Page 5 of 8~112 min topic

RAG quality audit

Debug scoring fluency instead of citation support in the RAG quality audit report

Page 5 reproduces and repairs the characteristic failure of the versioned RAG quality audit report: scoring fluency instead of citation support, or mixing docs versions in one report.

~14 min this pageDebugging

1Learn the idea

Read

Reproduce before you repair

Do not start with a speculative fix for the RAG quality audit report. Force the failure on purpose, save the before output, then change one cause at a time. Retries are allowed only for transient conditions—not for bad input that will fail forever on rag-quality-audit.

Read

Force the failure

fluency_score=0.99; citation_precision=0.5
if fluency_score>0.9 and citation_precision<0.8:
    print('FAIL: fluent but weakly cited')

Expected evidence: FAIL: fluent but weakly cited. If you cannot reproduce on demand, you do not yet control the failure mode for rag-quality-audit.

Read

Repair with a reviewable diff

After repair, rerun the exact reproduction command. Keep the failing fixture as a regression seed for the observability page. For the versioned RAG quality audit report, remember the claim you are restoring: audit records hit/miss per question with docs_version and fails release under threshold.

Read

Lab notebook: reproduce on command

Store a one-command reproduction for: scoring fluency instead of citation support, or mixing docs versions in one report. The command should use gold set with expected evidence IDs + candidate run output or a minimal mutant of it. Paste the failing output into notes/failure-before.txt (or your shell scrollback as copied text). After the fix, paste notes/failure-after.txt and keep both.

Retries belong only on transient faults. If the failure is bad input, a bad allowlist, or a logic bug in the RAG quality audit report, retrying will amplify cost without repairing trust around turn golden questions, expected evidence IDs, and citation checks into a release artifact.

Read

Worked judgment

Classify the failure as prevent, detect, contain, or recover—using this lab’s language, not a generic poster. For rag-quality-audit, the first fix should usually be detect+prevent at the boundary, because scoring fluency instead of citation support, or mixing docs versions in one report is cheaper to stop early than to explain in production prose.

Read

Why this stage matters for the RAG quality audit report

At the debugging stage for rag-quality-audit, the job is narrower than finishing a product demo. You are creating one progressive evidence piece about gold set with expected evidence IDs + candidate run output that later pages inherit without redefining success. Keep that fixture small enough to inspect by hand, keep outputs copy-pasteable as text, and refuse to narrate this baseline as if it were a production SLA: previous docs_version audit score.

For this page specifically, success looks like before/after evidence for the characteristic failure while still centering the user decision to turn golden questions, expected evidence IDs, and citation checks into a release artifact. If you cannot point to a file, command, or assertion that proves that for the RAG quality audit report, stay on this page instead of advancing.

Glossary: faithfulness · Glossary: recall@k · Cheatsheet: RAG quality · How-to: evaluate RAG quality

Previous · Next

Go deeper

Before you start

Why this matters

Describe the smallest fixture that triggers scoring fluency instead of citation support. Predict the first visible symptom (exception, wrong label, silent empty success). You will compare that prediction with the reproduction below.

In the wild

See how this idea shows up as a product and a company — then come back to the lesson. Skills transfer across vendors.

Check your understanding

Page assessment

Answer from memory. Completion is saved from this evidence, not from opening the next page.

1. Can you reproduce the failure with a one-command fixture?
2. Did you avoid retrying non-transient bad input?
3. Is before/after evidence saved as text (not only a screenshot)?
4. Does the repair restore the metric path toward: citation precision, unsupported answer rate, abstention correctness?

All responses are required.