Unverified50% confidenceTradeoffExact time
RAG中Chunk大小通常在256至1024 token之间，过小导致语义不完整，过大则引入噪声并稀释相关性
1
Sources
50%
Confidence
Long-term
Relevance
7/22/2026
First Seen
Sources
Related Claims
Unverified小模型长上下文理解质量随长度下降快，通常超过4K-8K token窗口后表现退化64% similarUnverified文档分块需权衡片段大小与语义完整性，过小则语义不完整，过大则引入噪音64% similarUnverified块太大会导致嵌入向量稀释、检索精度下降，块太小则可能丢失完整的语义单元63% similarUnverified量化对结构化输出（如 JSON 工具调用）的影响比对通用对话更敏感，因边界 token 概率分布较窄易受权重噪声漂移触发格式错误63% similarUnverifiedRAG文档切分粒度太大会引入噪声，太小则丢失上下文，是工程实践中需要反复调优的关键参数62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/593879API
curl https://kongchang.com/api/v1/knowledge/claims/593879MCP
get_claim(id=593879)