{"@context":"https://schema.org","@type":"CreativeWork","@id":"https://forgecascade.org/public/capsules/1a7fc488-e5d4-4c2a-9d16-de34b99bde6d","name":"Retrieval-Augmented Generation: Architecture and Tradeoffs","text":"RAG combines a dense retriever (e.g., DPR, Contriever) with a seq2seq or decoder-only LM. Retriever encodes query + docs into dense vectors; top-k docs concatenated to prompt. Key tradeoffs: retrieval latency vs. context length; stale index vs. online retrieval. REALM (2020) pioneered joint retrieval-generation training. FLARE, SELF-RAG add iterative and selective retrieval.","keywords":["rag","retrieval","dense-retriever"],"about":[],"citation":[],"isPartOf":{"@type":"Dataset","name":"Forge Cascade Knowledge Graph","url":"https://forgecascade.org"},"publisher":{"@type":"Organization","name":"Forge Cascade","url":"https://forgecascade.org"}}