Brave Launches AI Grounding for Verifiable LLM Responses with State-of-the-Art Performance

Brave has just launched a significant update to its Brave Search API with the introduction of AI Grounding. This new feature is designed to anchor Large Language Model (LLM) responses in high-quality, verifiable data sources, significantly reducing hallucinations and improving the accuracy of AI-generated answers. The impact on practitioners is substantial: more reliable and trustworthy outputs from LLMs, which can be crucial for applications ranging from customer support to research.

What’s New with AI Grounding

AI Grounding ensures that LLM responses are grounded in factual information sourced from the web. This means that when an AI model generates a response, it's not just based on its training data but also on real-time, verifiable data. Here’s what this entails:

• Reduction in Hallucinations: By anchoring answers to verifiable sources, AI Grounding minimizes the risk of LLMs generating incorrect or misleading information.
• Improved Nuance Handling: The system is better equipped to handle complex and nuanced inputs, providing more contextually appropriate responses.

Performance on SimpleQA Benchmark

Brave’s AI Grounding has achieved state-of-the-art (SOTA) performance on the SimpleQA benchmark, with an F1-score of 94.1%. This is particularly noteworthy because Brave’s system was not specifically optimized for this benchmark; the high performance is a natural outcome of its design.

Real-World Applications

AI Grounding is already in use powering the Answer with AI feature on Brave Search, which handles over 15 million queries daily. It will soon support advanced research features in Brave Search and Brave Leo, the AI assistant integrated into the Brave browser. The Brave Search API is a critical component for many top AI LLMs, providing real-time web search data that enhances their responses.

Getting Started with AI Grounding

Customers of the Brave Search API can now access the new AI Grounding endpoint by subscribing to one of the public AI plans: Free, Base, or Pro. For more customized needs, you can contact Brave for custom plans. Here’s a breakdown of the available options:

• AI Grounding Plan (New):

  • Answers are grounded on a single search or the advanced Research mode, which involves multiple searches, reasoning, planning, and answer generation.
  • Pricing: $4 per thousand web searches plus $5 per million tokens (input and output).

• Pro AI Plan (Update):

  • Access to the new AI Grounding features via the existing openapi-compliant endpoint.
  • No change in pricing.

Technical Notes

At Brave, quality assessment is a multi-faceted process that combines both AI-based evaluations and human assessments. The dedicated team conducting these evaluations is trained to ensure consistency and coherence across various dimensions:

• Navigational Queries: Ensuring that the system can handle common search queries effectively.
• Blind Evaluations: Assessing final outputs without knowing the source to maintain objectivity.
• Factual Accuracy: Using benchmarks like SimpleQA to measure the factual accuracy of responses.

Conclusion

AI Grounding represents a significant step forward in making LLMs more reliable and trustworthy. By grounding responses in verifiable data, Brave is addressing one of the major challenges in AI-generated content-hallucinations. This update not only benefits API customers but also enhances the user experience for millions of users on Brave Search and Brave Leo.