fjmatrix/flight-search-intent-eval

Flight Search Intent Eval

244 real flight search queries in English, German and Spanish, each paired with the structured search specification a model should produce from it.

This is the step before booking: turning one messy sentence into origin, destination, dates, duration, passengers, cabin and filters. That specification decides which itineraries a traveler is ever shown, so grading it is grading whether the model understood which flights they meant. τ-bench and similar benchmarks cover what happens after a search spec exists. This covers what produces it.

split	cases
en	150
de	50
es	44
total	244

Results

Current leaderboard across 9 models:

https://fjmatrix.github.io/flight-search-intent-eval/

A case

{
    "id": "ow-europe-seasia-mar",
    "lang": "en",
    "text": "Europe to south east Asia mid march",
    "tags": ["oneway", "fuzzy-location", "flexible-date"],
    "expect": {
        "action": "get_tickets",
        "search_type": "oneway",
        "trips": [
            {
                "departure": { "fuzzy": true, "names": ["Europe"] },
                "arrival": { "fuzzy": true, "names": ["Southeast Asia", "South East Asia"] },
                "departure_date": "03/10/26-03/20/26"
            }
        ]
    }
}

text is the query as a traveler wrote it. expect asserts only the fields that query actually determines — anything it does not mention is simply absent, and no grader is run for it.

Expected dates are written against an anchor date of 01/01/2026, which is also the date the model under test is told is today. Change one and you must change the other.

Tags

tag	cases
roundtrip	153	out and back, stated as either a return date or a length, never both
flexible-date	116	a window rather than a day
fuzzy-location	109	a region, a country, or a vibe — no single IATA code
oneway	59	no return stated
multi	12	two or more trips, each with its own date, in ascending order
invalid-date	7	a date that cannot be honored — already past, or no such day
invalid-others	3	reads as a flight query but cannot become a search

Grading

Eleven graders run per case. Each returns pass, fail, or not applicable — and not-applicable stays out of the denominator, so no model is rewarded or punished for a field its case says nothing about. A case passes only when every dimension it asserts passes.

grader	asserts
gradeAction gradeSearchType	search vs. invalid; oneway / roundtrip / multi
gradeOrigin gradeDestination	the right airports — or, for places with no IATA code, the right region
gradeDateRange gradeDuration	departure and return land inside the allowed window; the stay inside the allowed nights
gradePassengers gradeCabin gradeFilters	passengers, cabin, and the stops / price / bags / connection filters
gradeDateSanity	no departure before today, no return before departure, multi-city legs in order
gradeTripShape	trip count matches the search type; return_date and duration never both set

The last two take no expected value. They hold for any valid search, so every case is scored on them.

What keeps this from being brittle:

• Dates are windows. "03/15/26" demands that day. "10/01/26-10/31/26" accepts any day in October, because "somewhere warm in October" has no single right answer.
• Durations are ranges. "7" is exactly a week, "3-5" accepts 3 to 5 nights, "21-" is three weeks or more.
• Several airports can be right. {"any_code": ["ORD", "CHI", "MDW"]} — Chicago is Chicago.
• Regions are compared by meaning. "west europe", "beach towns in Southeast Asia" have no IATA code, so returned names are paired against expected ones by embedding cosine (≥ 0.80), one-to-one and order-free.

Running it

The harness, prompt, JSON schema and per-vendor adapters live in the GitHub repo:

https://github.com/fjmatrix/flight-search-intent-eval

npm install && npm run eval

Each run writes per-model JSON plus a self-contained HTML leaderboard.

License

MIT.