and write-up: github.com/yashharkawat/llm-extract. Sample documents are made up.