Fix Your Search, Fix Your RAG Output

Short summary

Most RAG quality problems are actually retrieval problems, not LLM problems. The article argues that mechanical chunking (fixed token windows) destroys semantic meaning and advocates for structural-then-semantic chunking using embedding-based valley detection. It also covers parent-child chunk relationships for scenarios where full-document retrieval is preferable to chunk-level retrieval.

• •RAG hallucinations usually stem from poor retrieval, not poor LLMs
• •Semantic chunking outperforms fixed-token chunking by preserving meaning boundaries
• •Parent-child chunk strategies let you retrieve precise chunks or full documents as needed

Generated with AI, which can make mistakes.