Eighteen ways to give a passenger the wrong answer.

Why it happens: ranking is not reconciliation. The flyer is short and entirely about the hike, so it often ranks first; nothing in top-k retrieval notices that a lower-ranked chunk says something else.

Why it happens: the cabin classes and the loyalty tiers share their names, Pioneer, Navigator and Pathfinder. To an embedding model "Pathfinder priority" is one idea, so a chunk about cabin benefits ranks as well as the one about the tier.

Why it happens: the documents define the two as "distinct in every way", but a passenger calls either one "my Orbital balance". A pipeline that retrieves only one of them answers for that one, with confidence.

Why it happens: a chunk that cuts through a table, or a model reading pipe-delimited rows, easily attaches a value to the wrong row. It is worse when the neighbouring row is on the same topic and both are marked firm.

Why it happens: "No EVA, no surface access" sits close to every EVA procedure in embedding space. If the procedure chunks outrank the exclusion, the generator fills in a helpful, wrong yes.

Why it happens: the free-cancellation rule is the best match for "how much do I get back", and it promises a full refund with total confidence. Only two words in it, "of booking", say which clock the 14 days run on.

Why it happens: the question never says "Ceres". The extraction terms list both caps by stop, and without the flyer that puts The Leap at Ceres the model has to guess which row applies. The first row, the inner system cap, is the easy pick.

Why it happens: top-k returns the best-matching paragraph. The credit option, the deadline, rebooking priority, the Points hold and refund timing sit in other sections and another document, and each scores lower on its own.

Why it happens: the answer is a subtraction that appears in no document, so retrieval can only supply its inputs. The same paragraphs give a round-trip figure next to each one-way figure, and a generator that takes the wrong pair still does the arithmetic perfectly.

Why it happens: no passage says a 2 pm group is cut short. The answer comes from putting the exit rule beside the tour length, and the deadline that sits closest to the question, the 30-minute muster, is the wrong one to measure against.

Why it happens: models round durations into familiar units, and "90 days" becomes "three months" before any counting happens.

Why it happens: the sheet states its scope at the top, several chunks away from the price table. The chunk with the number has no itinerary in it, so a passenger on another voyage gets a confident, wrong figure.

Why it happens: a four-letter acronym carries almost no meaning for an embedding model, so "stranded" pulls in the missed-departure passages instead. Keyword or hybrid retrieval finds the right one immediately.

Why it happens: the question shares no content word with the passage that answers it. "Put pen to paper" is "signed", "release form" is "waiver", and "the big untethered hop" is The Leap. Keyword retrieval has nothing to match, and an embedding model has to know the idiom. A guess lands on Ceres, where the hop happens.

Why it happens: the excursion descriptions are vivid and specific, so they outrank the dry sentence about layover length. The question never says which itinerary.

Why it happens: the list crosses a chunk boundary. Nothing tells the generator the list is incomplete, and two rules read like a finished answer.

Why it happens: the check-in card's table is cut in two. The Venus row lands in the second chunk, and the header that says which column is the deadline stays in the first. Retrieved alone, the row is a place and four numbers of minutes, with nothing to say which one is the muster.

Why it happens: a retriever always returns its k nearest chunks, even when none is relevant. USD 180 is Pioneer's excess-baggage rate per kilogram, the closest thing to a fare the corpus has.