Unverified50% confidenceFactExact time
During HNSW queries, the upper-layer search uses ef = 1, keeping only one candidate and performing pure greedy descent; the bottom layer uses a beam search of size ef.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe bottleneck of single-pass retrieval stems from vector similarity search compressing the user query into a single high-dimensional vector; when a question contains multiple semantic directions, this vector becomes an average of multiple semantics, causing sub-intents to be insufficiently represented.67% similarUnverified语义搜索和关键词搜索单独使用时约每4个结果漏检1个，两者结合漏检率降到约十分之一66% similarUnverifiedThe hnswlib documentation states that a higher ef makes search more accurate and slower, and that it cannot be less than k; this description can basically be carried over to understanding pgvector's hnsw.ef_search.64% similarUnverifiedThe main source of the 93% reduction in input tokens is the context compression stage of Discovery AI, which extracts only the fields and a small number of examples most relevant to the current query type.62% similarUnverified索引阶段和查询阶段必须使用同一个Embedding模型，否则向量空间不一致会导致检索完全失效62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1050490API
curl https://kongchang.com/api/v1/knowledge/claims/1050490MCP
get_claim(id=1050490)