Feedback

The Feedback page in the admin console is where every user-submitted comment about your Digital Twin shows up. It is the qualitative signal layer on top of Histories — Histories tells you what was said, Feedback tells you what users thought of it.

​

What feedback collects

Pria captures three kinds of signal whenever a user reacts to a Digital Twin response: Each feedback row is associated with a user, an instance, a timestamp, and (when the user submitted feedback from inside a conversation) a link back to the originating message in Histories.

​

Where to find it

Sign in as an admin and open Admin → Feedback from the side menu. The page lists every comment for instances you have permission to see, newest first.

​

Filtering and sorting

The filter bar lets you narrow the list to exactly the feedback you need to act on:

• Instance — restrict to a single Digital Twin instance (defaults to your home instance)
• Account — when you administer multiple instances, scope to a whole account at once
• User search — by email, first name, or last name
• Date range — last day, last week, custom range; useful when isolating regression after a configuration change
• Sort — newest first by default; switch to oldest first when working through a backlog

Filters persist as you paginate so you can review a large window without losing your place. Use the Reset affordance to clear all filters at once.

​

Drilling into a feedback row

Click the pencil icon on any row to open the feedback detail panel. You will see:

• User — name, email, account type, and link to their profile
• Instance — the Digital Twin instance the user was talking to
• Comment — the full text the user submitted
• Rating — thumbs-up or thumbs-down, when applicable
• Conversation link — opens the source conversation in Histories at the exact message so you can read the surrounding context
• Model used — which AI model produced the response being reviewed (useful when correlating spikes in negative feedback with a recent model change)
• Created at — exact timestamp

If the conversation has since been compacted, the original turn is still preserved — see the Compaction section in Histories.

​

Responding to feedback

You have several options once you have read a piece of feedback:

• Reach out to the user — the panel surfaces the user’s email so you can reply directly. For sensitive or complex issues, a personal email beats anything automated.
• Share with your team — copy the conversation link and circulate it for triage.
• Mark as resolved — update the feedback’s status once you have acted; resolved items can be filtered out of the active view.
• Delete — soft-delete (status → deleted) when the comment is spam, abusive, or accidentally submitted. Deleted rows are retained but hidden from the default view.

When the issue is operational rather than product-level (an outage, an upstream provider failure), escalate to the Praxis AI team at humans@praxis-ai.com with the conversation link.

​

Patterns to look for

Single thumbs-down events are noise. The value of the Feedback page is in patterns. Things to watch for:

• Sudden drop in ratings — a wave of thumbs-down within a few hours usually signals a configuration drift: a new instruction, a model change, or a retrieval source that has gone stale.
• Same complaint, multiple users — three users in a day all reporting “it forgot what we were talking about” is a memory or compaction signal, not three coincidences.
• Model regression — if you switched the instance’s default model recently and negative ratings climbed afterwards, the rollback is your first move.
• Prompt drift — when the Digital Twin starts answering off-topic, the Instructions field is the place to look; small edits sometimes have outsized behavioural effects.
• Topic clusters — if multiple users complain about the same subject (a specific course, policy, or feature), feed that back into your Onboarding Questions or Assistants.

​

Using feedback to refine Instructions and Assistants

Feedback is most valuable when it changes behaviour. Two tight loops:

1. Instructions tuning — when feedback consistently shows the Digital Twin missing a constraint (“don’t recommend third-party tools”, “always cite the source”), add a line to the Instructions field in Configuration. Re-test with a fresh conversation.
2. Assistant scope refinement — if users keep complaining that the wrong assistant answered their question, tighten the assistant’s description so the dispatcher picks more accurately. See Assistants.

Tie each change back to the feedback that prompted it (commit notes, a shared doc) so you can measure whether the next week of feedback improves.

​

Privacy considerations

Feedback rows always include the submitting user — there is no anonymous channel inside Pria today. Consider the following when reviewing:

• Treat comments as personal data. FERPA / GDPR principles apply: don’t paste raw feedback into public channels.
• Identified by default. Users see their own name attached when they submit; they should expect a human to read it.
• Consent for outreach. If you plan to email a user about their feedback, it is courteous to keep the reply scoped to the issue they raised.
• Cross-instance exposure. Admins of an instance see feedback for that instance only. Account-level admins can see across the account.

​

Bulk export for analysis

For longitudinal analysis, export the visible feedback set as CSV from the page toolbar. The export includes the rating, comment, user email, instance, model, and timestamp. Common downstream uses:

• Pivot tables in a spreadsheet (positive vs negative split by week)
• Sentiment classification in your analytics pipeline
• Sharing a redacted slice with the product team during weekly reviews

For richer analyses (linking feedback to retrieval sources, joining across multiple instances, slicing by model version), open conversations directly in Histories and use its export there.

​

Related

• Histories — the underlying conversation each feedback row points back to
• Sessions — who was active when the feedback arrived
• Configuration — where you tune the Instructions feedback most often calls out
• Assistants — refining assistant scope based on feedback patterns
• Entitlements — the feedbacks.list, feedbacks.edit, and feedbacks.delete permissions that gate this page