HHEM: Expanded Language Support

Removing language barriers

We’re thrilled to announce that the HHEM now supports more languages, expanding from 3 to 8. In addition to English, German, and French, you can now evaluate hallucinations in Portuguese, Spanish, Arabic, Chinese - Simplified, and Korean.

This upgrade eliminates the need for workarounds like translating data into supported languages, making HHEM a more inclusive and seamless tool for global teams. Now you can build AI applications that support more languages and HHEM will deliver the same fast and accurate Factual Consistency Score (FCS).

Building trust through HHEM

Vectara developed HHEM because trust is the cornerstone of every successful AI system and Retrieval-Augmented Generation (RAG) application. Ensuring the accuracy of large language models (LLMs) is critical to delivering a trusted, responsible AI assistant and agentic experience. HHEM is a dedicated model designed specifically to identify hallucinations, which is when an LLM generates content not found in or based on its source data.

Recognizing the global demand for accurate, trustworthy AI, we’re taking this meaningful step toward ensuring accuracy across more languages.

Empowering businesses and teams

Hallucinations are a top concern for CIOs navigating high-stakes AI deployments. With broader language support, HHEM strengthens its role as a trusted ally for businesses aiming to reduce risk and increase confidence in their AI systems.

But the benefits go beyond technical improvements. Expanding HHEM’s language coverage creates opportunities for global collaboration and reduces the friction of adopting Vectara in your multi-lingual environment so you can accelerate the development of your assistant and agentic AI applications. Whether you’re developing a multilingual chatbot, analyzing international data, or evaluating cross-border projects, HHEM now ensures a seamless experience for everyone involved.

Improved latency and context window

In this new HHEM release, we’ve also upgraded the context window to 16k from 8k and reduced the latency. For a context of 4,096 (4k) tokens (both the retrieved results and the summary), the latency is under 500 milliseconds (ms) or 0.5 seconds (s); for 8,192 (8k) tokens, the latency is under 1,500 ms or 1.5 seconds; and for 16,384 (16k) tokens, the latency is around 5,000 ms or 5 seconds.

Conclusion

The future of AI trust is multi-lingual and this expansion is just the beginning. At Vectara, we’re committed to meeting your needs and pushing the boundaries of what’s possible with RAG applications. By removing barriers and broadening our platform’s capabilities, we’re empowering you to build AI systems that inspire confidence across languages and industries.

For the latest documentation on HHEM and how to use it, have a look here.

As always, we’d love to hear your feedback! Connect with us on our forums or on our Discord or on our community. If you’d like to see what Vectara can offer you for retrieval augmented generation on your application or website, sign up for an account!