Your RAG system isn't hallucinating. Your data is.
Nine times out of ten, a 'hallucinating' enterprise AI assistant is faithfully summarizing documents that are outdated, duplicated, or contradictory. Fix retrieval's inputs before blaming the model.
A pattern we see constantly: a company pilots an LLM assistant over its internal knowledge, the answers come back wrong, and the verdict lands immediately: "the model hallucinates, AI isn't ready."
Then you open the source documents, and the truth is more uncomfortable. The model summarized exactly what it was given. What it was given was wrong.
The three failure modes that masquerade as hallucination
Stale truth. The pricing page from 2023 and the pricing page from 2026 both live in the knowledge base. Retrieval finds the older one; it has more inbound links and better keyword overlap. The model answers confidently, and incorrectly, with last year's prices.
Duplicate drift. The same policy exists in four places: the wiki, a PDF, a slide deck, and an email thread. Three of the four were never updated. Majority vote goes to the wrong answer.
Contradiction without arbitration. Two departments documented the same process differently, and both are "official." The model has no way to know which one wins, so it blends them, producing an answer that matches neither.
None of these are model failures. They're data governance failures, surfaced at conversational speed.
What actually fixes it
- Curate before you index. Treat the corpus as a product. Assign owners, kill duplicates, and mark authoritative sources explicitly. A smaller, cleaner corpus beats a bigger, dirtier one every time.
- Attach freshness metadata and use it. Every chunk should carry a last-verified date, and retrieval should prefer recency for volatile topics (pricing, policy) while tolerating age for stable ones (how-to guides).
- Build the eval set first. Before tuning anything, collect 50–100 real questions with verified answers. Without this, every "improvement" is a vibe.
- Close the loop. Give users a one-click way to flag wrong answers, and route those flags to document owners, not just to the ML team. The fix is usually an edit, not a fine-tune.
The reframe
The organizations getting real value from AI assistants aren't the ones with the fanciest retrieval pipelines. They're the ones that treated the project as what it actually is: a forcing function for the knowledge management they'd been deferring for a decade.
The model is the easy part. It always was.