Photo credit: venturebeat.com
Contextual AI has introduced its new grounded language model (GLM), asserting that it achieves the highest level of factual accuracy in the AI sector. According to the company, it surpasses major competitors such as Google, Anthropic, and OpenAI on an important truthfulness benchmark.
Founded by the innovators behind retrieval-augmented generation (RAG) technology, Contextual AI reports that its GLM obtained an 88% factuality score on the FACTS benchmark. This positions the GLM ahead of Google’s Gemini 2.0 Flash at 84.6%, Anthropic’s Claude 3.5 Sonnet at 79.4%, and OpenAI’s GPT-4o at 78.8%.
Despite the revolutionary impact of large language models in enterprise software, issues of factual inaccuracy, commonly referred to as hallucinations, pose significant barriers to broader business use. In response, Contextual AI has developed a model that is significantly focused on enhancing accuracy to meet the stringent needs of enterprise RAG applications.
“We recognized that RAG — retrieval-augmented generation — would be a crucial aspect of the solution,” said Douwe Kiela, CEO and co-founder of Contextual AI, in a recent interview. “Our goal is to effectively elevate RAG practices to the next level.”
Unlike widely-used models such as ChatGPT or Claude, which cater to a range of applications from creative narratives to technical content, Contextual AI zeroes in on high-stakes enterprise scenarios where factual accuracy is paramount.
“In enterprise environments, particularly in regulated sectors, any misinformation is intolerable,” Kiela emphasized. “While a general-purpose model may suffice for creative needs, a more precise tool is crucial in applications where inaccuracies carry significant consequences.”
How Contextual AI Makes ‘Groundedness’ the New Standard for Enterprise Models
The term “groundedness” refers to the necessity for AI systems to respond based strictly on given context. This standard has become increasingly vital in industries like finance, healthcare, and telecommunications, where there is a pressing need for AI to either provide trustworthy information or admit uncertainty.
Kiela illustrated this principle with an example: “If a standard model is given a recipe that states, ‘this is usually true,’ it might still present it as an absolute fact. Our model, however, captures that context effectively, signaling that the information may not always apply.”
Expressing uncertainty, such as stating “I don’t know,” is critical for enterprise applications, Kiela remarked, adding that this feature can be immensely beneficial in professional settings.
Contextual AI’s RAG 2.0: A Revolutionary Integration
The platform incorporates an innovative method termed “RAG 2.0,” which aims to transcend the limitations of traditional RAG systems by optimizing all components as a cohesive whole.
“Conventional RAG setups often use standalone models for different tasks, which can create a disjointed system,” according to a company statement. “This can lead to a suboptimal combination of generative AI tools.”
In contrast, Contextual AI endeavors to improve the efficiency of its system by jointly refining its components. “Our innovative retrieval component intelligently assesses questions before devising a retrieval strategy,” Kiela explained.
This fully synchronized system is enhanced by what Kiela describes as “the best re-ranker globally,” ensuring that the most relevant information is highlighted prior to being processed by the grounded language model.
Extending Capabilities: Chart Reading and Database Connectivity
While the new GLM primarily focuses on text, Contextual AI’s platform now also accommodates multimodal content, including charts, diagrams, and structured data from top platforms like BigQuery, Snowflake, Redshift, and Postgres.
Kiela pointed out, “Significant enterprise challenges often arise at the intersection of structured and unstructured data. The most intriguing opportunities lie where these data types converge.”
The platform is already adept at handling complex visualizations, including circuit diagrams relevant to the semiconductor industry, as noted by Kiela.
Future Directions: Enhancing Business Tools
Looking ahead, Contextual AI intends to roll out its specialized re-ranking component shortly after the GLM’s release, alongside expanded features for document comprehension. The company is also exploring experimental features aimed at increasing the agentic capabilities of its offerings.
Established in 2023 by Kiela and Amanpreet Singh, who previously contributed to Meta’s Fundamental AI Research (FAIR) and Hugging Face, Contextual AI secures its clientele from notable enterprises, including HSBC, Qualcomm, and the Economist. The company positions itself as a facilitator for organizations striving to achieve tangible returns on AI investments.
Kiela summarized this opportunity, stating, “Companies pressured to demonstrate AI ROI should consider specialized solutions that genuinely address their needs. Although our grounded language model may feel less dynamic than standard models, it excels in ensuring contextual accuracy and reliability.”
Source
venturebeat.com