Unverified50% confidenceFactExact time
When chunks are split across semantic boundaries, with the definition in one chunk and the example in another, only one of them is recalled, and the LLM gets only half the answer.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe recursive character text splitter balances the contradiction between document chunks being too long to exceed the context window and too short to lose semantics by attempting to split at natural boundaries (paragraphs, sentences).69% similarUnverifiedSegmented generation should be split along business semantic boundaries rather than hard-cut by character count, otherwise the structural consistency of the concatenated result is difficult to guarantee.68% similarUnverifiedThe fixed-size chunking strategy splits by N characters (or tokens), retaining an overlap of M characters between adjacent chunks, without awareness of structure, sentences, or semantics.67% similarUnverifiedRecursive character chunking is LangChain's default approach, which sequentially attempts to split at hierarchical boundaries such as paragraphs, sentences, and words in order to prioritize semantic integrity.66% similarUnverifiedFixed word-count splitting may hard-cut in the middle of a sentence and disrupt semantics, so an overlap mechanism is commonly introduced (e.g., 500 characters per segment with 100 characters of overlap between adjacent segments) to preserve cross-segment information.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1048027API
curl https://kongchang.com/api/v1/knowledge/claims/1048027MCP
get_claim(id=1048027)