Builder · RAG Quality System
Delta Support Audit
Nightly RAG audit of support articles for factual drift.
Builder & triager · 2026 · Next.js 15 / TypeScript · OpenRouter → Claude Sonnet · Upstash Redis + Vector · BGE-large-en-v1.5 embeddings · RAG · Slack + Notion
The problem
Support content drifts. Freshdesk articles quietly fall out of step with the product docs and guides that are the real source of truth, and nobody notices until a customer is misled.
The system embeds 344 support articles (1,067 chunks) into a vector store and compares them against the canonical docs for contradictions and coverage gaps.
How it works
RAG detectors classify each article for conflicts, contradictions and coverage gaps. P0 issues page a Slack channel; the full report lands in Notion. A Vercel cron runs the whole sweep at 04:00 IST.
On a 100-article trial it caught 14/14 P0 issues (100% true-positive) with a combined P0+P1 false-positive rate of 5–9%, at $2.45 and four minutes per run.
Colophon
Next.js 15 / TypeScript · OpenRouter → Claude Sonnet · Upstash Redis + Vector · BGE-large-en-v1.5 embeddings · RAG · Slack + Notion · Vercel Cron
