Skip to content
WORDS WORK
← back to the library
AI docsspec sample · fictional client, real writing

Docs for an AI API that admits it’s probabilistic

Client
Parse (spec)
Year
2026
Niche
AI docs

[ The brief ]

Parse extracts structured data from messy PDFs with an LLM under the hood. AI docs usually oversell certainty. The spec brief: document the model like the honest, slightly unpredictable colleague it is.

[ Craft decisions ]

  • →Confidence scores documented as a decision tool, with thresholds and actions.
  • →Prompt guidance shows a weak prompt and a strong one; diffs teach faster than rules.
  • →Non-determinism addressed head-on: what varies, what doesn’t, how to test around it.

[ The sample ]

Working with confidence scores

Every extracted field returns a confidence between 0 and 1. This is not decoration. It’s the API telling you how to route the result. In production systems we recommend three lanes: ≥0.95 auto-accept, 0.80-0.95 human review queue, <0.80 reject and re-request with a hint (below). Teams that skip the middle lane end up rebuilding it after their first invoice with a handwritten total.

Hints: prompting Parse like you mean it

Weak hint: “extract the total.” Parse will find a total, possibly the subtotal, possibly the previous balance. Strong hint: “extract the final amount due after tax, usually the largest currency value in the bottom third of the last page.” Location, relationship, and disambiguation: that’s the whole craft.

{
  "fields": [{
    "name": "amount_due",
    "hint": "final amount due after tax; bottom third of last page",
    "type": "currency"
  }]
}

What’s deterministic and what isn’t

Identical documents with identical hints return identical values at temperature 0, but confidence scores may drift ±0.02 between model versions. Pin a model version in production (model: parse-2026-05), test against our golden-document set before upgrading, and never assert exact confidence values in your test suite. Assert the lane, not the number.

[ Read it like an editor ]

“Assert the lane, not the number” is the kind of sentence support teams frame. AI docs earn trust by documenting uncertainty as a feature with a workflow, not a footnote.