L0S2 · 数据与检索S5 · 评测与观测Verified: 2026-11-17
RAG 切块边界割裂(Boundary Collapse)与长尾召回诊断原语
量化固定字数切片把“表头与数据行、条款前提与例外条款”从中间腰斩导致的检索假阴性,通过上下文保留率、父子块扩展命中率及长尾切片 Recall@K 解耦检索与生成故障。
适用场景 (Best For)
长合同条款、技术 API 手册、财报附注等“局部字句高度依赖上文标题与定义”的高精度企业 RAG 检索诊断。
人机边界 (Human-AI Boundary)
对于高频错题切片,人工抽查并标记“缺表头/指代丢失/切碎跨页表格”原因,反哺 Stage 2 结构化切块正则或父子块策略。
常见反模式与避坑 (Anti-Patterns)
- Debugging LLM prompt instructions for weeks when 70% of incorrect RAG answers never retrieved the supporting sentence in Top-20 chunks
- Splitting Markdown or HTML tables mid-row using blind fixed-character `RecursiveCharacterTextSplitter` without injecting section path prefixes
延迟成本与评测门禁
Offline ingestion audit + lightweight runtime parent-chunk fetch (<5ms KV lookup) · Parent-child chunk index multiplier + optional contextual-header enrichment tokens at ingest time
- Span Integrity & Section-Header Retention: >= 95% gold evidence spans remain intact inside single retrieved parent contexts (Evaluate evidence span overlap against chunk boundaries on multi-paragraph gold passages)
- Retrieval Attribution Separation (Miss vs Hallucination): Pre-reranker Recall@20 >= 0.94 before tuning generation prompts (Log misses by document category (dense text, tables, footnotes))
💰 商业 ROI 模型与证据等级
production_practice📊 同类技术选型与基准对比
- Hierarchical Parent-Child + Section-Path Header: matches tight child vectors for high precision while returning rich parent windows to LLM context.
- Blind Fixed 512-Token + 50 Overlap: fastest baseline setup, repeatedly severs cross-sentence qualifications and multi-row table semantics.
⚡
想检验本题的肌肉记忆?进入 3 阶闪卡主动回忆盲测
隐藏答案遮罩,自测:审题锚点 ➔ 数理公式机理 ➔ 工业标杆代码与连环追问。3 分钟攻破认知盲区!
上下游关联知识链路
[L1] Stage 2 · 增量 CDC 文档同步、内容哈希幂等去重与向量墓碑清理模式[L1] Stage 2 · 语义指标层约束 Text-to-SQL 与只读沙箱校验模式[L2] L2 业务场景 · 企业法务合同条文比对、风险批注与 Playbook 溯源审核助手[HubA] 企业情报枢纽 · Amazon AWS/Bedrock AI、电商搜索与 Leadership Principles 解析[L0 Card] Parent-Child 与句窗分块策略[L0 Card] Parent-Child 父子切块策略[L0 Card] Sentence-Window 句窗口切分[L0 Card] Semantic 语义断崖自适应切分[L0 Card] RAG 三元组评估 (Ragas/TruLens)
站内交互工具与信源依据
- Evaluating Chunking Strategies for Retrieval-Augmented Generation (paper, 2026-11-17)
- Anthropic — Introducing Contextual Retrieval (engineering_blog, 2026-11-17)