The quality loop is a continuous cycle for improving RAG answers: collect feedback on real answers, use Compare Models to test alternatives, and use deep query tracing to see exactly why an answer came out the way it did — then fix the cause and repeat. Citation-backing makes an answer checkable; the quality loop is how you act on what the checking reveals.
Why grounding alone isn’t enough
Retrieval-augmented generation reduces fabrication by constraining answers to your sources. But “grounded” is not “correct.” Retrieval can pull the wrong passage; a poorly structured document can retrieve badly; a model can over-reach. If you stop at grounding, you have reduced the risk without building any way to see or improve it. The quality loop closes that gap.
The three parts, working as a cycle
Feedback collection lets users signal when an answer helped or missed, on the real questions they ask — your ground truth, drawn from production. Compare Models tests alternatives on the same queries, so you decide with evidence on your corpus. Deep query tracing is the diagnostic core: it shows which passages were retrieved, what was passed to the model, and how the answer was formed. When an answer is wrong, the trace tells you why.
How the loop runs
Observe → diagnose → act → confirm
- Observe — feedback flags answers that missed.
- Diagnose — tracing reveals the cause (retrieval, content, or model).
- Act — fix the content, adjust retrieval, or switch the model via Compare Models.
- Confirm — watch feedback on the same kinds of question to verify the fix.
Most quality problems trace back to inputs — sprawling, duplicated or poorly structured documents retrieve noisily. Tracing surfaces that, so the fix is often improving the source content rather than touching the model at all.
Why this is also a compliance asset
See how citation-backed RAG works and the quality-loop page for the product detail.
Frequently asked questions
Isn't citation-backing enough on its own?
Citations make an answer verifiable, which is necessary but not sufficient. The quality loop is what lets you find systematic problems and fix their causes, rather than catching errors one at a time. The two work together: grounding for trust per answer, the loop for improvement over time.
Do I need data scientists to run this?
No. Feedback is a user action, Compare Models is a built-in comparison, and tracing is a readable view of what happened — designed for the team operating the platform, not a research group. The skill it asks for is curiosity about why an answer missed, not machine-learning expertise.
What usually needs fixing — the model or the content?
More often the content and how it is structured than the model. That is why tracing matters: it points you to the real cause instead of letting you guess.
Sources & further reading
- RAGSuite — the quality loop (Compare Models, tracing, feedback) — the feature set in the product
- How citation-backed RAG works — the grounding the loop improves