Repository navigation
v0.5.1
- Fixed: a
uniqueItemscheck of a list too costly to hash is answered
as not checked before any item is hashed, so the answer takes no
longer than reading the list. v0.5.0 hashed such a list first, which
a list of a megabyte made take seconds on a slow machine.
v0.5.0 - 2026-10-06
Its tag was cut and its release pipeline stopped before it published, so its changes are released here.
-
Changed: a
uniqueItemscheck of a list of more than 20 items is
priced by the comparisons the validator makes, which are the pairs of
items whose hashes are equal, and no longer as if every item met every
other. A list of distinct objects is checked up to 65,535 small
items, where it was refused past about 500; items built to hash alike
are still priced past the bound and not checked. -
Added:
GET /readersmarks a reader that reads a page from the text
its file carries and calls no model withtext_layer: true, so a
console can keep it out of a picker of models while the routing
policy still tries it first. -
Fixed: an extraction no longer reads page furniture. Blocks of kind
page_header,page_footerandpage_numberare left out of the
text a model is given and out of what a citation may name, so a
running header is not taken for the document's title. -
Added:
PUT /parses/{parse}/fields/{name}asks an extraction again
under its name, with another schema or with the same one after it
failed. Only an extraction that has ended is asked again; one that is
pendingis409 conflict. It is reset topendingwith the new
request and nothing of its last result, the result it replaced is
removed, and its task is queued as a new extraction's. A name the
parse has no extraction of is asked as a new one. Migration
000007_ask_againadds a count of askings to an extraction. Pods of
v0.4.0 that are running keep working beside it; one that starts after
it does not come up, so roll every role.