Issues

Subscription and extraction job views ignore the active group; switching it should switch what is listed

openradx/radis#394 · opened

on Nov 10, 2026

Delete a subscription when its owner is removed from its group instead of pausing it

openradx/radis#393 · opened

on Nov 10, 2026

Add a Filtering feature: a job type that answers decision questions per report with an open-weight decision model instead of extracting values

openradx/radis#383 · opened

on Nov 3, 2026

Labelling gate answers the group's name instead of its question: NO for 70 % of abdominal studies

openradx/radis#373 · opened

on Oct 23, 2026

Reports edited through the ORM get a new tsvector but keep the old embedding

openradx/radis#370 · opened

on Oct 21, 2026

No CA-bundle setting for the LLM and embedding clients; internal TLS only works through an env var no code reads

openradx/radis#369 · opened

on Oct 21, 2026

The labelling scope query materialises millions of rows even for .exists(); the nightly guard took 3.6 s and the count 11.6 s on staging

openradx/radis#368 · opened

on Oct 21, 2026

Default EMBEDDINGS_WORKER_CONCURRENCY of 2 makes the first-run backfill take ~32 hours instead of ~7

openradx/radis#367 · opened

on Oct 21, 2026

Extraction CSV export turns negative numbers into text and omits the identifiers needed to join the rows back

openradx/radis#366 · opened

on Oct 21, 2026

Subscription jobs have no page or admin, so a failed refresh cannot be seen anywhere

openradx/radis#365 · opened

on Oct 21, 2026

per_page=0 raises ZeroDivisionError and negative values raise EmptyPage in the search view

openradx/radis#364 · opened

on Oct 21, 2026

SearchForm.clean() raises KeyError when the age validation fails, turning the form's own error into a 500

openradx/radis#363 · opened

on Oct 21, 2026

POSSIBLE labels surface as badges and filter hits although two thirds of them are wrong; some PRESENT labels contradict the report's negation

openradx/radis#362 · opened

on Oct 21, 2026

1

Any save of a label or gate question, including a no-change save, invalidates up to the whole corpus

openradx/radis#361 · opened

on Oct 21, 2026

Full-text ranking cost scales with the number of matching reports; common terms take seconds on every page

openradx/radis#360 · opened

on Oct 21, 2026

Silent fallback to FTS-only makes identical searches return different result sets seconds apart, with no indication

openradx/radis#359 · opened

on Oct 21, 2026

Cancelling an extraction job during preparation is overwritten: the status is written back from a stale in-memory object

openradx/radis#358 · opened

on Oct 21, 2026

EmbeddingBackfillRun counts 'doing' queue rows as live, so its stall detector never fires after a worker death

openradx/radis#357 · opened

on Oct 21, 2026

Embedding backfill overwrites reports edited while a batch is in flight (lost update, window = the whole batch)

openradx/radis#356 · opened

on Oct 21, 2026

1

Extraction output field names are not validated: a leading underscore drops the field silently, 'model_config' crashes every task

openradx/radis#355 · opened

on Oct 21, 2026

German compound words are not matched: single-word searches miss 26–66 % of the reports that contain the finding

openradx/radis#354 · opened

on Oct 21, 2026

Query sanitizer glues 'V.a. Blutung' into 'Va Blutung' and 'L4/5' into 'L45', losing 96–98 % of matches for common shorthand

openradx/radis#353 · opened

on Oct 21, 2026

1

Nested parentheses make the query parser exponential; a 21-character query pins a server thread for minutes

openradx/radis#352 · opened

on Oct 21, 2026

Subscription notification mail embeds patient ID, age, sex and study description in plaintext

openradx/radis#351 · opened

on Oct 21, 2026

Extraction output schema is built with nullable=False, so the model fabricates values for data the report does not contain

openradx/radis#349 · opened

on Oct 21, 2026

1

Navigation Menu

Repository navigation

Issues

Subscription and extraction job views ignore the active group; switching it should switch what is listed

Delete a subscription when its owner is removed from its group instead of pausing it

Add a Filtering feature: a job type that answers decision questions per report with an open-weight decision model instead of extracting values

Labelling gate answers the group's name instead of its question: NO for 70 % of abdominal studies

Reports edited through the ORM get a new tsvector but keep the old embedding

No CA-bundle setting for the LLM and embedding clients; internal TLS only works through an env var no code reads

The labelling scope query materialises millions of rows even for .exists(); the nightly guard took 3.6 s and the count 11.6 s on staging

Default EMBEDDINGS_WORKER_CONCURRENCY of 2 makes the first-run backfill take ~32 hours instead of ~7

Extraction CSV export turns negative numbers into text and omits the identifiers needed to join the rows back

Subscription jobs have no page or admin, so a failed refresh cannot be seen anywhere

per_page=0 raises ZeroDivisionError and negative values raise EmptyPage in the search view

SearchForm.clean() raises KeyError when the age validation fails, turning the form's own error into a 500

POSSIBLE labels surface as badges and filter hits although two thirds of them are wrong; some PRESENT labels contradict the report's negation

Any save of a label or gate question, including a no-change save, invalidates up to the whole corpus

Full-text ranking cost scales with the number of matching reports; common terms take seconds on every page

Silent fallback to FTS-only makes identical searches return different result sets seconds apart, with no indication

Cancelling an extraction job during preparation is overwritten: the status is written back from a stale in-memory object

EmbeddingBackfillRun counts 'doing' queue rows as live, so its stall detector never fires after a worker death

Embedding backfill overwrites reports edited while a batch is in flight (lost update, window = the whole batch)

Extraction output field names are not validated: a leading underscore drops the field silently, 'model_config' crashes every task

German compound words are not matched: single-word searches miss 26–66 % of the reports that contain the finding

Query sanitizer glues 'V.a. Blutung' into 'Va Blutung' and 'L4/5' into 'L45', losing 96–98 % of matches for common shorthand

Nested parentheses make the query parser exponential; a 21-character query pins a server thread for minutes

Subscription notification mail embeds patient ID, age, sex and study description in plaintext

Extraction output schema is built with nullable=False, so the model fabricates values for data the report does not contain