Known-item search
Known-item search is a search where the person or program looks for one specific item that already exists and is known to them, such as a paper by its title, a court opinion by its case name or a filing by its number. A search counts as successful only if that exact item appears in the results, usually near the top.
The term comes from library science: Don Swanson’s studies of library catalogs (1968–1972) described it as a search for a specific work that the user knows to exist. AI agents run the same kind of search when they look up one particular source, such as a paper or a court case. In NEEDLE, the Scholar and Legal verticals and the filing lookups in Finance are known-item tasks, scored by matching identifiers such as an arXiv ID, a legal citation or an SEC accession number.
Example
An agent has to cite the paper titled “Attention Is All You Need”. It searches for that exact title. The search succeeds if the paper’s own page is among the results. A list of other good papers about transformers counts as a miss, because the agent needed this one.
Related terms
Written by Keenable.
Published
