What the generator decides

The generator reads Model._meta, counts a few rows, and fills in seven ModelAdmin attributes. Knowing which rule produced a line tells you which flag to reach for when you disagree with it.

Attributes it fills in

list_display

Every local field on the model, in declaration order, with fields inherited through a parent link left out.

list_filter

Fields that make sense as a sidebar filter. See Filters and raw ids.

raw_id_fields

Relations with too many rows to render as a select box.

autocomplete_fields

Many-to-many relations, which the admin renders as a search box rather than a multi-select list.

search_fields

Fields whose name appears in the search field names, name and slug by default. Extend the list with -s.

prepopulated_fields

Pairs like slug=name, emitted when both fields exist on the model. Extend with -p.

date_hierarchy

A single date field for the drill-down navigation above the changelist.

Filters and raw ids

Four field types are filter candidates from their type alone: DateField, DateTimeField, BooleanField and ForeignKey.

A foreign key is the interesting case, because a select box listing fifty thousand related rows is worse than no widget at all. The generator counts the related table and sorts the key into one of three outcomes:

Related rows

Outcome

Fewer than --list-filter-threshold (25)

list_filter

At or above --raw-id-threshold (100)

raw_id_fields

Between the two

Neither, because it is too big to filter on and small enough to pick from a list

Any other field also becomes a filter when it has few enough distinct values. The generator runs a bounded DISTINCT query per field and keeps the field when the count stays at or below the list filter threshold. Text, JSON, binary and file fields are skipped, because DISTINCT on those columns fails on PostgreSQL, Oracle and SQL Server.

Date hierarchies

A date hierarchy is cheap on a small table and slow on a large one, so the generator only emits one when the model holds fewer rows than --date-hierarchy-threshold (250). It then takes the first name that matches, preferring created_at, then updated_at, then joined_at.

Extend the candidates rather than renaming your fields:

./manage.py admin_generator blog -d published_at

Many-to-many relations

Every many-to-many relation becomes an autocomplete_fields entry by default. Autocomplete needs search_fields on the related model’s admin, so generate the related app too rather than only the one you started with.

Narrow it to specific fields, or turn it off entirely:

./manage.py admin_generator blog --auto-complete tags
./manage.py admin_generator blog --disable-auto-complete

With autocomplete off, a many-to-many relation may be emitted as a raw id field instead, which keeps the change form usable without a search endpoint.

Running without a database

Row counts drive the filter, raw id and date hierarchy decisions, so all three change shape when you pass -n:

./manage.py admin_generator blog --no-query-db

Every type-based filter candidate then becomes a list_filter entry, no field is promoted to raw_id_fields on size, and no date_hierarchy is emitted. This is the mode for a machine with no database access, and for a development database whose row counts say nothing about production.

Note

The counts come from whichever database the command connects to. Running against a nearly empty development copy is what produces a select box that turns into a thirty thousand row dropdown once deployed.