opinion

Aug 16, 20245 mins

Emerging TechnologyGenerative AITechnology Industry

CIOs are so desperate to stop generative AI hallucinations they’ll believe anything. Unfortunately, Agentic RAG isn’t new and its abilities are exaggerated.

Agentic RAG is gaining significant attention as a method to reduce or eliminate hallucinations from generative AI tools. However, it may not effectively decrease hallucinations and could introduce new issues.

Agentic RAG, which stands for retrieval augmented generation, is not inherently problematic; it can work well for some users but may be underwhelming, costly, and labor-intensive for others.

This system is designed to integrate additional databases and data sources, providing a broader information range for generative AI algorithms. However, using AI to manage AI can complicate matters without guaranteeing better results.

Experts Alan Nichol, CTO at Rasa, and agentic specialist Sandi Besen shared their insights.

“Agentic RAG is an unnecessary buzzword,” Nichol stated. “It simply means adding a loop around your [large language models] and retrieval calls. The market is in a strange place where adding a simple ‘while’ loop or ‘if’ statement is presented as a revolutionary method. State-of-the-art web agents only achieve a 25% success rate, which is unacceptable in software development.”

“Companies should focus on building business logic in standard code,” he advised. “They can utilize LLMs to convert user input into structured formats and paraphrase search results, making them sound more natural.”

Nichol believes Agentic RAG is often unsuitable for enterprise data analytics. “It’s the wrong way to approach the problem,” he remarked. “A well-performing RAG is merely a basic search engine enhanced with some LLM capabilities.”

IT professionals should stop assuming that adding another LLM call will resolve hallucination issues, he added. “People are hoping this approach will magically address the root problem.”

What is the root problem? Data quality.

Nichol frequently encounters enterprises with poorly constructed retrieval systems due to unclean data. “Cleaning outdated information is tedious and unexciting, yet companies often add more LLM calls to mask their data issues,” he explained. “This approach will burden the LLM and yield subpar results.”

“It won’t resolve your problem but may create an illusion of a solution.”

Besen, an applied AI researcher at IBM, concurs that agentic methods can reduce hallucinations but acknowledges they may not always be the optimal enterprise strategy.

Besen warns that increasing complexity in a generative AI system can lead to unforeseen challenges. “Adding more agents inherently increases variability,” she noted. “However, with proper architecture and adequate prompting, the chances of hallucinations can decrease as evaluation and reasoning are integrated. For example, one agent can retrieve content while another assesses its relevance to the original question. Traditional RAG lacked this natural language reasoning check.”

As with any programming endeavor, success is not guaranteed. “There are ways to achieve success and ways to fail. The key is to align expectations with the technology’s capabilities,” Besen stated. “An agent’s effectiveness is limited by the language model it utilizes.”

Nevertheless, Besen emphasized that, despite vendor claims, even the best implementation of agentic RAG cannot entirely eliminate hallucinations. “It is impossible to completely eradicate hallucinations at this time, but a reduction is feasible.”

IT executives must determine whether they can tolerate the uncertainty and occasional inaccuracies. “If you seek consistent outcomes, avoid generative AI,” Besen advised. When considering the acceptance of occasional hallucinations, she suggested reflecting on how one would respond to an employee or contractor making mistakes. “Would you be comfortable with an employee being incorrect 10% of the time?”