Moon Base 7
The Moon
The staging hub. Tranquility Base, the Lunar Golf Links and a rover tour, but only on the itinerary with a long enough layover.
Trap context
A test dataset for retrieval-augmented generation
A travel company built for things to go wrong.
Pelorus Space is 28 documents from Pelorus Voyages, an interplanetary travel operator that doesn't exist. Its policies, guides and brochures are written so that retrieval fails in known, named ways, and a right answer can be told apart from one that only sounds right.
Documents
28
PDFs · ~48,800 words
Frozen chunks
426
Stable ids at 1000 / 200
Failure types
18 shown
Conflicts, tables, negation, gaps…
Corpus version
2.1
Locked ·
pip install rag-connectorWhy a fictional company
No model has read these documents.
A language model can answer questions about a real airline from what it learned in training, so a correct answer tells you nothing about whether retrieval worked. Nobody has trained on Pelorus Voyages. Where The Leap waiver is signed, or which Navigator earns Points faster, is written only in these 28 files. A right answer had to be retrieved, and a wrong one points at the retriever or the generator, not at the model's memory.
The fiction is also deliberately detailed. The documents describe a working operator: one vessel, three itineraries, three cabin classes, a loyalty programme, and a stack of policies that refer to one another. Lorem ipsum can't produce a flyer that differs from the guide by one number, a rule that applies only to groups of eight or more, or a dated notice that overrides three older documents.
The world inside the corpus
Seven worlds, one vessel,
twenty-eight documents.
Every voyage stages through Moon Base 7, then runs inward to Mars and Venus, outward to Ceres and the Saturn system, or both. Each world has a long destination guide, and the signature excursions have their own flyers. Flyer and guide describe the same thing, and they don't always agree.
Mars Colony 1
Mars
Excursions to Cydonia, Hellas and Olympus Mons. The Caldera Rim flyer wants the fitness test passed 24 hours before the hike; every guide says 48.
Trap conflict
The Aphrodite
Venus
A floating resort 52 km up. The Glass Floor Tour (age 8) and the Gondola Descent (age 16) sit in neighbouring rows of the same table.
Trap table
Ceres Waystation
Ceres
The Leap: 100 metres up at 0.03g, with an age limit, a waiver that can only be signed back at Moon Base 7, and a platform that closes early.
Trap synthesis
Excursion Base Alpha
Titan
Ice sports at −179 °C. The muster tables give 75 minutes in the deadline column; the 95 that actually applies sits in the notes, because decontamination comes first.
Trap qualifying context
Saturn system
Enceladus
A flyover 25 km above the geysers with no surface access, stated plainly in a corpus full of EVA procedures.
Trap negation
What is in it
What an operator publishes.
Formal terms, long guides, and the short literature that restates them: flyers, fact sheets, a check-in card, an FAQ. The overlap is deliberate, as it is in any real document set, and so is the occasional disagreement. 39 Markdown tables hold facts, some of which appear nowhere else. Every document carries a publisher, edition and issue date.
| Kind | Docs | Covers |
|---|---|---|
| Terms and policy | 3 | Booking Terms and Conditions, the Orbital Points guide, and a dated notice that changes the companion discount |
| Passenger guides | 6 | Handbook, medical and fitness, packing, welcome aboard, departure planning, onboard charges for one itinerary |
| Travel guides | 8 | A complete guide for each of the three itineraries and each of the five destinations |
| Fact sheets and flyers | 9 | An at-a-glance sheet per itinerary, a flyer per signature excursion, family age limits |
| Card and FAQ | 2 | The check-in and muster card, and the FAQ page |
How RAG fails
Rarely with a blank.
Usually with confidence.
A retriever finds something relevant, and the generator writes a fluent answer from it. The corpus is built to exercise eighteen named failures. Six of them:
CONF · conflicting sourcesTwo documents disagree“How long before the Caldera Rim hike must I pass the fitness test?”The flyer says 24 hours, every guide says 48. Top-k picks one and never mentions the other.
DISAM · distractorA near-twin answers“I’ve booked a Pathfinder suite. Do I get early access to the Venus gondola?”No. Priority goes to the Pathfinder loyalty tier, a different thing with the same name.
TABLE · structured tableThe wrong row“Can my 12-year-old do the glass floor tour?”Yes, from age 8. The 16 in the next row is the Gondola Descent.
NEG · negation“No” retrieved as “yes”“Can I do an EVA at Enceladus?”No surface access at all. Every EVA procedure in the corpus pulls the other way.
QUAL · qualifying contextRight number, missing caveat“How much is the nightly service charge for a Navigator cabin?”USD 22, but only on the Outer Reaches Tour. The sheet says so once, far from the price.
UNANS · unanswerableThe corpus doesn't say“What does the Outer Reaches Tour cost?”No fares are published anywhere. The nearest number is an excess-baggage rate.
The corpus
Read the documents themselves.
The corpus is 28 illustrated PDFs with wordless artwork and ruled tables, the kind of mixed, overlapping document set a real operator publishes. Open any of them, then see which questions each one is built to get wrong.
Where the files live. RAG Connector contains the post-extracted and pre-chunked versions of this corpus (426 chunks with stable ids), which the Pelorus Query demo uses. This is corpus version 2.1 and it is locked.