[source]
This constructs a table by extracting rows from the complete data set whose column entries
match specified values.
Columns are referenced using the first few letters of their name followed by a colon (:)
and then the value to match.
For example, the filter desc:raspberrypi would extract all rows having
the string raspberrypi in their Description column.
You can use logical operators (&&, ||, and !) and comparison operators (<, >, ==) along with parentheses for grouping. For example, this would filter for repos having combined stars and forks greater than 100 and a size of at least 10000:
(stars: + forks: > 100) && size:>=10000
This is for "quick-and-dirty" searching. It will do a live search of all columns for the text you enter. As you type in this field, rows of filtered data that have no matching text in any column will be hidden, leaving only matching rows.
Describe what you are looking for in plain English and press <Enter>.
Each repo's description and your query are converted into vectors by a small language model
that runs entirely inside your browser, and the repos are ranked by how closely their
vectors align with yours. Unlike Filter and Search, this matches on meaning
rather than on literal text, so cheap board that talks to WiFi can turn
up a repo whose description never uses any of those words.
The first search on a set of repos is slow. It downloads the language model (about 25 MB) and then computes a vector for every repo in the table, which runs at roughly a thousand repos per minute. Progress is reported to the left of this icon. The vectors are then cached in your browser, so every later search over those same repos returns immediately.
Only the repos currently in the table are ranked, so use Filter first to cut a large topic down to a workable size before searching.
The results appear in a sortable Rank column at the left of the table, with rank 1 being the best match. Emptying this field, or changing the filter or the topic, clears the ranking.
