A technical deep dive into how Data → Domain → Sentiment → Themes → Trends → Segments → Risk → Synthesis → QA actually works.
Why nine and not one
A single prompt asked to do everything degrades on every axis at once. Splitting the work lets each stage carry its own validation, its own retries and its own confidence score — and lets the QA agent reject a stage without discarding the run.
The stages
- Data — schema detection, language, quality flags
- Domain — knowledge pack binding and concept resolution
- Sentiment — aspect extraction and polarity
- Themes — governed taxonomy assignment
- Trends — period-over-period comparison
- Segments — cohort splits
- Risk — complaint-risk scoring
- Synthesis — narrative and recommendations
- QA — confidence scoring and rejection
Want to try this on your own data? Run a free VOC health check — no credit card, EU-processed.
Start freeKey takeaways
- Stage boundaries are where validation becomes possible.
- A rejected stage should never silently pass downstream.
About the author
Marc Dubois
Principal Data Scientist, InsightNarrator
Builds the multi-agent analysis pipeline and the taxonomy governance layer. Formerly NLP research in telco churn modelling.
Connect on LinkedIn