Unverified50% confidenceFactExact time
The model was mostly able to ignore irrelevant injections: 10 ordinary questions about push notifications received saved answers for different features, and none of the replies mentioned it.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSemantic caching bypasses the model call entirely when a new query is sufficiently similar to an already-answered query, but carries the trade-off of returning incorrect answers to subtly different questions.76% similarUnverifiedThe decision model takes a state and up to five typed questions (noul/choice/score) and returns answers and confidence in a single forward pass, generating no text, so output_tokens is 0.74% similarUnverifiedWhen the retrieval context does not contain the answer, many language models fill the gap with plausible-sounding content, which is exactly the hallucination users notice; fixes include instructing the model to answer only from the provided context, giving it a clear way to opt out, requiring a citation for every claim, setting a retrieval score threshold, and logging rejected questions.74% similarUnverifiedOne-third of the questions had no answer in the documents, and all models answered 'not found' for these unanswerable questions across all document scales, with none fabricating an answer.73% similarUnverified不提供上下文时，本地小模型对领域问题会拒绝回答或产生幻觉，引入 RAG 后能给出准确答案71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1027005API
curl https://kongchang.com/api/v1/knowledge/claims/1027005MCP
get_claim(id=1027005)