Eval Dashboard · runs in CI on every push

Grounded, or it doesn't ship.

The agentic-rag-mcp golden dataset runs through the real pipeline (plan → retrieve → synthesize → self-critique) on every push. A groundedness, citation, or refusal regression fails the build. Source →

20/20
CASES PASSED (100%)
0
FAILED
3m 57s
LAST CI RUN
real
PIPELINE · NOTHING MOCKED
✓ citation presence ✓ groundedness · LLM-as-judge ✓ refusal correctness ✓ latency

Answerable12

grounded answer, cited [n]
PASSdefinition of groundedness
PASSworst failure mode
PASScorrect behaviour on missing context
PASScitation format used by this project
PASSgolden dataset definition
PASSgolden dataset sizing
PASSthree case categories
PASSregression-capture rule
PASSjudge vs deterministic checks
PASSjudge calibration
PASSCI-scored dimensions
PASSCI behaviour without API keys

Refusal5

out-of-corpus → declines
PASSgeneral-knowledge question absent from corpus
PASSpricing question absent from corpus
PASSsports question absent from corpus
PASStooling question absent from corpus
PASScosts question absent from corpus

Adversarial3

false premise → corrected
PASSpremise inverts the refusal principle
PASSpremise freezes the golden dataset
PASSpremise merges judge and system model
Latest green run · commit fcea396 · pgvector service in CI View the live CI →