User Management
Dashboard accounts. Admins manage users here and see the Health and Qdrant views.
Activity Summary
Key counts for the selected time window, across all users. The origin filter applies to the panels below, not to these cards.
Active Users
--
Conversations
--
In selected window
Messages
--
User → assistant pairs
Ratings
--
Reports
--
Issue reports
Errors
--
Application errors
Usage, ratings and reports
Ratings and reports count on the day the feedback arrived, not the day the question was asked.
OpenRouter Credits
Live balance and top spending models — spend is for this API key only, not the whole account.
Health & Freshness
Service probes and latest observed events across the system.
Model Deprecation
Provider-reported expiration dates for every configured model role.
Storage
Postgres database size and per-collection Qdrant footprint. Qdrant sizes are measured on disk from the qdrant volume (including segment/WAL preallocation); the logical vectors + payloads size is shown as secondary.
Health Incidents
Recorded outages of the upstream LLM and embedding providers over the last 30 days (newest first).
Activity by User
Top users ranked by total activity, with rating quality and error count.
Recent Questions
Each row pairs a question with its answer, rating, report, planner reasoning, and consulted documents.
Improvements
Open feedback and improvement ideas awaiting a fix. Add your own items with the plus button; mark an item as implemented once it is addressed.
Support Requests
Individual support requests users raised. A conversation may have several requests; each row is one request. Open a request to read the full transcript and mark it resolved. Add requests received by mail, phone or in person with the plus button.
Source:
Qdrant Projection
2D UMAP projection of the collection's vectors, colored by content type. Hover a point for metadata; search to highlight matches.
UMAP-1
UMAP-2
2D UMAP projection of embeddings — points close together are semantically similar; axis units (UMAP-1 / UMAP-2) are arbitrary.
Benchmark Cases
Question set the benchmark runs against: reference answer, up to 5 expected documents, the expected agents and an optional thematic category. Edit rows in place, or import/replace the whole set from the Excel format. Collections group cases into named subsets a run can be scoped to.
Agents:
Run Benchmark
Each case runs against every listed model (up to 3), repeated per the configured repetitions, scored 0–10 (answer quality minus routing and document penalties). Repetitions and penalties live in config.yml. Runs execute in the backend; progress appears live below as a per-case results table.
Score Distribution
Per-model final scores (0–10) with one dot per scored unit colored by case category. Hover a dot for details, click to open the full result, toggle categories from the legend.
Run History
Past benchmark runs, newest first. Open a run to restore its full per-case results; delete removes the run and its results.
Digest
Ask about chatbot usage in plain language — answered live from the database, read-only.