Question 1

Why one yes/no question per pair instead of one multiple-choice per column?

Accepted Answer

The README says a multiple choice must pick something even when nothing fits and hides ties, while a matrix of booleans lets a column map to nothing at all and the per-column guard question surfaces columns that were dropped only because no pair cleared the threshold.

Question 2

Do I need an API key to try the demo?

Accepted Answer

Only for stage 2. Without an OpenRouter key the demo still starts and the deterministic stage still runs, with everything that needs judgement returned unmapped, which shows what a purely deterministic importer gives users.

Question 3

What does one real mapping run cost?

Accepted Answer

A 23-column warehouse export against a 10-column orders table mapped 10 of 10 columns with 253 questions in one call, taking 915 ms and $0.0012, measured against jev 1.13 through OpenRouter. The README publishes a per-file table across nine sample exports rather than a single best number.

Question 4

What does the model actually see?

Accepted Answer

The state carries your destination schema with column name, type and a description of what it holds, plus each incoming column name and three example values trimmed to 100 characters. The descriptions are what separate neighbouring fields, so work_email beats personal_email because it says explicitly not a personal one, and the example values often do more work than the names, since Shopify's Name column goes to order_ref at 93% because its values are #1042 and not Marit Halvorsen.

Question 5

How do the two stages fit together, and what threshold applies?

Accepted Answer

Stage 1 does typography and nothing else, locking strict equality on normalized case, accents and separators for free, instantly and never wrongly, which settles 7 of 10 columns on a clean HubSpot export before any network call. What is left goes to stage 2, then a greedy one-to-one resolution above MATCH_THRESHOLD, default 0.75, returning mapping, suggestions, ignored, matrix and orphan. First Name and first_name lock; Company Name and company do not.

Question 6

What comes back, and how do I wire it into my own product?

Accepted Answer

mapping entries carry method and p, where an exact lock has no probability and faking one tells the user nothing; suggestions rank every unmapped destination; ignored is a visible list rather than a silence; and orphan flags the ignored columns the model thinks do belong somewhere. lib/jev.mjs is the only file that talks to a service, so you pass your own decide function taking state, questions, model and signal, which is why npm test needs no key or network. Keep your own parser and build [{ name, samples }], three samples per column.

Question 7

What do the nine sample files each test?

Accepted Answer

Three per destination table, and each exists to break something. The HubSpot export is the easy deterministic case, the Eventbrite file is missing and extra columns plus Attendee being a full name where the destination wants first and last, and the French CRM tests two emails at 87% against 54%. On the e-commerce side the Shopify Name trap makes Name the order number rather than a person, and on HR an identity provider export pits personal_email against company_email.

Jev Table Import Mapper

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community