RAG and Knowledge Grounding

Retrieval-Augmented Generation is the most powerful architectural intervention against hallucination: it shifts the model from trying to recall facts to reading and citing facts. This lesson explains how it works, where it still fails, and how to measure faithfulness.

Why RAG Reduces Hallucination

From Lesson 3: hallucination often occurs because the model does not have the correct fact in its parametric knowledge, so it generates a plausible-sounding token sequence instead. RAG addresses this at the root: rather than asking the model to recall a fact from training, it supplies the fact directly in the context window, then asks the model to read and synthesize what it was given.

This shifts the problem from "recall from uncertain training memory" to "read and faithfully report from a document you have in front of you." The latter is a task LLMs perform far more reliably than the former.

The Retrieval-Faithfulness Tension

Even with retrieved context in the window, models can still hallucinate. This is the retrieval-faithfulness problem: the model was supplied with the correct information but generated output that does not faithfully represent it. This happens for several reasons:

• Lost-in-the-middle effect: If the relevant chunk is buried in the middle of a long context, the model may under-attend to it. (See Lesson 3.)
• Parametric override: When the retrieved document conflicts with what the model "knows" from training, the model sometimes favors its parametric knowledge over the context. This is more common when the document says something surprising or counter-intuitive.
• Synthesis errors: When asked to synthesize across multiple retrieved chunks, models may blend or confuse information from different chunks, producing a hallucinated synthesis.
• Incomplete retrieval: If the retrieval step does not surface the right chunk, the model answers without the needed fact, and may hallucinate to fill the gap rather than acknowledging the limit.

Chunking Strategies for Faithfulness

The way you split documents into chunks has a direct effect on hallucination rate in RAG systems. Poorly chunked documents create retrieval gaps and synthesis problems.

Citation-Forcing Patterns

The strongest RAG anti-hallucination prompt pattern is citation forcing: require the model to cite the specific chunk that supports each claim in its answer. Uncited claims are flagged as potentially hallucinated.

Faithfulness Evaluation Metrics

Faithfulness evaluation measures how well the model's output is supported by the retrieved context. The field has converged on a set of metrics for this assessment:

When RAG Still Does Not Help

RAG is not a universal solution. Three categories of hallucination are largely unaffected by retrieval:

• Reasoning hallucinations: RAG provides facts, not reasoning. If the model makes a logical error with the correctly-retrieved facts, RAG cannot prevent it. Use chain-of-thought prompting (Lesson 5) for reasoning tasks.
• Synthesis hallucinations: When the answer requires combining information from many chunks, the synthesis step itself can introduce errors. Keep synthesis tasks narrow: fewer sources, more focused questions.
• Out-of-distribution queries: If the user's question is not well-served by any retrieved chunk, the model may answer from parametric memory rather than acknowledging the retrieval gap. Add explicit "no relevant information found" handling to your RAG pipeline.