[KongchangAI]
Unverified50% confidenceTradeoffExact time

文档切片过大会引入噪声，过小则丢失语义完整性，固定大小切片通常为512~1024 Token的滑动窗口并保留上下文重叠区

1
Sources
50%
Confidence
Long-term
Relevance
7/11/2026
First Seen

Sources

AI Agent Learning Roadmap: From Zero to Real-World Projects
bilibiliAI大模型小海7/9/2026

Related Claims

Unverified固定长度切片按token数截断，通常配合10到20%滑动窗口重叠，chunk size一般设置在256到512 token72% similarUnverifiedVector embeddings typically map text into high-dimensional spaces of 512 to 4096 dimensions.64% similar

Cite This Claim

Stable URI
https://kongchang.com/claim/486572
API
curl https://kongchang.com/api/v1/knowledge/claims/486572
MCP
get_claim(id=486572)