temporal filtering

**Temporal filtering** is the **retrieval filtering technique that limits candidates by publication or validity time windows** - it helps systems prioritize evidence that is current for time-sensitive questions. **What Is Temporal filtering?** - **Definition**: Time-based constraints applied to documents or chunks during retrieval. - **Time Signals**: Uses created dates, updated timestamps, effective dates, and expiry metadata. - **Window Types**: Supports relative windows such as last 30 days and absolute ranges by calendar date. - **Pipeline Role**: Combines with semantic ranking to balance recency and topical relevance. **Why Temporal filtering Matters** - **Freshness Control**: Reduces outdated evidence in domains with fast-changing facts. - **Regulatory Accuracy**: Ensures responses reflect valid policy versions at answer time. - **User Intent Match**: Many queries imply current-state answers even without explicit date terms. - **Noise Reduction**: Old historical records can dominate retrieval unless constrained. - **Trust Preservation**: Time-aligned evidence lowers visible answer contradictions. **How It Is Used in Practice** - **Date Normalization**: Standardize all timestamps into one canonical timezone and format. - **Recency Boosting**: Blend hard filters with rank boosts for newer but still relevant documents. - **Evaluation by Epoch**: Benchmark retrieval quality separately for stable and volatile knowledge areas. Temporal filtering is **essential for recency-sensitive RAG workflows** - time-aware retrieval improves factual currency and reduces stale-answer risk.