Unverified50% confidenceFactExact time
Re-chunking and re-embedding an entire document wastes compute and loses the continuous identity of the content, making it impossible to answer questions about the historical versions of a clause.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe author believes that only by being free from frameworks can ID invariants be strictly enforced end-to-end, as mainstream frameworks such as LangChain's Document objects undergo multiple serializations during chunking, retrieval, and passing, leading to metadata loss.63% similarUnverifiedRAG将文档切成固定大小chunk，跨文档、跨时间的逻辑关联在切片时被破坏，向量相似度检索无法重建这些隐性联系62% similarUnverifiedFailure of vector indexes to reflect data changes in a timely manner leads to two types of problems: stale retrieval (generating answers based on outdated documents) and ghost data (source documents have been deleted but vectors remain in the index and are still retrieved).61% similarUnverifiedWhen exploration goes down a wrong branch, the Rewind command should be used to truncate the trailing invalid turns in order to preserve the prefix cache, rather than using the Compact command, which rewrites history and destroys the cache.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/908861API
curl https://kongchang.com/api/v1/knowledge/claims/908861MCP
get_claim(id=908861)