RAG with Partial Contradictions Degrades Quality More Than No Retrieval

Providing retrieved context that partially contradicts the correct answer produces lower quality outputs than providing no retrieval context at all, due to model uncertainty and hedging behavior.

Evaluated 250 QA pairs with three conditions: no retrieval, accurate retrieval, and partially contradictory retrieval (correct answer present but with 1–2 contradicting passages also included). No-retrieval accuracy: 71%. Accurate retrieval: 89%. Contradictory retrieval: 58% — 13 points below baseline. Models hedge, average positions, or follow the wrong source in contradictory contexts.

Contradiction detection must be a retrieval pipeline stage, not optional. Before passing chunks to LLM, run a consistency check: if retrieved passages contradict each other on key claims, either resolve via re-ranking or flag to LLM explicitly: "Note: sources disagree on X. Apply skeptical synthesis." Dropping contradictory chunks is better than including them silently.