The Configuration screen contains tabs where administrators can define prompts that are automatically injected into Compass's automated workflows. This allows organizations to customize AI behavior across scans without modifying individual diagnostic or definition settings.
Accessing Configuration
Navigate to Configuration from the Compass sidebar. The link is visible to users with Compass Author-level access or higher. The screen has six tabs: Global Prompts, Scan Presentation Guidance, Entities, Articles, Usage, and Library. The Entities and Articles tabs are documented on their own pages.
Scan Presentation Guidance
The Scan Presentation Guidance tab lets administrators create and manage named presentation prompts. When creating a scan, users select one of these prompts to control how the AI builds the scan's automatic presentation.
The tab displays a list of all prompts, showing each prompt's name and a preview of its text.
Creating a Prompt
Each guidance prompt has two fields:
- Name — A short label for this guidance (e.g., "Clinical Leadership Focus" or "Cost Trends Only").
- Prompt — The text that will be appended to the AI's starting prompt when building a presentation.
Deleting a prompt removes it from the scan creation dropdown, but a deleted prompt already selected on a running or completed scan will still be used for that scan's presentation. Deletion only prevents future selection.
How It Works
When creating a scan, the user selects one of these prompts from the Presentation Guidance dropdown, and the choice is saved on the scan. Once the scan finishes all of its diagnostics and begins building the presentation, the saved prompt is appended to the AI's starting instructions. If no guidance is selected, the default scan presentation behavior is unchanged.
Example Uses
- Direct the AI to emphasize specific themes (e.g., "Focus on cost trends and utilization patterns").
- Request specific chart types or layouts (e.g., "Use bar charts for comparisons across entities").
- Set a preferred tone or audience (e.g., "Write for a clinical leadership audience").
- Exclude certain topics (e.g., "Do not include pharmacy analysis in the overview slides").
Global Session Addendum
The Global Session Addendum field on the Global Prompts tab accepts free-form text that is appended to the very end of the system prompt for every Compass session — after the starting request from the user. This provides a way to inject organization-wide instructions that apply to all AI investigation sessions.
Override Behavior
The Global Session Addendum can be overridden at two levels. At both of these narrower scopes the field is labeled simply Session Addendum, since the value applies only to that scan or session rather than to the whole organization.
- Scan scope: When creating a new scan, a "Session Addendum" textarea is available. The value defaults to the global setting but can be edited. Every session in that scan will use the scan-level value instead of the global default.
- Session scope: When starting an individual session (outside of a scan), a "Session Addendum" textarea appears below the main question field if a global value has been configured. The user can edit it before starting the session.
If an override is provided (at either scope), it replaces the global value for that session. If the override is cleared, the session will fall back to the global setting.
Example Uses
- Enforce response formatting (e.g., "Always include date ranges in the finding's title").
- Set guardrails (e.g., "Do not make clinical recommendations").
- Require specific analysis steps (e.g., "Always compare against the benchmark population").
How It Works
When a session starts, Compass looks for a per-request override from the scan or the session form and falls back to the global setting if there is none. Whatever it resolves to is appended as the final section of the system prompt, after the user's starting request. If the field is empty and no override is provided, no additional text is appended.
Scan Followup Transition Guidance
The Scan Followup Transition Guidance field on the Global Prompts tab accepts free-form text that is injected when a user sends their first follow-up message to a completed scan session. This provides a way to give the AI transition context — for example, informing it that the investigation phase is over and it is now responding to interactive questions from the user.
How It Works
When a user sends a follow-up message to a scan session that has already completed, the guidance is injected as a system message ahead of their question. It is injected only once per session — subsequent follow-ups do not repeat it. If the field is left empty, no transition guidance is injected.
Example Uses
- Inform the AI that it is now in a follow-up conversation rather than an autonomous scan ("You are now responding to follow-up questions from the user about your earlier investigation.").
- Provide guardrails for follow-up interactions ("Do not re-promote findings unless the user explicitly asks you to.").
Usage
The Usage tab provides an approximation of inference cost by showing the total character count of AI conversation logs for a selected month. Character counts are drawn from the conversation logs stored in the client database (session logs and presentation logs).
Reading the Usage Table
Select a month from the dropdown. The table shows one row per user who owned sessions or presentations during that month, with character counts broken out by type:
- Presentation — Characters from presentation conversation logs.
- Scan Session — Characters from sessions that belong to a scan-originated diagnostic.
- Explorer Session — Characters from sessions that belong to an Explorer diagnostic.
- Other Session — Characters from sessions that belong to other diagnostics (manually created).
- Total — The sum across all types for that user.
A totals row at the bottom sums each column across all users.
Notes
- Usage is attributed to the owner of the session or presentation, which may not reflect the actual user who drove all of the conversation (e.g., in shared diagnostics).
- Deleted sessions and presentations are included, since their logs were still generated and consumed inference.
- The month filter is based on when the session or presentation was created, not necessarily when it was used. A session created in one month but used in the next will appear under the creation month.
- Character count is a proxy for inference cost, not an exact measure of tokens consumed.
Library
The Library tab compares canonical diagnostic definitions, explorer personas, and articles maintained by Ursa Health against the definitions, personas, and articles saved in the local database. This lets administrators see which items are up to date, which are missing locally, and which have diverged from the canonical version.
The library is organized into three sections:
- Diagnostics — Grouped by domain and sorted alphabetically within each group.
- Explorer Personas — Listed below diagnostics without domain grouping.
- Articles — Listed below explorer personas, sorted alphabetically.
Each row shows the definition's name, description, and a status indicator:
- Match (green check) — The local definition is identical to the library version. No action is needed.
- Missing (orange warning) — The library definition does not exist locally. Click Import to create it.
- Diverged (yellow branch) — The local definition exists but differs from the library version. Click the expand chevron to view a line-by-line diff of the differences. Once expanded, click Overwrite from Library to replace the local definition with the library version.
Syncing a Definition
Importing a missing item creates it locally — for a definition, with the library's configuration, sessions, and metadata; for an article, with its name, summary, and body.
For a diverged item, expand the diff panel first to review the changes. The diff shows lines added in the library version (green) and lines removed from it (red). Overwrite from Library then replaces the local item entirely with the library version.
Limitations
The Library tab does not track whether local changes were intentional customizations or whether divergence is due to a library update. It performs a full-text comparison — any difference, however small, results in a Diverged status. The Overwrite from Library action replaces the entire local definition; there is no selective merge.
Access Control
- Authenticated users can view the global prompts and scan presentation guidance prompts.
- Compass Authors (and above) can update the global prompts, create/edit/delete scan presentation guidance prompts, view the Usage tab, and use the Library tab to import or overwrite definitions.