The vault follows a reference the question never named
A vault's answer is only as wide as what one search can reach, and some
questions are not shaped to be reached in one search. "Who took custody of
the paper arguing that the cure for faction is to extend the sphere?"
names an argument and a person, and the only thing joining the essay to
the custody register is a reference code that appears in neither the
question nor the asker's vocabulary. The join key is in the first result,
not in the query — so no amount of ranking, and no larger k, gets there.
So when a first search surfaces a reference code the question did not
contain, the vault searches again for that code, and the answering model
receives both halves.
Three properties are load-bearing. The second search is issued as the same
person, so it resolves the same compartment predicates as the first and
can never reach a document the asker could not already read — a hostile
document that plants a code buys itself a search over the asker's own
vault, not a disclosure. A question that surfaces nothing to follow pays
nothing at all: no extra search, no extra model call, no extra wait. And
the extraction is done by pattern rather than by asking a model what to
look up next, because a model asked that question is one more surface for
a document to give orders through.
One limit is worth knowing before you rely on this. It happens when you
ask through Think, or through the vault's tool interface — not on the
query page, where the model is shown exactly the passages you retrieved
and can see for yourself, and where adding a search on the server would
put text in front of it that you were never shown. So the same two-part
question follows a reference on one surface and not on the other.
A second limit is deliberate: only reference codes are followed, never
people's names. A name has no shape that separates it from any other pair
of capitalised words, so following names would mean following noise.
That the second search can never reach a compartment the asker could not
already reach is asserted against a real vault rather than here — a
scenario that mocked a compartment boundary would prove nothing about the
one that exists.
An answer spans two documents the question could not name
- Given an essay that carries a reference code
- And a register that records who received that code
- When a person asks a question the essay answers only half of
- Then the answering model receives the register as well as the essay
A question with nothing to follow costs nothing extra
- Given a document that carries no reference code
- When a person asks a question that document answers on its own
- Then the vault searches once and not twice
An administrator can turn the second search off
- Given a deployment whose second-hop bound is set to zero
- When a person asks a question that would otherwise follow a reference
- Then the vault searches once and the prompt is what it was before