How do you get confidence scores from document extraction with an LLM?

The trap: a confidence field in your schema

Adding a `confidence` field to your schema feels like the answer, but the model just generates a plausible number — it has no measured signal behind it. See why LLM confidence scores are unreliable for the full explanation.

What actually works

Use a model that exposes its specialist encoder's output. Interfaze returns a `precontext` array containing the raw OCR result — per-word text, confidence, and bounding boxes. You map each extracted value back to those words to attach a real confidence and location to every field.

How to wire it up

Because Interfaze implements the OpenAI chat-completions API, migration is a base URL, an API key, and a model string — the request body is otherwise identical. The confidence and bounds come back in the response's `precontext` field, next to your structured `object`.

The honest tradeoff

Interfaze runs full OCR internally, so it consumes more input tokens and is slower — and usually costs more per call — than a raw general model. You pay that for an answer you can verify. If your task tolerates silent errors, you may not need it. If a wrong field is expensive — KYC, lending, insurance, claims — that is exactly what it's for.

FAQ