Your RAG system is not hallucinating. It is failing to retrieve.
Enterprise AI Management · October 2026 · PolyQ Team

The diagnosis shifted in 2026. Practitioners increasingly report that the dominant failure in enterprise retrieval is not the model inventing things — it is the retriever handing the model the wrong passages and the model faithfully summarising them.
“The model hallucinated” is a comfortable explanation because it puts the fault somewhere you cannot reach. It is also, in a retrieval system, usually wrong. If the passages handed to the model do not contain the answer, a confident wrong response is the system working exactly as built.
- Retrieval failure is the common case: The 2026 practitioner consensus puts retrieval quality, not generation, at the centre of enterprise RAG failures. That reframes the fix: better chunking, hybrid dense and sparse search, query reformulation and multi-hop retrieval do more than a larger model.
- Citation drift is the subtle one: A system can cite the correct document and still extract a passage that does not support the claim it is attached to. Many pipelines have no mechanism tying each individual claim to an identifiable span of source text, which is precisely what an auditor will ask for.
- A regulated domain shows the ceiling: Published evaluations of production legal research assistants reported hallucination on a meaningful share of queries — around 17% for one major product and roughly a third for another. These were built by serious vendors over curated corpora, which is the point.
- Agentic retrieval became the default pattern: Specialised agents handling retrieval and validation in parallel, with self-correction loops, is now the dominant architecture rather than a research curiosity. It adds cost and latency in exchange for catching the retrieval miss before the user sees it.
- Governance has to precede retrieval: Access control and metadata are not a layer you add afterwards. Without them the retriever is as likely to surface something the asker should not see as something they need.
Blaming the model is the explanation that requires nothing of you. Blaming the retriever is the one that comes with a list of things to fix, which is why it is the more useful diagnosis even when both are partly true.
If you are evaluating a grounded assistant, ask the vendor to show you a failed answer and point at which retrieved passage caused it. The ones who cannot are not instrumented.