‹ Culture 3.14 · What we ask ourselves
Question · Information
How does a system know it has found enough evidence?
A system can retrieve ten relevant passages and still have nothing to support the claim it is about to write.
Why it matters
Retrieval measures similarity, not sufficiency. A search engine returns whatever is closest to the question, even when the closest thing is not enough to answer it.
When the system writes the answer anyway, the result is correct in form and unfounded in substance. It is the hardest failure to spot by reading the output.
What we know so far
It works reasonably well to require every claim to be anchored to a specific passage, and to discard the claim when the anchor does not appear. It also helps to split the question into sub-questions and check the coverage of each one separately.
Score thresholds, on the other hand, have behaved worse than we expected: they shift with the corpus and with the language.
What remains open
We do not know how to distinguish “there is no evidence” from “the evidence exists but we have not retrieved it”. For the reader, both situations produce the same abstention, and they do not deserve the same answer.
It also remains open how much evidence is enough when the question asks for a synthesis rather than a fact.