SSD-0018 · Status: stable · Version 1.0 · Reviewed 12 November 2026 · Wikidata Q121362277
Definition
Retrieval-augmented generation (RAG) is a technique in which a language model answers a query by first retrieving relevant documents, then generating its answer conditioned on them.
檢索增強生成（RAG）是一種技術：語言模型在回答查詢前，先檢索相關文件，再以這些文件為條件生成答案。
In practice
For publishers this is the mechanism that decides whether content is even eligible to be cited. Retrieval happens before generation, so a page that cannot be retrieved cannot be quoted, however good it is. Retrieval is usually embedding-based: passages are matched by semantic similarity to the query rather than by keyword overlap, which favours clear, self-contained passages over pages whose meaning depends on surrounding context. It also explains a common complaint — a correct page is missed because the passage that answers the question never appears as a coherent unit.
對內容發佈者而言，這是決定內容是否有資格被引用的機制。檢索發生在生成之前，所以無法被檢索到的頁面不可能被引用，無論它多好。檢索通常基於嵌入：段落按與查詢的語意相似度匹配，而不是關鍵詞重疊，這有利於清晰、自足的段落，不利於意義依賴上下文的頁面。它也解釋了一個常見抱怨：一個正確的頁面被錯過，因為回答問題的那段話從未以一個完整單元出現。
What it is not
- Not a search engine. It is a generation method with a retrieval step attached.
- Not fine-tuning. Nothing is trained; documents are supplied at inference time.
- Not a cure for hallucination. Retrieval reduces invented claims; it does not eliminate them.
Status
Stable as a technical term, with a published research lineage and a clear mechanism. What remains unstable is how each commercial answer system implements it — chunking strategy, ranking, and whether citations are shown at all.
Identity
- Wikidata: Q121362277 — retrieval-augmented generation
- English aliases: RAG; retrieval augmented generation; grounded generation
- 中文正名：檢索增強生成（香港、台灣）；中國大陸作「检索增强生成」
- Related field: Q816826 — information retrieval
Related terms
Sources
- Wikidata, item Q121362277 “retrieval-augmented generation” (accessed 12 November 2026).
- Wikidata, item Q816826 “information retrieval” (accessed 12 November 2026).