This piece walks through a failure mode most enterprise RAG pipelines have but no one is measuring: retrieval works, the right documents come back, and the model still answers wrong because two of those documents contradict each other. A preliminary earnings figure and the audited restatement. The June policy and its November reversal. Your v1.2 API docs and the v2.0 update. The model picks one — usually the older one, because position and lexical alignment win — and reports it with 80% confidence.
What it gets right: the diagnosis. Confidence scores and retrieval metrics will not flag this. Hallucination detectors will not flag this. The conflict is detected nowhere because no pipeline stage is looking for it.
The actionable kernel for an Architect or data lead: audit your top-k retrievals on queries you know touch versioned content — financial restatements, policy revisions, deprecated APIs. If conflicting documents are landing in the same context window, you have this problem in production right now, and you cannot see it. The fix is one pipeline stage between retrieval and generation: pair-wise contradiction detection on what was retrieved, then a resolution rule (recency for versioned docs, surface-both for genuine disputes). The author's reference implementation runs on a CPU in 220 MB — proof that the architectural gap matters more than model size.
Read it before your next RAG design review: