RAG Is Blind to Time — I Built a Temporal Layer to Fix It in Production

Standard RAG systems rank documents by cosine similarity with no awareness of when information was true, causing outdated documents to surface ahead of current ones. This post describes a temporal reranking layer inserted between the vector retriever and the LLM that addresses three distinct time problems: hard-removing expired facts, boosting active time-bounded events, and using exponential decay to prefer newer versions of updated documents. The design uses two orthogonal axes — validity state (EXPIRED/VALID/TEMPORAL) and document kind (STATIC/VERSIONED/EVENT) — combined in a scoring formula that blends semantic similarity with decay score, recency score, validity multiplier, and an event relevance gate. Content-type-specific half-life profiles (from 1 day for breaking news to 36,500 days for mathematics) prevent over-penalizing timeless content. Additional features include document pairing for weak-but-useful sources, confidence tiers for conflicting documents, query-adaptive temporal weighting, and version-chain deduplication. The implementation adds 15–30ms overhead and requires only a created_at timestamp on documents. Full code is available on GitHub.