Manage users, app settings, and view usage statistics
The code is shown once, here. Only its hash is stored, so it cannot be read again afterwards.
| Name / Email | Role | Status | Joined | Actions |
|---|---|---|---|---|
| Loading... | ||||
Configure OpenID Connect for SSO with Azure AD, Okta, Keycloak, PocketID, Google, etc.
Callback URL:
Environment variables take precedence over DB settings. Storing password in DB is convenient but consider using env vars for production security.
1. Discover & test a model the gateway offers and add it → 2. Roster is everything added, with the default and on/off switches → 3. Availability is which roster models each feature may use.
Press Search gateway to list what it offers for this kind. The box filters by name.
Press Test on a model above or on the roster, or type an id the gateway does not advertise.
Press Test on a model above, or type an id the gateway does not advertise. A picture is generated, so this costs one request.
Used wherever nobody chose a model explicitly. Only enabled models can be the default.
Loading...
Added with + Add under Discover & test. Under Availability, tick one to offer it in chat and pick one per workflow.
Loading...
Speech and transcription are not on this roster. There is one voice and one model for each, chosen with Make default under Discover & test.
Only models on the roster appear here. If one is missing, add it under Discover & test first. One Save availability at the bottom applies everything in this card.
Leave every box empty to offer only the chat model above.
Leave every box empty to offer only the workflow's own image model.
The model that writes a deck never sees it, so it cannot tell that a slide overflowed or that a long list wants two columns. With a reviewer chosen, each generated deck is rendered, looked at, and its layout corrected once. Wording and figures are never changed — a review that alters them is discarded.
Needs to be able to see images. One image per slide is sent on every presentation generated, so this costs money each time — a cheap capable vision model is the right choice here, not the best one available.
Display only. The AI prompt, retrieval and the stored answer are byte-for-byte identical either way, so turning this off cannot change what the model says. Citations are hidden, not removed — turning it back on restores them.
UTF-16 code units. Leave empty to use CLINICAL_ASSISTANT_CONVERSATION_CHARS.
This sends text outside the building. When a resource author turns it on, the model may send a short search query to the provider below. Only the query leaves — never library excerpts, never the generated content, never anything about the user — but a query written while drafting clinical material can still carry clinical detail, and the provider keeps its own logs. SearXNG is the only option here that you host yourself.
Not needed for SearXNG. Shown masked once saved.
Only used when the provider is SearXNG. Its /search?format=json endpoint must be enabled.
A separate source from the web providers above, and a separate tick box for authors. It returns structured records — title, journal, year, PMID — so a reference can be exact rather than reconstructed from a page title. Queries go to NCBI, not to a commercial provider.
Optional. PubMed works without one; a key raises the rate limit from 3 requests per second to 10, which matters for bulk work rather than for one person writing a resource. Create one at account.ncbi.nlm.nih.gov/settings — sign in, then API Key Management → Create an API Key.
NCBI asks callers to identify themselves so they can contact you before blocking you. Optional but courteous.
Answers where the assistant wrote a citation number that no returned source matched. Those markers are never turned into links, so a reader sees plain text — this is where they are recorded. Consistent entries usually mean retrieval is returning fewer sources than the answer assumes.
Reset all announcement, feature flag, and email settings back to factory defaults. SMTP configuration and custom models are preserved.