Traditional RAG systems struggle with context loss when breaking documents into chunks, causing important information to be missed. Hybrid search combining semantic and keyword methods helps but doesn't solve the fundamental problem of scattered context across large documents.

When text chunks lose their surrounding meaning, retrieval becomes unreliable. Simply increasing chunk size or overlap creates new problems like higher storage costs and altered semantic meaning. More sophisticated approaches like Hypothetical Document Embeddings and Document Summary Index have also failed to deliver substantial improvements.

Contextual retrieval, introduced by Anthropic in 2024, solves this by preserving each chunk's surrounding context. The method generates helper text for every chunk using an LLM, situating it within the original document. This approach maintains semantics, keywords, and context simultaneously, dramatically improving RAG pipeline accuracy. By ensuring chunks retain their full meaning during retrieval, contextual retrieval represents a significant advancement for handling complex documents where meaning spans multiple sections.