Vectara's new Corrector cuts enterprise AI hallucinations to 0.9%

Enterprise teams love generative AI until an agent invents a financial figure or misquotes a policy document. In regulated fields, one hallucinated paragraph can kill a deal.

Enterprise AI assistant platform Vectara has rolled out a new Hallucination Corrector. The tool is directly integrated into its platform to catch and fix unreliable model responses before they cause damage.

Reasoning models are thinking more and lying more

Traditional LLMs carry an average hallucination rate between 3% and 10%. But the new wave of reasoning models—which break problems into step-by-step logic chains—are surprisingly prone to making things up.

Vectara's own research reveals that DeepSeek-R1, a reasoning model, hits a 14.3% hallucination rate, compared to 3.9% on its predecessor DeepSeek R3. OpenAI's GPT-o1 lands at 2.4%, up from 1.5% on GPT-4o. New Scientist published similar findings, noting elevated error rates across several reasoning architectures.

Vectara founder and CEO Amr Awadallah pointed out that while models are improving, they still fall short of accuracy requirements in strict fields like finance, healthcare, and law.

Spotting the lie and patching it at runtime

In initial testing, Vectara claims the Hallucination Corrector brought enterprise hallucination rates down to roughly 0.9%.

The system works alongside the company's Hughes Hallucination Evaluation Model (HHEM), which compares outputs against source files at runtime. HHEM scores statements from 0 (complete hallucination) to 1 (perfect accuracy). It pulled over 250,000 downloads on Hugging Face last month.

When a factual clash occurs, the corrector outputs an explanation detailing the hallucination, along with a revised version using minimal text changes to restore accuracy.

Enterprises can automatically surface the corrected text to end users, or keep the original flagged summary while experts review the logs to fine-tune their internal guardrails. For gray-area responses that are technically misleading rather than completely false, the system can tweak wording based on custom risk thresholds.

Why it matters

Every enterprise wants autonomous AI agents, but legal and compliance teams are terrified of rogue hallucinations. For developers building RAG setups or internal assistants, having an automated safety check at runtime makes high-stakes deployment far less risky.

We think Vectara is tackling the right problem at the right time. Instead of spending billions training bigger frontier models, focusing on the evaluation and correction layer targets the exact headache enterprise buyers face today—especially as reasoning models run wild with lengthy chain-of-thought traces.

That said, 0.9% is still not 0%. In medical and legal settings, even a tiny failure rate demands human oversight. Runtime document comparison also inevitably introduces extra latency and API overhead, so teams will need to decide which workflows truly justify the added check.