Online evaluation
Every live answer is scored automatically after it is sent: four RAGAS metrics, answer time and cost. Scores range 0–1, colour-coded: ≥ 0.85 0.70–0.85 < 0.70. Hover any metric name for its definition.
Summary
Answer time
How long engineers waited for an answer, measured at the API (not the scoring time).
Cost per answer
From the usage ledger.
Daily average (last 30d)
Knowledge Gaps
Topics users asked about that FIKA could not answer confidently, clustered by phrase. Read-only — uploading documents that cover the top topics closes the gaps.
| Topic | Count | Sample question | Last seen |
|---|---|---|---|
| Loading… | |||
Answers given with no signal (last 30 days)
What the device's offline guards decided, from each device's turn log after it checks in: the share of turns per verdict, for tuning the thresholds. No question text unless set below.
| Verdict | Turns | Share |
|---|---|---|
| Loading… | ||
Metric health — automatic diagnosis
Recent evaluations
| When | User | Question | Faithfulness | Answer relevance | Cited precision | Retrieval relevance | Answer time | |
|---|---|---|---|---|---|---|---|---|
| Loading… | ||||||||
Offline evaluation
Each release is scored against a fixed golden dataset of questions with known answers, and compared with the previous release against quality gates.
No release scorecards recorded yet.