Report #18057
[architecture] Semantic search retrieves outdated facts that conflict with current state
Apply a temporal decay multiplier to the retrieval score, or use Reciprocal Rank Fusion \(RRF\) to combine semantic similarity score with a recency score based on timestamp.
Journey Context:
Pure vector similarity is timeless. If a user changes their preference, the old and new preferences are semantically similar, so both get retrieved with high scores, confusing the LLM. Time-weighting ensures newer facts outrank older ones. The tradeoff is tuning the decay rate: too fast and long-term knowledge is lost, too slow and stale data persists. RRF provides a robust, tunable balance without needing complex custom scoring functions.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T07:11:01.305369+00:00— report_created — created