HalfLife adds temporal reranking to RAG

This is a practical fix for a real failure mode: many RAG systems already retrieve the right chunk, then lose it in ranking because “more complete” older text scores better. HalfLife’s value is that it treats freshness as a first-class ranking signal instead of pretending retrieval alone will solve stale answers.

• –The core idea is modest but useful: infer dates from text, classify query intent, then fuse semantic and temporal scores during reranking
• –That makes it a better fit for messy corpora where metadata is incomplete or unreliable
• –The approach is heuristic, so it will work best when year extraction and intent classification are good enough to steer obvious “today/current/latest” queries
• –The repo claims it can be used as a drop-in postprocessor, which makes experimentation cheaper than rebuilding retrieval pipelines
• –This sits in the broader RAG tooling gap between semantic relevance and time-aware relevance, which is still under-served in most stacks