RAGとは？
Retrieval-Augmented Generation（RAG）は、外部の知識源を統合することにより、大規模言語モデル（LLM）の精度と信頼性を高めるAIフレームワークです。
RAGは、情報検索とテキスト生成の強みを組み合わせたものです。まず、ユーザーのクエリに基づいて、知識ベースまたは外部データソースから関連情報を取得します。次に、この取得した情報がLLMのプロンプトを補強し、より情報に基づいた正確な応答を生成するためのコンテキストと事実を提供します。このアプローチは、LLMの内部パラメータへの依存を減らし、特に専門的または最新の知識に関して、誤った情報やハルシネーションされた情報を生成するリスクを軽減します。このプロセスには通常、知識ベースのインデックス作成、ユーザーの質問によるクエリ、関連ドキュメントの取得、そしてそれらのドキュメントを元の質問とともにLLMにフィードして回答を生成することが含まれます。
graph LR
Center["RAGとは？"]:::main
Pre_logic["logic"]:::pre --> Center
click Pre_logic "/terms/logic"
Rel_retrieval_augmented_generation["retrieval-augmented-generation"]:::related -.-> Center
click Rel_retrieval_augmented_generation "/terms/retrieval-augmented-generation"
Rel_rag_pipeline["rag-pipeline"]:::related -.-> Center
click Rel_rag_pipeline "/terms/rag-pipeline"
Rel_reinforcement_learning["reinforcement-learning"]:::related -.-> Center
click Rel_reinforcement_learning "/terms/reinforcement-learning"
classDef main fill:#7c3aed,stroke:#8b5cf6,stroke-width:2px,color:white,font-weight:bold,rx:5,ry:5;
classDef pre fill:#0f172a,stroke:#3b82f6,color:#94a3b8,rx:5,ry:5;
classDef child fill:#0f172a,stroke:#10b981,color:#94a3b8,rx:5,ry:5;
classDef related fill:#0f172a,stroke:#8b5cf6,stroke-dasharray: 5 5,color:#94a3b8,rx:5,ry:5;
linkStyle default stroke:#4b5563,stroke-width:2px;
🧠 理解度チェック
🧒 5歳でもわかるように説明
Imagine you're asking a super-smart robot a question, but it only knows things it learned a long time ago. RAG is like giving the robot a quick peek at a library book (the knowledge base) before it answers, so it can tell you the latest and most correct information!
🤓 Expert Deep Dive
RAG architectures fundamentally decouple knowledge acquisition from model inference by augmenting a generative Large Language Model (LLM) with an external, dynamically queried knowledge source. The core components typically include:
- Document Indexing: A corpus of documents (e.g., articles, PDFs, web pages) is processed and embedded into a vector space using a pre-trained embedding model (e.g., Sentence-BERT, OpenAI's
text-embedding-ada-002). These embeddings are stored in a vector [database](/ja/terms/vector-database) (e.g., Pinecone, Weaviate, FAISS) for efficient similarity search. - Retrieval: Upon receiving a user query, the query is also embedded into the same vector space. A similarity search (e.g., Approximate Nearest Neighbor - ANN) is performed against the vector database to identify the
kmost relevant document chunks (or passages) based on cosine similarity or dot product. - Augmentation & Generation: The original user query and the retrieved document chunks are concatenated into a single prompt. This augmented prompt is then fed into the LLM. The LLM uses this context to generate a response that is grounded in the retrieved information.
Mathematically, the retrieval step can be viewed as finding document embeddings $d_i$ such that their similarity to the query embedding $q$ is maximized:
$i^* = \arg\max_i \text{sim}(q, d_i)$
where $\text{sim}(u, v)$ is a similarity function like cosine similarity: $\frac{u \cdot v}{\|u\| \cdot \|v\|}$.
This approach mitigates the 'knowledge cut-off' problem inherent in LLMs and significantly reduces hallucination by providing factual grounding. Advanced RAG techniques explore re-ranking retrieved documents, query expansion, and fine-tuning the retriever and generator components jointly (e.g., REALM, DPR).