Back to AI Systems Engineer Mind Map
中文·English
🚀 AI Systems EngineerID: aie-chunking-hierarchical-parent-child

Parent-Child & Sentence-Window Chunking

Parent-Child 与句窗分块策略
🎯Core Definition
Hierarchical Parent-Child & Sentence-Window Chunking resolves the fundamental trade-off in RAG systems between retrieval precision and generative context completeness; standard fixed-size chunking forces a compromise (small chunks lack context for generation, while large chunks dilute embedding representations); hierarchical chunking decouples indexing from synthesis: 1) Parent-Child Chunking: slicing a large parent document (2,000 characters) into dense child chunks (200 characters); the vector database indexes only child chunks for high semantic matching precision, but upon retrieval automatically swaps children for full parent chunks via `parent_id` foreign keys before feeding the LLM; 2) Sentence-Window Retrieval: embedding individual sentences, then expanding a surrounding ±K\pm K sentence window dynamically around retrieved matches.
💡Use Cases
Legal contract search, technical API documentation RAG, and financial equity research synthesis.
Key Problems Solved
Standard chunks clip critical premise context, causing hallucinated completions; Parent-Child chunking achieves fine-grained search indexing with broad, unclipped context generation.
🎯5 High-Frequency Exam Points
1
Diagram the end-to-end Parent-Child dataflow from hierarchical chunking, child embedding, vector lookup to parent payload hydration?
2
Analyze the trade-off of sentence-window expansion radius KK on context coherence versus introducing irrelevant token noise?
3
Explain Semantic Chunking algorithms that detect semantic boundaries via cosine distance drops between consecutive sentence embeddings?
4
How do advanced document parsers preserve intact Markdown table structures and code blocks across chunk boundaries?
5
How does injecting document hierarchy breadcrumbs into child chunk metadata boost vector retrieval accuracy?
Updated 2026-08-14
🎯
Test Your Knowledge: Practice Questions for "Parent-Child & Sentence-Window Chunking"
Single choice pitfall questions with instant feedback and mistake tracking.
🚀 Start Card Practice
Previous CardLangSmith-Style Tracing & Self-HealingNext CardHybrid Search (BM25 + Dense) & RRF

🔗 More AI Systems Engineer Knowledge Cards

AIE vs MLE Competency ModelPrompt Chains & Jailbreak DefenseStructured Output & Constrained DecodingEval Benchmarks: RAGAS & SWE-bench