Unverified50% confidenceFactExact time
When a retriever recalls documents based solely on semantic similarity, it may return both outdated and current documents simultaneously, causing the LLM to answer based on stale information.
1
Sources
50%
Confidence
Long-term
Relevance
9/26/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedStale embeddings silently winning retrieval means that if old-version vectors are not cleaned up or down-weighted in time, they may be preferentially recalled due to greater semantic similarity, returning outdated answers without raising errors.71% similarVerifiedThe semantic caching mechanism vectorizes and stores historical responses to similar requests, and when semantic similarity exceeds a threshold, it directly returns cached results to reduce upstream API call frequency and costs69% similarUnverified语义缓存命中判断通常设置预设相似度阈值（例如0.9），超过则直接返回缓存答案跳过LLM调用69% similarUnverifiedLLMs exhibit Positional Bias, tending to focus more on the beginning and end of the input sequence, with information in the middle easily being forgotten.68% similarUnverifiedWhen context becomes too long, history should be organized in segments while retaining recent results, and if necessary a new web conversation with progress should be continued; when the response format is wrong, return it to the model for correction, and if it is still wrong, stop—never execute by guessing.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/951749API
curl https://kongchang.com/api/v1/knowledge/claims/951749MCP
get_claim(id=951749)