Traditional RAG systems fail because fixed-size chunking strips chunks of their surrounding context, leading to pronoun ambiguity, orphaned comparisons, and missed semantic relationships. Contextual retrieval addresses this by prepending LLM-generated context summaries to chunks before embedding, so the vector representation captures both content and its place in the document. Combining contextual embeddings with BM25 keyword search (hybrid retrieval) outperforms either method alone. Additional improvements include semantic or structural chunking instead of fixed token splits, two-stage reranking to distinguish relevance from mere similarity, and graph-based retrieval to model explicit entity relationships. Common pitfalls include over-optimizing embeddings while neglecting chunking quality, ignoring metadata filters, using single-shot retrieval, and lacking fallback strategies when no relevant chunks are found.
