Your LLM Is Hallucinating Facts.

RAG hallucination occurs when a Retrieval-Augmented Generation system produces fabricated or inaccurate information despite having access to a knowledge base. It can happen at three stages: retrieval (wrong documents retrieved), augmentation (conflicting or excessive context), or generation (the model ignores retrieved documents and generates based on training data patterns instead).

RAG can reduce certain types of hallucination by grounding model outputs in retrieved documents, but it does not eliminate hallucination. It shifts the failure point from the model’s training data to the retrieval pipeline — and introduces new failure modes at retrieval, augmentation, and generation stages. Without proper safeguards like confidence thresholds and post-generation verification, RAG systems can still hallucinate confidently.

Hallucination rates vary significantly depending on the domain, retrieval quality, chunk size, embedding model, and evaluation methodology. Published benchmarks report rates ranging from 3% to over 20% depending on task complexity. The critical point is that aggregate accuracy metrics hide which component is failing — you need to measure retrieval precision and generation faithfulness separately to understand your true hallucination risk.

LLM hallucination is a fundamental consequence of how language models work. They optimise for what sounds plausible, not for what is true. They sample outputs based on statistical patterns from training data. When the model lacks sufficient context or encounters ambiguity, it generates confident, grammatically perfect responses that may be completely fabricated. In RAG systems, this is compounded by retrieval failures and context window limitations.

Prevention requires a multi-layered approach: hard confidence thresholds at retrieval to filter irrelevant documents, corrective RAG with fallback mechanisms, post-generation claim extraction and evidence linking, hybrid retrieval with reranking, and a three-tier evaluation framework measuring retrieval precision, answer faithfulness, and response relevance separately. No single technique is sufficient — effective prevention addresses all three failure points.

Corrective RAG is an approach that adds verification and fallback mechanisms to the standard RAG pipeline. When retrieved documents fall below a confidence threshold, the system triggers fallback behaviour — such as querying alternative knowledge sources, reformulating the search query, or returning “I don’t have enough information” instead of generating a potentially hallucinated response. It prioritises accuracy over always providing an answer.

In the context of RAG systems, the three types correspond to the three failure points: (1) Retrieval hallucination — the system retrieves wrong or irrelevant documents, causing the model to generate from insufficient context; (2) Contextual hallucination — conflicting or excessive documents in the prompt cause the model to blend, invent, or pick arbitrarily; (3) Generation hallucination — the model ignores retrieved context and generates based on training data statistical patterns instead of source documents.

Post-generation claim extraction is a verification technique that breaks every model response into individual factual statements, then attempts to link each statement back to specific source evidence in the retrieved documents. Statements that cannot be grounded in source evidence are flagged as potential hallucinations. This catches fabricated claims before they reach users and provides an auditable trail of which sources support which outputs.

In AI, hallucination and confabulation are often used interchangeably to describe when a model generates false information presented as fact. Some researchers distinguish them: hallucination refers to generating content with no basis in input or training data, while confabulation refers to filling gaps in knowledge with plausible but fabricated details — similar to the neuropsychological term. In practice, both result in the same outcome: confident, incorrect outputs that users cannot easily distinguish from accurate ones.

Several evaluation frameworks specialise in measuring RAG system quality: Ragas (Retrieval Augmented Generation Assessment), TruLens, and DeepEval are among the most widely used. These tools measure three key dimensions: retrieval precision (are the right documents being retrieved?), answer faithfulness (does the response accurately reflect the retrieved documents?), and response relevance (does the answer address the actual question?). Measuring these separately reveals which component is failing.

Yes. Organisations are legally responsible for decisions and outputs generated by their AI systems, whether those outputs are accurate or fabricated. In healthcare, hallucinated treatment recommendations can cause patient harm. In legal services, fabricated case citations — as demonstrated in the June 2023 law firm incident — can result in sanctions and malpractice claims. In finance, incorrect AI-generated advice can lead to regulatory penalties. The liability exists regardless of whether the hallucination originated from the model, the retrieval system, or the augmentation process.

Hybrid retrieval combines multiple search strategies — typically dense vector search (semantic similarity) and sparse keyword search (BM25/TF-IDF) — to retrieve documents. A reranking model then reorders the combined results by relevance. This approach addresses the weakness of single-strategy search: dense retrieval may miss exact keyword matches while sparse retrieval may miss semantically related content. Hybrid retrieval with reranking significantly improves retrieval precision and reduces the likelihood of hallucination caused by missing critical documents.