Practical Checklist to Fix Common Retrieval-Augmented Generation Failures

Retrieval-augmented generation (RAG) systems frequently fail not at the generation stage but during retrieval, where wrong context leads to confidently incorrect outputs. Engineers building research and analysis pipelines recommend chunking documents on semantic boundaries rather than fixed token windows, and combining keyword search with vector search using reciprocal rank fusion to improve recall. A cross-encoder reranker should then narrow top candidates to the most relevant chunks before passing them to the language model. Structured metadata filters, such as date or document type, should be applied before semantic search to avoid relying on embeddings alone for precision. Developers are also advised to build a test set of real questions with known source passages and continuously measure retrieval accuracy and answer faithfulness across any system changes.