L0S2 · Data & RetrievalS5 · Eval & LLMOpsVerified: 2026-11-17
RAG Chunk Boundary Collapse & Lost-in-Context Retrieval Audit Primitive
Diagnostic primitive auditing how naive fixed-token chunking severs table headers, legal exceptions, and co-references across split boundaries, decoupling retriever miss rate from generator hallucination.
Best For
Long contracts, API reference manuals, and financial filings where child chunks lose crucial section headers, table units, or entity antecedents.
Human-AI Boundary
Annotators tag failed retrieval traces (severed header, broken pronoun, split table) to drive deterministic Stage 2 ingestion rules.
Anti-Patterns to Avoid
- 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
Latency/Cost Profile & Eval Metrics
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))
💰 Business ROI Model & Evidence Level
production_practice📊 Benchmark & Alternative Comparison
- 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.
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
Bidirectional Taxonomy Links
[L1] Stage 2 · Incremental CDC Document Sync, Content-Hash Dedup & Tombstone GC Pattern[L1] Stage 2 · Semantic-Layer Grounded Text-to-SQL & Read-Only Sandbox Repair Pattern[L2] L2 Scenario · Enterprise Legal Contract Redline & Risk Clause Auditor[HubA] Company Intelligence Hub · Amazon AWS AI, Applied Scientist & LP Loop[L0 Card] Parent-Child & Sentence-Window Chunking[L0 Card] Parent-Child Chunking[L0 Card] Sentence-Window Chunking[L0 Card] Semantic Chunking[L0 Card] RAG Triad Evaluation (Ragas)
Internal Tools & Primary Sources
- Evaluating Chunking Strategies for Retrieval-Augmented Generation (paper, 2026-11-17)
- Anthropic — Introducing Contextual Retrieval (engineering_blog, 2026-11-17)