TraceDogTraceDog
ExperimentsOpen sourceAboutContactGitHub
DashboardRead docs
← Traces / Detail

Trace decision

Needs review

Grounding is borderline and should be reviewed.

Grounding0.17Weak
Risk0.83High
Reliability0.29Low
Recommended actionReview retrieved evidence before shipping
gpt-4o-mini1766ms totalNormans2026-03-26T04:47:22.190408+00:00
Why it scored this way
Why this needs review

Retrieved chunks align weakly with the response after answer-aware scoring (best hybrid 0.17 < 0.35).

  • Best grounding score: 0.17
  • Sentence match: 0.13
  • Keyword overlap: 0.25
Full scorer narrative

Retrieved chunks align weakly with the response after answer-aware scoring (best hybrid 0.17 < 0.35).

Likely causes

  • Retrieval returned irrelevant passages.
  • The answer uses wording that still does not match any sentence closely.

What we measured

  • Hybrid grounding (per chunk): best 0.17 (strong ≥ 0.52)
  • Mean hybrid across chunks: 0.17
  • Best raw sentence match: 0.13 (short answers often score low vs. whole paragraphs)
  • Lexical overlap with sources: 0.25
  • Blend: standard blend (70% best-sentence + 30% keyword)
  • Weak / strong cutoffs: 0.35 / 0.52
Evidence
Prompt & response
Grounding vs thresholds0.17
weak < 0.35review 0.35–0.52strong ≥ 0.52
Blend contribution
Sentence 0.13 (70%)Keyword 0.25 (30%)
Confidence trend

Series padded from this trace’s score — batch view shows drift.

Failure mix (this trace)

One bar — fleet-wide % needs aggregate metrics from the API.

  • Hallucination risk 0%
  • Low grounding 0%
  • Failures 100%
  • Healthy 0%
Execution runtime

Total 1766ms

Retrieval441ms✓
LLM1325ms✓
  • ✓Retrieval441ms
  • ✓LLM1325ms
Debug