News from the AI & ML world
DeeperML
@siliconangle.com
//
Vectara has announced the launch of its Hallucination Corrector, a new guardian agent integrated within the Vectara platform designed to significantly improve the reliability and accuracy of AI agents and assistants. This innovative approach aims to reduce AI hallucinations to below 1%, addressing a critical challenge in enterprise AI deployments where accuracy is paramount. The Hallucination Corrector builds upon Vectara's existing leadership in hallucination detection and mitigation, offering organizations a solution that not only identifies unreliable responses but also provides explanations and options for correction.
The Hallucination Corrector functions as a 'guardian agent,' actively monitoring and protecting agentic workflows. It leverages the Hughes Hallucination Evaluation Model (HHEM), a widely-used tool with 4 million downloads on Hugging Face, to compare AI-generated responses against source documents and identify inaccuracies. When inconsistencies are detected, the Corrector provides detailed explanations and corrected versions of the responses, ensuring minimal changes are made to maintain accuracy. This capability is particularly beneficial for smaller language models, enabling them to achieve accuracy levels comparable to larger models from Google and OpenAI.
According to Vectara, initial testing of the Hallucination Corrector has shown promising results, reducing hallucination rates in enterprise AI systems to approximately 0.9%. The system not only identifies and corrects factual inconsistencies but also provides a detailed explanation of why a statement is a hallucination. While the corrected output is automatically used in summaries for end-users, experts can utilize the detailed explanations and suggested fixes to refine their models and guardrails, further mitigating the risk of hallucinations in AI applications. The Hallucination Corrector represents a significant step towards building trusted AI and unlocking the full potential of generative AI for enterprises.
ImgSrc: d15shllkswkct0.
References :
The Hallucination Corrector functions as a 'guardian agent,' actively monitoring and protecting agentic workflows. It leverages the Hughes Hallucination Evaluation Model (HHEM), a widely-used tool with 4 million downloads on Hugging Face, to compare AI-generated responses against source documents and identify inaccuracies. When inconsistencies are detected, the Corrector provides detailed explanations and corrected versions of the responses, ensuring minimal changes are made to maintain accuracy. This capability is particularly beneficial for smaller language models, enabling them to achieve accuracy levels comparable to larger models from Google and OpenAI.
According to Vectara, initial testing of the Hallucination Corrector has shown promising results, reducing hallucination rates in enterprise AI systems to approximately 0.9%. The system not only identifies and corrects factual inconsistencies but also provides a detailed explanation of why a statement is a hallucination. While the corrected output is automatically used in summaries for end-users, experts can utilize the detailed explanations and suggested fixes to refine their models and guardrails, further mitigating the risk of hallucinations in AI applications. The Hallucination Corrector represents a significant step towards building trusted AI and unlocking the full potential of generative AI for enterprises.
ImgSrc: d15shllkswkct0.
Share:
References :
- aithority.com: Vectara Announces Significant Step Toward Greater Reliability & Accuracy for AI Agents and Assistants with Launch of Vectara Hallucination Corrector
- SiliconANGLE: Vectara launches Hallucination Corrector to increase the reliability of enterprise AI
- AI News | VentureBeat: Guardian agents: New approach could reduce AI hallucinations to below 1%
Classification:
- HashTags: #Vectara #AIHallucinations #GuardianAgent
- Company: Vectara
- Target: Enterprises
- Product: Hallucination Corrector
- Feature: Hallucination Correction
- Type: AI
- Severity: Informative