Command line

Run inside the backend container (docker compose exec backend cara ...) or locally (see Development). Every command runs in one transaction: if it fails, nothing it did is saved, and the error says what to fix.

Command	Does
cara user-add <name> [--password P] [--superuser]	create an account; the password is prompted for when not given
cara dataset-add <folder> [--name N] [--path P] [--encoder E]	register a dataset: images, query sets, embeddings in index/ (see Datasets for the options)
cara datasets	list datasets: images, query sets, encoders
cara encode <dataset> --encoder E [--batch-size 64] [--num-workers 4]	embed the images and text queries that have no E embedding yet (needs the encode extra; use a GPU for large models; see Encoding)
cara study-create <cara.yaml> --dataset D [--query-set S] [--queries N] [--seed 0] [--member <name>]...	create a study on a dataset: sample the queries, build the pools, group them, add the annotators
cara load <folder> [--image-dir D] [--member <name>]...	dataset-add and study-create for a single-folder study; a dataset of the same name is reused
cara member <study> <name> [--role admin\|annotator\|adjudicator]	add someone to a study, or change their role
cara pilot <study> --count N [--seed 0]	choose pilot queries (assign them to annotators in the study's Pilot tab)
cara pilot <study> --count N --labels QRELS --annotator NAME [--then NAME...]	record a pilot from known labels (drawn from the queries in the file); --then assigns the other queries to the real study's annotators
cara judge <study> --labels QRELS --annotator NAME [--first N]	record a user's judgments of their first N queries from known labels (demos, tests)
cara simulate <study> [--qrels FILE \| --pilot] [--max-queries 30]	the settings simulator, with common settings and the typical annotator; prints a table
cara build-pools <study>, cara cluster <study>	rebuild a study's pools or groups (before judging starts)