Retrieval-Augmented Generation

Retrieval-augmented generation (RAG) combines a language model with external documents retrieved at query time. The model generates from both the query and retrieved evidence.

A simple RAG pipeline is:

$$D_q=\mathrm{Retrieve}(q,\mathcal{D})$$ $$y\sim p_{\theta }(y\mid q,D_q)$$

where $q$ is a query, $\mathcal{D}$ is the document collection, and $D_q$ is the retrieved subset.

RAG has two separable quality gates:

$$\text{retrieval quality}\ne \text{answer faithfulness}$$

Good retrieval can still produce an unsupported answer if generation ignores or distorts evidence.

In practical systems, retrieval often includes query rewriting, chunking, hybrid retrieval, reranking, and context packing.

Checks

• Are retrieved documents relevant, current, and source-attributed?
• What is the retrieval unit: document, section, paragraph, chunk, table, or code block?
• Is query rewriting improving recall without changing intent?
• Is retrieval evaluated separately from generation?
• Are untrusted retrieved documents protected against prompt injection?
• Does the answer cite or link the evidence it used?
• Is the corpus public, private, or mixed?
• Are unsupported generated claims marked or removed?

Related

• Evidence-grounded generation
• Embedding retrieval
• Chunking
• Hybrid retrieval
• Query rewriting
• Context packing
• Prompt injection boundary
• LLM Wiki
• Prompt injection
• Verification loop
• Systems papers