A closer look at where Retrieval-Augmented Generation is heading. Each piece digs into a strand of recent research, cites the key papers, and draws out what it means for the systems you are building today.
General leaderboards like BEIR and MTEB told you which retriever was good in the abstract. A wave of 2024 to 2026 benchmarks asks a harder question: good at what, for whom? A tour of domain, task, and contamination-resistant RAG evaluation, and how to use it.
Read RAG evaluation gets specific: the 2026 wave of domain benchmarks
A new line of work trains the model itself to decide when and what to retrieve, interleaving reasoning with live search. A look at RL-trained search agents like Search-R1, and what they change for RAG.
As context windows stretch into the millions of tokens, does retrieval still earn its place? A look at the evidence on long-context models versus RAG, and why the answer is usually both.
Most RAG pipelines OCR a PDF, then chunk the text, and lose the tables, charts and layout on the way. A newer approach embeds the page image itself. A look at ColPali, ViDoRe and OCR-free retrieval.
Embedding search excels when the answer looks like the question. When relevance needs reasoning, not surface similarity, even top retrievers fall apart. A look at the BRIGHT benchmark and what it means for RAG.
Read Reasoning-intensive retrieval: where dense search breaks
A fast first-stage retriever casts a wide net; a slower reranker sorts what it catches. How cross-encoders and late-interaction models like ColBERT sharpen RAG retrieval, and when the extra stage pays off.
Read Reranking and two-stage retrieval: precision after recall
Vector search retrieves passages that look similar to the query. For multi-hop and whole-corpus questions, structure helps. A look at graph-based retrieval, from Microsoft GraphRAG to knowledge-graph grounding.
Read GraphRAG: when a knowledge graph beats plain vector search
A RAG system can retrieve the right passage and still write an answer the passage does not support. Measuring that gap, faithfulness, is its own hard problem. A look at RAGTruth, LLM-as-judge, and what to trust.
Read Measuring RAG faithfulness: is the answer actually supported?
Retrieval fetches from a fixed corpus; memory decides what an agent should carry forward from its own past. A look at write, consolidate and forget systems like MemGPT and Mem0, and their limits.
Naive RAG retrieves and hopes. A newer line of work lets the model decide when to retrieve, judge whether the passages are any good, and correct course. A look at Self-RAG, Corrective RAG and where this is heading.
Read Self-correcting RAG: retrieval that knows when it is wrong