?? How to Handle Hallucinations in LLM Deployment

Generative AI, particularly Large Language Models (LLMs) like GPT, revolutionizes interaction by crafting responses specifically tailored to each question or prompt. This adaptability makes it invaluable in various sectors, from customer service to content creation. By analyzing vast data, these models generate relevant, context-aware responses, enhancing efficiency and user engagement.

?? The Hallucination Hurdle: A Threat to Business Credibility

However, the phenomenon of "hallucination" in LLMs poses a significant challenge. Hallucination refers to instances where the AI presents false or misleading information as fact. For businesses, this can be catastrophic. Imagine an AI confidently providing incorrect financial advice or misrepresenting a product. Such errors not only mislead customers but also erode trust in the AI's reliability, potentially diminishing its utility.

Example: If an AI inaccurately states a company's return policy, it could lead to customer dissatisfaction and operational disruptions.

?? Understanding the Source of Hallucinations: Why It Happens

Hallucinations in LLMs primarily stem from two sources:

1. Contextual Information Deficit: When the prompt lacks sufficient information, the model may default to its pre-trained knowledge base, leading to responses that are not grounded in the specific context of the prompt.Example: If asked about a current event without specific date references, the model might provide outdated information based on its pretrained data.
2. Confusion from Mixed Document Contexts: When multiple document chunks are combined in a single prompt, the mixing of facts and contexts can confuse the model. This may result in misinterpretation and false responses.Example: If a knowledge chunk of info about Product A is mixed together with another chunk that contains Return Policy of Product B - this may cause the model to misinterpret Product B's Return Policy as being that of Product A, providing customer with false information.

?? Measuring Hallucination: Ensuring Factual Consistency

To tackle hallucinations, we employ tools like Ragas (Evaluation | Ragas), specifically on measuring an AI's response's "faithfulness" metrics, which measures the factual consistency of the answer to the context base on the question.

??? Strategies to Prevent and Mitigate Hallucination

1. Prompt Engineering: The first line of defense is to set strict guidelines into its response synthesizer prompt that the AI should rely solely on information given from within the context section of its prompt (which is dynamically retrieved as part of RAG workflow), rather than its pre-trained knowledge, to formulate responses.Example: For a medical query, the AI is programmed to reference only to the context of medical document chunks given within the context of the prompt and is not allowed to use its pretrained knowledge to prevent it from suggesting any medication that might not be authorized within the scope of the bot.
2. Reranker Confidence: Document chunks retrieved from RAG workflow should also be filtered by confidence or relevancy score (ideally reranker score). This would prevent irrelevant document from mixing into the context of the prompt and confusing the model.
3. Validator Mechanisms: Tools like Ragas uses a separate set of LLM-powered evaluator to assess the relevance and truthfulness of both the question posed and the AI's response, ensuring alignment and accuracy.Example: If a user asks about climate change, the validator checks the response against provided context to make sure that the information is written within the context.
4. Providing Citations: While not a preventative measure, citing sources offers users a way to verify information independently, serving as a user-side check against misinformation.Example: When discussing historical events, the AI provides citations from recognized history websites or digital archives.

?? The Reality of Hallucination Rates

In practice, hallucination rates in LLMs can start as high as 20%. With continuous optimization, these can be reduced to around 1-2%. However, this residual rate implies that hallucination will still happen to your end users sooner or later. Therefore, we strongly suggest businesses implementing LLMs, like GPT, to transparently communicate the potential for errors. This helps set realistic expectations, ensuring users understand the need for occasional fact-checking and the limitations of current AI technologies, much like the disclaimers used in self-driving car technologies.