Dense Passage Retrieval (DPR) is a neural open-domain QA retriever introduced by Karpukhin et al. (EMNLP 2020) built as a dual-encoder of two independently parameterized BERT models: a question encoder $E_Q$ and a passage encoder $E_P$. Each text is mapped to a single dense vector, and the question-passage similarity score is the dot product $\mathrm{sim}(q,p) = E_Q(q)^\top E_P(p)$. Training uses a contrastive objective with in-batch negatives (and optionally one hard BM25 negative per question), so the model learns to pull gold question-passage pairs together while pushing unrelated passages apart. At inference, all corpus passages are pre-encoded once and indexed for maximum inner product search (e.g., with FAISS), so retrieval reduces to a nearest-neighbor lookup on the query embedding. DPR is the canonical 'flat dense retrieval' baseline that strict-parity graph-RAG comparisons hold fixed alongside cutoff, candidate pool, and split policy.