Keenable raises $26M from Accel and Conviction. Techcrunch
Glossary
Plain-language definitions of the terms behind web search for AI agents, from vector search to nDCG.
A
Search run by an AI agent that plans queries, reads results and searches again until the task is answered.
B
When a benchmark’s test items leak into a model’s training data or are known in advance, so the score no longer reflects real ability.
C
Pulling the main text out of a web page’s HTML and leaving out menus, ads and other boilerplate.
How recently a search index last re-fetched its pages, and so how current its results are.
D
How up to date the results a search API returns are compared with the live web.
H
Search that combines keyword matching and vector similarity in a single ranking.
I
Whether a search engine builds its own index or returns results drawn from another engine’s index.
K
A search for one specific item that already exists, such as a paper found by its title or a court case found by its name.
L
Giving a language model retrieved sources so its answer can be checked against them instead of coming from training data alone.
Using a language model to rate results or answers instead of human reviewers.
M
Mean reciprocal rank: the average of 1 divided by the rank of the first correct result across queries.
N
A metric that scores how well a search engine orders its results, from 0 to 1.
Dropping pages that are almost identical to ones already kept, so an index or result list does not repeat itself.
Keenable’s open-source benchmark of web search APIs for AI agents, run on fresh queries every hour or day.
P
Search limited to what the web looked like at a chosen past date.
Q
Rewriting a search query into several variants to find pages a single query would miss.
R
The share of all relevant results that appear in the top k results.
S
Special commands, such as quotes or site:, that narrow and refine web search results.
Search that matches a query to documents by meaning rather than by exact keyword overlap.
T
Looking for an item the searcher has seen but cannot name, so the query describes it from memory instead of giving its title.
U
In NEEDLE, a synthetic engine that pools every engine’s results per query and sets the best score the whole field could reach.
V
Finding items by the similarity of their embeddings rather than by matching words.
W
A searchable copy of web pages, built by crawling, that a search engine queries instead of the live web.
A programmatic interface for querying a search index and getting structured results, so applications and AI agents can use live web content.
