Small decisions. Interesting possibilities.Submit contentSubmit

How it uses Jev

When benefit and category are missing but merchant and description are supplied, Formanator sends those two fields plus the valid benefit/category pairs returned by Forma. Jev chooses one pair or `no_match`; a missing match or confidence below 0.5 falls back to the configured language model. Receipt images stay with the vision-capable LLM path.

What Jev decides

Choose a claim categoryExample answers · not a recorded Jev response · Source ↗
Question 1 · category
YOUR APP
INSTRUCTION

Which Forma benefit and category pair best describes this expense claim?

STATE

Merchant: Open University. Description: MBA tuition fee. Forma returns two benefit/category pairs: Flexible Reimbursement Account / University Program, and Wellness and Lifestyle / Gym Membership.

JEV · CHOICE
  1. category_0_0
  2. category_1_0
  3. no_match

App workflow

  1. Supply claim context

    The caller provides the merchant and description; the command selects the Jev category provider when configured.

  2. Load valid choices

    Formanator builds one Choice option for each benefit/category pair returned by Forma and adds a `no_match` option.

  3. Ask Jev

    It sends only merchant, description, and those candidate pairs to TypeSafe System One.

  4. Validate and continue

    The code checks the answer type, confidence range, and selected option. A valid match at confidence 0.5 or higher is used; otherwise it falls back to the configured LLM provider.

Why it is interesting

The integration separates extraction from classification: Jev receives concise text and the live valid option set, then application code validates the selected key and applies an explicit confidence fallback.